Note: Descriptions are shown in the official language in which they were submitted.
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
TOUCH-TYPABLE DEVICES BASED ON AMBIGUOUS CODES AMD
METHODS TO DESIGN SUCH DEVICES
FIELD OF INVENTION
This invention relates to the design of touch-
typable devices, and the use of touch-typable devices in
computing and telecommunications, and more particularly
to touch-typable devices based on strongly touch-typable
ambiguous codes and substantially optimal ambiguous
codes.
BACKGROUND OF THE INVENTION
Since the invention of the typewriter more than 100
years ago, keyboard engineering has been an active field
of research and development, resulting in many competing
designs. With the growth of personal computing and
telecommunications, the number of keyboard designs has
multiplied as designers attempt to accommodate the wide
variety of constraints and to exploit opportunities these
new technologies present. Nonetheless, much of the
variability of prior-art keyboard designs is not due to
this variety of constraints and opportunities. Rather, it
results from an incomplete appreciation on the part of
keyboard designers of the constraints inherent in the
problems they are trying to solve. It also reflects the
lack of general, effective methods for optimizing with
respect to these constraints. The present state of the
art is thus represented by a plethora of partial
solutions. These ills are cured by the keyboard design
methods taught by the present invention. To illustrate
the many facets of this invention, the optimizing methods
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
are applied to the design of a variety of device
embodiments, each preferred as the substantially optimal
solution of a given set of design constraints.
The instant invention relates to touch-typable
devices. Touch typing, like playing a musical
instrument, is a manual skill which is difficult to
learn. Once learned, it is difficult to modify the
acquired motor patterns. This difficulty places strong
constraints on keyboard design. The familiar Qwerty
keyboard (and its close variants such as the Azerty
keyboard used in France) owes its dominance to ingraining
and overlearning of the motor patterns involved in touch
typing. Thus, the wide established base of the Qwerty
keyboard has created a barrier to entry to improved
keyboards, such as the Dvorak keyboard. Indeed such
keyboards have gathered but a limited user community.
Due to its large number of keys, the Qwerty keyboard
is unsuitable for handheld and smaller typable devices.
The advent of such devices opens a niche for keyboard
designers. A new design in this niche which becomes
dominant will likely conserve its dominant position even
if more optimal designs appear later, due to the
intrinsic tendency of repetitive motor patterns to become
fixed. This prospect imposes an enormous burden of
responsibility on keyboard designers to avoid saddling
future generations of keyboard users with suboptimal
designs.
There are two main approaches in the prior art
toward reducing the number of input means required to
encode a given set of symbols 1) chording methods, in
2
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
which a combination of input means are activated to
encode each symbol, and 2) ambiguous codes, in which a
combination of symbols are encoded by each input means.
Chording methods have not met with practical success
since they have been heretofore difficult to learn to
operate, and few are willing to make the time investment
required. Thus, only ambiguous codes, or ambiguous codes
in combination with chording methods, hold any real
promise as a solution to this problem.
OBJECTS OF THE INVENTION
It is thus an object of the present invention to
provide methods for the design of substantially optimal
ambiguous codes for typable devices.
It is a further object of this invention to provide
methods for the design of strongly-touch-typable
ambiguous codes for typable devices.
A further object of this invention is to provide
keyboards suitable for touch typing on both full-sized
and miniature keyboards.
A further object of this invention is to enable
sending alphanumeric messages from ordinary phones or
two-way pagers to other such devices, without human
intervention, and thus inexpensively.
A further object of this invention is to provide
touch-typable personal digital assistants.
3
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
A further object of this invention is to provide
keyboards which are typable by the driver of a vehicle
without unnecessarily distracting the driver.
A further object of this invention is to provide for
typable communication devices which are inexpensive to
manufacture and work with standard telephone
communication systems.
A further object of some of the preferred
embodiments of this invention is to facilitate the
transfer of typing skills of touch typists trained on
conventional keyboards to novel keyboards through partial
conservation of the layout of the conventional keyboard
in the layout of the novel keyboard.
A further object of this invention is to provide
general methods to produce ambiguous codes which have a
substantially minimal lookup error rate.
A further object of this invention is to provide
general methods to produce ambiguous codes which have a
substantially minimal query rate.
A further object of this invention is to provide a
device to reduce typing injuries.
A further object of this invention is to provide
a handheld computing device which is twice foldable.
4
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
A further object of this invention is to provide a
one-handed keypad suitable for implementation on a hand-
held computer.
A further object of this invention is to provide a
one- and two-handed keypad suitable for implementation on
a hand-held computer or a desktop keypad.
A further object of this invention is to provide
keyboards which are Qwerty-like.
A further object of this invention is to provide
easily learnable chording keyboards.
A further object of this invention is to provide
synergistic hybrids of chording and ambiguous keyboards.
A further object of this invention is to provide a
touch-typing-oriented querying mechanism for typable
devices embodying ambiguous codes.
A further object of this invention is to provide a
touch-typing-oriented disambiguation mode for typable
devices embodying ambiguous codes.
A further object of this invention is to provide a
hybrid chording/ambiguous code keyboard fully compatible
with the standard telephone keyboard.
A further object of this invention is to provide
ergonomic assignments of symbols to modes.
5
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
A further object of this invention is to provide a
substantially transparent touch-typable interface for
typable devices comprising touch screens.
A further object of this invention is to provide
optimization across a set of natural languages.
A further object of this invention is to provide a
device typable using the one hand holding the device,
with reduced scanning time.
Still further objects of this invention will be
described in the detailed description below.
BRIEF DESCRIPTION OF THE DRAWINGS
The detailed description of the preferred
embodiments of the present invention will be discussed
with reference to the drawings, a brief description of
which follows.
Figure 1 shows an overview of the optimization
considerations for producing a typable device according
to the present invention.
Figure 2 shows a flowchart for the construction of
devices based on strongly touch typable ambiguous codes.
Figure 3 shows a flowchart for the construction of
ambiguous codes satisfying at least one ergonomic
criterion, and optimized with respect to these ergonomic
criteria.
6
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
Figure 4 shows a flow chart of particular embodiment
of the method of Figure 3 using a random search
optimization method.
Figure 5 shows the distribution of lookup error
probability for randomly chosen ambiguous codes on
several selected keys.
Figure 6 shows the distribution of query probability
for randomly chosen ambiguous codes on several selected
number of keys.
Figure 7 shows a flow chart for directed random walk
optimization.
Figure 8 shows a flow chart of the construction of
strongly touch-typable ambiguous codes.
Figure 9 plots lookup error rate vs. number of keys
for randomly chosen, and substantially optimized
ambiguous codes.
Figure 10 plots query rate vs. number of keys for
randomly chosen, and substantially optimized ambiguous
codes.
Figure 11 shows lookup error rate vs. query rate for
some substantially optimized ambiguous codes on a range
of number of keys.
Figure 12 shows a table relating levels of strong
touch typability to the number of keys required to
7
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
achieve that level, for several different optimization
methods.
Figure 13 shows a flow chart of the method for
synthesizing encoding symbols.
In order to help the reader appreciate the unity of
the present invention in the face of the multitude of
apparatus embodiments which are required to clearly and
distinctly point out the broad scope and various aspects
of the invention, a table summarizing these embodiments
and their major features is shown in Figure 14.
Figure 15 shows a smart-card embodiment with 16 keys
devoted to encoding of letter symbols.
Figure 16 shows a smart-card embodiment with 9 keys
devoted to encoding of letter symbols.
Figure 17 shows a keyboard embedded in a steering
wheel.
Figure 18 shows a telephone with a substantially
optimal code on 10 keys.
Figure 19 shows an example reduced-ambiguity
alphabetically ordered ambiguous code in application to a
portable telephone.
Figure 20 shows a Qwerty-like keyboard, optimized
with respect to lookup error rate and query rate, while
8
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
respecting the ordering of letters on each row of the
Qwerty keyboard.
Figure 21 shows an alternate Qwerty-like keyboard.
Figure 22 shows an ambiguous keyboard embodied in a
standard numeric keypad layout.
Figure 23 shows an ergonomic touch-typing-oriented
disambiguation mechanism.
Figure 24 shows a flow chart for a method to allow
queries to be answered in a touch-typing oriented manner.
Figure 25 shows a one-handed embodiment of a
keyboard designed for conservation of typing skills
between one-handed and two-handed keyboards.
Figure 26 shows a two-handed embodiment of a
keyboard designed for conservation of typing skills
between one-handed and two-handed keyboards. In this
case, the two-handed keyboard is weighted for maximum
similarity in typing motions between the two keyboards.
Figure 27 shows a two-handed embodiment of a
keyboard designed for conservation of typing skills
between one-handed and two-handed keyboards. In this
case, the two-handed keyboard is evenly weighted between
the two hands.
Figure 28 shows an integrated mouse/keyboard.
9
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
Figure 29 shows a top view of a twice-foldable
information appliance in the unfolded state.
Figure 30 shows a bottom view of a twice-foldable
information appliance in the unfolded state.
Figure 31 shows a twice-foldable information
appliance in the once-folded state, revealing an
additional functionality.
Figure 32 shows a twice-foldable information
appliance in the twice-folded state, revealing yet
another functionality.
Figure 33 shows a twice-foldable information
appliance in a detached state, allowing two-handed
typing.
Figure 34 shows a typical personal digital assistant
with a touch screen.
Figure 35 shows a typical personal digital assistant
with a potentially transparent keyboard.
Figures 36A, B and C show three modes for a 16-key
keyboard.
Figure 37 shows a standard telephone layout.
Figure 38 shows a hybrid chording/ambiguous code
keyboard embodied in a telephone.
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
Figure 39 shows the distribution of lookup error
rates and query rates fbr all hybrid chording/ambiguous
codes of a specified structure, compared to the lookup
and query rate of the standard ambiguous code.
Figure 40 shows a flow chart for the creation of
multi-level strongly touch-typable ambiguous codes.
Figure 41 shows a flowchart for the creation of a
specific embodiment of a multi-level strongly touchable
ambiguous code.
Figure 42 shows a typable device suitable for
implementation of the multi-level ambiguous code of
figure 41.
Figure 43 shows the device of figure 42, operating
to display the first level of a multi-level ambiguous
code.
Figure 44 shows the second-level code of a multi-
level ambiguous code.
Figure 45 shows the device of figure 42 operating to
display part of the second level code of a multi-level
ambiguous code.
Figure 46 shows the sequence of operating states of
the device of figure 42 while used in combination with a
multi-level ambiguous code to type the word "think".
11
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
Figure 47, like figure 46 shows the operating states
of the device of figure 42 while used in combination with
a multi-level ambiguous code to type the word "think". In
this case, however, the operation of a visual cache to
reduce scan time is show as well.
DETAILED DESCRIPTION OF THE INVENTION
Definitions and Basic Notions
This section collects definitions of words and
concepts which will be used in the following detailed
specification.
Language Given a set of symbols, one can construct
sequences of symbols, and assign probabilities to the
sequences. The set of symbols, sequences of symbols, and
the probabilities assigned to the sequences will be
referred to here as a language. For clarity of
discussion, and without limiting the scope of this
invention, the languages we will refer to are written
natural languages, such as English, and though for
concreteness we may refer to symbols as "letters" or
"punctuation", it will be understood by those of ordinary
skill in the art that symbols in this discussion may be
any discrete unit of writing, including standard symbols
such as Chinese ideograms or invented symbols such as the
name of the artist formerly known as Prince.
12
CA 02353862 2001-06-08
WO 00135091 PCT/US99/29343
Keyboard/Input Means A keyboard is a component of a
communications and/or computing device which transforms
physical movement by an operator into symbol sequences.
Keyboards comprise at least one input means which
is responsible for the transformation of some subset of
the physical motions operative to activate the keyboard
into some subset of symbol sequences.
The physical movement used to operate a keyboard
is typically in the form of motion of fingers and/or
thumb or of a hand-held stylus. This definition extends
to other bodily motions, such as head, tongue, or eye
motions which might serve to signal a choice of symbol
from the keyboard. A device comprising a keyboard
according to this definition will be referred to as a
typable device.
By "typable device" we understand not just the
physical device containing the keyboard, but the entire
communication system in which this typable device is
embedded, the limits of that system defined by dependence
on the underlying ambiguous coding scheme. In the case
of a typable device in which input symbols appear
directly on a display which is physically part of the
typable device, the limits of the system are clear and
defined by the physical perimeters of the device. In
more general cases, in which for example the typable
device includes a telephone handset sending information
to a central computer, the central computer begin
responsible for decoding or otherwise acting on the
textual information communicated from the handset, then
the "typable device" must be understood to include the
13
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
central computer, as configured to operate in the
required manner by software built in view of the
teachings of this invention.
It will be appreciated that each of the at least one
input means comprised in a keyboard can take a wide
variety of physical manifestations. The essential
feature of an input means is that it permits an operator
to select a subset from the set of symbols to be encoded
by the keyboard. With this appreciation, and in order to
increase the readability of this present specification,
the word "key" will often be used interchangeably with
the words "input means".
Typing is the process of sequentially selecting at least
one input means in order to select sequences of subsets
of symbols from the set of symbols which can be encoded
by the keyboard. It is to be appreciated that well-known
handwriting recognition software permits a kind of typing
in which the input means translates a collection of
drawing motions into the selection of a subset of a set
of symbol s .
Touch typing is the process by which the symbol
sequences are generated from the keyboard using only or
predominately kinesthetic rather than visual or auditory
feedback.
Strongly correlated symbols and symbol sequences It is
well known that different letters appear in words with
different frequencies. For instance, in the previous
sentence, the letter "e" appeared 11 times, while the
14
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
letter "z" did not appear at all. This is also true of
pairs of letters, triples of letters, and so on. It is a
related fact that words do not all occur with the same
frequency. The 3-letter word "the" is very common in
English, while the 3-letter word "zap" is rather
uncommon. These statistical irregularities can be used
in the design of ambiguous codes. Indeed, statistical
irregularities have been exploited in keyboard design at
least since the invention of Qwerty.
We are particularly concerned with symbols and
symbol sequences whose distribution in typical samples of
text is substantially strongly correlated with the
distribution of other symbols or symbol sequences, such
symbols will be called strongly correlated symbols. For
example, the symbol "." often used in English and other
languages to indicate the end of a sentence may be a
strongly correlated symbol since the distribution of
sentence length is not random in typical text. In Hebrew,
the symbol ".11 is correlated as well with particular
letter symbols since Hebrew uses a different symbol for
some letters occurring at the end of a word, and ends of
sentences are correlated with ends of words.
Reference Statistics The reference statistics on
symbols sequences used to measure the correlation between
symbols are typically estimated by analysis of a
reference corpus. A reference corpus is a large
collection of text chosen to represent some aspect of
language. As is well known to linguists, there are
significant, fundamental problems in constructing corpora
to represent general features of a language, as opposed
CA 02353862 2001-06-08
WO 00/35091 PCT/11S99/29343
to features pertaining to particular classes of text or
particular types of writers. These problems are beyond
the scope of the present invention. Here we refer
throughout to a set of reference statistics gleaned from
analysis of the British National Corpus, one of the
largest corpora existing at the present time for analysis
of English. Choosing a corpus is a necessary step toward
gathering results which permit various methods and
embodiments to be compared. Nothing in this particular
choice should be construed as limiting the scope of this
invention. In particular, the choice of a corpus of
English language texts is an arbitrary choice. The same
analysis could be performed for any other written natural
language.
Encodings and Decodings In the United States, the keys
on a telephone keypad are often labelled with letters as
well as numbers, typically with the key corresponding to
the number 2 also corresponding to the letters a, b, and
c, the key corresponding to the number 3 also
corresponding to the letters d, e, and f, and so on in
the standard ordering of letters in English.
Thus, the sequence of key presses associated to the
digit sequence 233 also corresponds to the letter
sequences add, bee, and bed, all of which are English
words, as well as various meaningless letter sequences
such as cff. Here, a sequence is considered to have
meaning if it appears in a reference list of meaningful
sequences. Thus all of these letter sequences,
meaningful or not, are associated with the same digit
sequence. We will say that the sequence of key presses
16
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
233 is an encoding and the sequences add, bee bed, eff
and so on are decodings of the encoding 233. When no
confusion will arise, "decodings" may be used to mean
"meaningful decodings". The set of symbols used in
decodings, in this example, letters in the alphabet, will
be referred to as decoding symbols, or simply symbols if
no confusion will arise, and the set of symbols used in
encodings, in this example, digits, will be referred to
as encoding symbols.
Ambiguous codes Ambiguous codes as such are well known
in the art. On the standard telephone keypad used in the
United States, there are 12 keys, 10 of which encode a
digit, and several of these, typically 8, encode in
addition 3 or 4 letters of the alphabet, arranged in
alphabetic order. These assignments produce an ambiguous
code which we will call the standard ambiguous code.
This code is abc def ghi jkl mno pqrs tuv wxyz.
Since several letters are encoded on each key, some
method of disambiguation must be used to decide which of
the several letters is intended by the operator. In
typical applications, such as a voice response system,
the intended letter is found by comparing the input
sequence with a list of stored responses. In the event
that several of the stored responses correspond to the
input sequence, the user is presented with a list of
these responses, from which he or she must choose. The
order in which these choices are presented may be
arbitrary, or may depend on the frequency by which each
response is the correct response, with the responses
presented in decreasing order of frequency.
17
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
Standard keyboards. There are essentially three standard
keyboards in wide use: the Qwerty keyboard and its close
variants, the 12-key telephone keypad and the typically
17-key numeric keypad and its close variants. It is a
unique advantage of this invention to provide keying
methods useful both on the standard telephone and numeric
keypads, as well as on a specially designed keypads
disclosed herein.
Strong touch typability A weakly touch-typable device
is a device in which the assignment of symbols to keys is
essentially fixed; only relative to such a device can a
typist develop physical reflexes for encoding particular
symbols using particular motor patterns. We will say
that a typable device is strongly touch typable if it is
1) weakly touch typable, 2) based on an ambiguous code,
and yet 3) such that in a normal mode of operation, a
touch typist can use the typable device to produce text
at an acceptable level of accuracy without being unduly
distracted from the touch typing task to intervene in the
disambiguation process.
Strong touch typability is a matter of degree; it is
a measure of touch typability which depends on a host of
factors, some pertaining to the individual typists, some
pertaining to the uses to which the typable device will
be typically put, and some pertaining to the structure of
the typable device itself. For a given touch typists,
for instance, a given typable device may be sufficiently
strongly touch typable for some typing tasks, but not for
others.
18
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
It is to be appreciated that two of the key elements
of the definition of strong touch typability, accuracy of
the text produced and the distraction of the touch-typing
user, depend on a number of factors, including:
= the disambiguation means,
= the context of the use of the machine, for
instance while driving or while sitting at a
desk,
= the kind of text to be typed, which determines
in part the level of accuracy required,
= the reference statistics,
= the skill of the individual typist,
= individual preferences, and
= the means by which the attention of the user is
drawn to the disambiguation mechanism (for
instance, a voice-synthesis mechanism which
speaks the words or words in a query to the
user may or may not be more distracting than a
bell or a flashing light).
Though strong touch typability, like temperature, is
a matter of a degree, it is, like temperature, perfectly
well-defined. The strong touch typability of a typable
device can be quantitatively measured, with respect to
any user or group of users, once these various factors
are fixed, using standard experimental protocols, well-
known to those skilled in the art. Furthermore, two
components of strong touch typability can be measured
directly from an ambiguous code: lookup error and query
error. Thus, numerical values of strong touch typability
can be assigned without any direct reference to a
19
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
population of users, but only with reference to the
ambiguous code in question.
Like temperature, there is a lower bound to strong
touch-typability. It is clear that a device which
requires user intervention after every word or even after
every three words in order to disambiguate cannot be
considered strongly touch-typable with respect to any
touch typist engaged in any task. The lower bound of
strong touch typability can be expressed in terms of
continuity of attention. If a user's attention must be
substantially continually focused on the operation of the
disambiguation mechanism to produce text which is
disambiguous to an acceptable level, then the device is
not strongly touch typable.
The practical lower bound of strong touch typability
pertains to a user of average skill in the art of touch
typing, and is higher than the theoretical lower bound
just described. In order to bring numerical as well as
conceptual precision and definiteness to the inventive
notion of strong touch typability, numerical values are
assigned to strong touch typability in terms of values
of lookup error and query error. This numerical
characterization serves to further distinctly point out
the differences between the inventive methods and devices
of the present disclosure and all prior art methods and
devices.
A strongly touch typable ambiguous code is an
ambiguous code on which strongly touch typable devices
may be based.
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
Feedback Devices In devices which permit the user to
intervene at various points in the decoding of symbol
sequences generated using ambiguous codes, some manner of
sensory feedback to the user is required. Typically, this
feedback will be in the form of a graphical
representation of symbols, however, feedback could take
many forms, such as auditory, tactile, or even olfactory.
Ergonomic Factors Design of keyboards implementing
ambiguous codes involves satisfying many constraints.
These may include reduction of lookup error rate,
reduction of query rate, selection of a number of keys
consistent with the size of the desired keyboard,
compatibility with existing keyboards such as the Qwerty
keyboard, phone keypad, or numeric keypad, regularity of
partition structure, anatomic fidelity, minimal mode-
changing key use, partition structure, compatibility
between one- and two-handed typing, and conservation of
conventions such as alphabetic ordering. Other
constraints include: the ergonomics of disambiguation
mechanisms, the ergonomics of the encoding of weakly
correlated symbols, look-and-feel, and availability of
computing resources at the sending and receiving ends of
a communication system utilizing ambiguous codes.
Lookup Error measures the error committed by a
disambiguation mechanism which disambiguates by
systematically selecting the most-probable (meaningful)
decoding from the set of possible decodings of an
ambiguous sequence. Thus, the lookup error rate of a code
is the sum, over all possible decodings which are not the
most probable decoding of an ambiguous sequence, of the
21
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
reference probability of the possible decodings. In the
case of word-based disambiguation, these sequences begin
and end with a"space" symbol, that is, are words. Lookup
error is the probability that the most-likely decoding is
not the correct one. The lookup error is conveniently
expressed as a rate, the lookup error rate, in units of
words per lookup error. The lookup error rate is the
reciprocal of the lookup error probability.
Query probability is the sum, over all (meaningful)
decodings which are not unique (meaningful) decodings, of
the reference probability of said decodings. This gives
the probability that a given word will have more than
one meaningful decoding, and therefore a query must be
made of the user as to decide which of these decodings to
use. The reciprocal of the query probability is the
query rate, expressed in units of words per query. The
query rate gives the average number of words entered
between queries.
Substantial Optimality A code will be said to be
substantially optimal with respect to a property if it is
among the best codes with respect to that property given
other constraints imposed on the code. For example, a
code on 20 keys may have a lower value of the lookup
error rate than a code on 2 keys, and yet the code on 2
keys may be substantially optimal with respect to the
lookup error rate given the constraint that the code be
on 2 keys. Substantially optimal ergonomic codes will be
defined as codes which are simultaneously substantially
optimal with respect to each of a collection of ergonomic
22
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
constraints. Such constraints include but are not limited
to key number, lookup error rate and query rate. For
these three constraints, pairs of constraints are
correlated. Lookup error rates tends to increase with
query rate, and both lookup error and query rates tend to
increase as key number decreases. The best value
possible for a given criterion when this given criterion
is the sole optimization criterion may be better than the
best possible value obtainable when some other criterion
must be optimized as well. Thus, the ergonomic
constraints relevant to a given design must be decided
upon and their importance weighted as an initial step of
the optimization methods.taught by this invention.
It is to be emphasized that the optimality of an
ambiguous code cannot be discussed in the absolute, but
must be evaluated relative to a set of reference
statistics for the language to be encoded. Indeed, given
any ambiguous code, it is possible to construct a set of
statistics such that code is optimal with respect to the
constructed statistics.
Given a set of reference statistics, an estimate of
the optimality of a given code can be obtained from
experiments comprising the generation of random codes, as
well be discussed in more detail below.
Disambiguation Methods Substantial optimality for
ambiguous codes is well defined only in reference to a
chosen disambiguation method. A code which is
substantially optimal with respect to one disambiguation
23
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
method may not be substantially optimal with respect to
another method of disambiguation.
At least two disambiguation methods are well known
in the art. These are word-based and block-based
disambiguation. In word-based disambiguation, a list of
words along with their probabilities is used to choose
among alternate decodings of a given encoding in the
ambiguous code. For instance, all words in the list which
are meaningful decodings of a given encoding may be
compared, and the word with the largest probability
selected. Block-based disambiguation is similar, except
that the list contains fragments of text up to some size,
along with the probability of the fragments.
Both word-based and block-based disambiguation
methods are special cases of a more general framework,
which we will call sequence-based disambiguation, in
which a list of sequences is associated with a
probability, and disambiguation is effected by reference
to this list. It is to be noted that the "space" symbol
which defines word boundaries in languages such as
English is for the purposes of this discussion no
different from any other symbol. One can define a list of
sequences and sequence probabilities in which said
sequences include the "space" symbol, and thus extend
beyond word boundaries. One can go further and define
sequences which include a wildcard symbol and thus define
lists of sequences which contain arbitrary subsequences,
which subsequences may or may not correspond to words in
the language. In this way, arbitrarily complex
representations of a language can be built up, and can
24
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
be used in a disambiguation method. For instance,
syntactic and semantic relationships between subsequences
can be brought in to resolve conflicts between possible
interpretations of an ambiguously coded sequence. For
clarity, we focus this specification on well-known word-
based disambiguation, unless otherwise specified. It
will be appreciated by those skilled in the art that the
methods taught by this invention do not depend on word-
based disambiguation; any other disambiguation method can
be used.
Partitions A partition of an integer n is a set of
integers such that the sum of the elements of the set is
equal to n. Typically, a given integer admits many
partitions, e.g. the integer 5 has the partition 3:2, but
also the partition 2:2:1. Algorithms for generating all
the partitions of an integer are well known to those
skilled in the art. Most prior art codes use an even-as-
possible partition. That is, a partition in which, to
the extent possible given the number of keys in relation
to the number of letters to be encoded, the number of
letters per key is the same. As will be further expanded
below, this choice is a sensible choice with respect to
some ergonomic considerations, it may be sub-optimal with
respect to others.
There are two genera of ambiguous codes for which
exclusive rights are herein claimed. These are 1)
strongly touch-typable ambiguous codes, and 2)
substantially optimal ambiguous codes. Ambiguous codes
may be substantially optimal but not strongly touch
typable, strongly touch typable but not substantially
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
optimal, neither substantially optimal nor strongly touch
typable, or both substantially optimal and strongly touch
typable.
The disclosure begins by pointing out how to make
ambiguous codes in both of these genera, and identifying
whether a code is contained in either of the genera. it
then explains how to use codes in both of these genera to
make typable devices, and how these codes may be used to
solve various design problems confronting the designer of
typable devices.
The best mode for practicing this invention depends
on the constellation of design criteria which are to be
optimized according to the teachings of the invention.
Thus, several particular, practically relevant and
useful, situations are chosen to illustrate the range of
the methods and devices taught by the present invention.
The range of machines which can be built by persons
skilled in the art according to the teachings of this
invention extends considerably beyond the specifics of
the preferred embodiments presented herein. Various
extreme or particular cases of design constraints are
solved in these embodiments. Given the teachings embodied
in these cases, it will be clear to one skilled in the
art how to combine features appropriately in order to
solve intermediate or hybrid design problems.
One embodiment is optimized with respect to lookup
error rate exclusively. This embodiment is designed for a
machine with limited memory and computing power, such as
26
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
a smart card. With such a machine, computing resources
may not be available to support a complex querying
mechanism for user intervention in the disambiguation
process. Thus this machine uses one of the simplest
possible disambiguation mechanisms, which comprises
systematic selection of the most-probable decoding of any
given encoding.
Another embodiment is optimized with respect to
query rate exclusively. This embodiment is designed for
use by the driver of a vehicle, such as an automobile.
Though computing power may be available to support a
complex querying mechanism, use of such mechanism should
be kept to a minimum, so as to distract the driver as
little as possible from driving.
A next embodiment provides a phone keypad optimized
with respect to both lookup error rate and query rate,
and which is compatible with the layout of standard
telephone keypads.
Another embodiment is optimized with respect a
conventional criterion: preservation of alphabetic order.
Letters are arranged on the standard touch-tone keypad in
alphabet order. it is possible to preserve the
alphabetic ordering of the conventional telephone keypad
and yet reduce lookup and query rates by optimizing over
partitions.
27
CA 02353862 2001-06-08
WO 00/35091 PCT/US99129343
Optimization over partitions leads to an additional
embodiment in which keyboards with substantially optimal
query and lookup error are exhibited which preserve as
well as possible the traditional Qwerty keyboard
arrangement.
A further embodiment illustrating the design of
keyboards which correspond as well as possible to
conventional designs is a keyboard based on an ambiguous
code and corresponding to a numeric keypad.
For many applications, a keyboard which can be
ergonomically operated in both an ambiguous and a non-
ambiguous fashion is desired. To this end, it is
preferable to chose ambiguous codes on a number of keys
which divides the number of symbols to be encoded. A
number of keys equal to % the number of symbols is
particularly preferred. The desirable consequences of
this preferred choice are exhibited by this next
embodiment.
Another embodiment shows how keyboards can be
optimized for cross-platform compatibility. In this
embodiment two keyboards, a one-handed keyboard and a
two-handed keyboard, are designed to be operated in
potentially rapid alteration, in such a way that touch-
typing motions used to operate one of the keyboards
transfer seamlessly to touch-typing motions used to
operate the other keyboard. This keyboard has the
additional advantage of having the potential to reduce
typing injuries, as well as other objects and advantages,
as will be described in the detailed specification.
28
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
The above-mentioned embodiments taken together show
that different keyboard uses imply different kinds of
optimality, which, since a given user may need keyboards
for several different uses, in turn implies that
mechanisms must be provided for several different
solutions to co-exist in a single device. A surprising
solution to this problem, made possible by the small
typable device sizes achievable with ambiguous codes, is
the twice-foldable personal digital assistant described
in this embodiment.
The embodiments discussed up to this point involve
both hardware and software specification. However, it
is possible to achieve many of the objects of this
invention using a purely, or predominantly, software
solution. An example software solution is worked out in
detail to show how specifics of existing hardware can be
incorporated using appropriate software to achieve some
of the objects of this invention.
A final set of embodiments synergistically unites,
for the first time, two alternative approaches to
producing keyboards with a small number of keys: chording
methods and ambiguous-code methods.
First it is shown that by ergonomic construction of
chording patterns, coupled with optimization of lookup
error rate and query rate, an ambiguous code on n keys
can be made to behave like an ambiguous code on m
substantially larger than n keys. When, in particular,
this method is applied to the standard ambiguous code,
the 8 letter keys of the standard ambiguous code gain the
29
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
properties of substantially optimal codes without
chording on 13 keys. Comments on how to extend ambiguous
code creation discussed throughout with reference to
English can be extended to other languages. For
concreteness, this discussion is carried out with respect
to this embodiment, but the comments apply generally to
all embodiments.
Second, it is shown that combining a divide and
conquer method with the methods exemplified by the
previous embodiment, the number of input means can be
further reduced, in this example 4 input means are used
to operate an ambiguous encoding with 16 elements. The
number 4 is chosen so that a handheld device embodying
this code can be operated using the fingers and thumb of
the hand holding the device.
Operational overview of a strongly touch typable device.
Figure 2 shows an operational overview of a strongly
touch typable device based on an ambiguous code. Such a
device possesses input means, the manipulation of which
140 causes sequences of encoding symbols to be generated
141. A strongly touch typable ambiguous code is used to
map these sequences of encoding symbols to sequences of
decoding symbols in step 142. These sequences of encoding
symbols may then be selectively output, either on a
display for direct observation by the user of the device,
or in some electronic form for further processing,
transmission, or storage 142.
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
It should be pointed out that the ambiguous code of
the device of figure 2 overview satisfy other ergonomic
criteria in addition to strong touch typability.
Construction of substantially optimal codes The
methodological steps to optimize an ambiguous code with
respect to a set of ergonomic criteria are explained in
reference to figure 3. In summary, the steps are as
follows:
= 2000 selecting a set of statistically
correlated decoding symbols to be represented
in an ambiguous code, comprising the sub steps
of
- 2007 selecting a set of reference
statistics,
- 2008 analyzing the statistical
correlation of symbols relative to
the statistics selected in
step 2007
= 2001 selecting a disambiguation method.
= 2002 selecting the number of encoding symbols.
= 2003 selecting the ergonomic criteria with
respect to which the code should be
substantially optimal.
= 2004 weighting the importance of ergonomic
criteria selected in step 2003.
0 2005 selecting an optimization method.
31
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
~ 2006 applying the optimization method selected
in step 2005, whereby substantially
optimal ambiguous codes are produced.
It will be appreciated that the steps 2000 through
2003 might be applied in any order, and that the
application of one of these steps could influence the
choices at other of these steps. Details regarding the
application of these steps will now be explained.
Step 2000 selecting a set of statistically
correlated symbols to be represented in an ambiguous
code. This step is comprised of the substeps of 2007
selecting a set of reference statistics, and 2008
analyzing the statistical correlation of symbols relative
to the statistics selected in step 2007. The goal of
these steps is to identify those symbols which are
capable of being represented ambiguously. All
disambiguation methods work by exploiting correlations
between symbols to make predictions about which sequence
of decoding symbols should be associated with a given
sequence of encoding symbols. If a decoding symbol is
distributed randomly throughout all texts to be encoded,
then it cannot be represented in an ambiguous code, since
no predictions can be made about a randomly distributed
symbol. Typically, for any natural language, the symbols
used to encode that language (for instance, letters in
the case of English, ideograms in the case of Chinese)
are sufficiently statistically correlated that an
effective ambiguous code for these symbols can be
designed. There may be other symbols, such as punctuation
symbols, which are significantly statistically correlated
32
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
with each other, and with the letters or ideograms used
to write the language. The fine details of steps 2007 and
2008 depend on the natural language to be represented.
Analysis of statistical correlation of symbols used in
written natural language is a well-known art to
linguists.
Step 2001, selecting a disambiguation method. As
has already been mentioned, there are currently at least
two well-know methods of disambiguation, block-based and
word-based disambiguation. Both of these methods use
statistical,correlation between symbols to make
predictions about which sequence of decoding symbols to
set in correspondence with a given sequence of encoding
symbols. Both block-based and word-based methods can be
augmented though use of higher-level information about a
language, such as its syntax and semantics. The goal of
this present method is to construct an ambiguous code
si.ich that, relative to the selected disambiguation
method, optimal selection of a decoding sequence to
correspond to each encoding sequence. Therefore, the
details of the selected disambiguation method can
influence the detailed nature of the ambiguous code to be
thus designed. This method will be illustrated with
respect to the selection of word-based disambiguation as
the disambiguation method, though other disambiguation
methods will also be discussed.
Step 2002, selecting the number of encoding symbols.
The selection of the number of encoding symbols is
crucial to the design of a typable device based on
ambiguous codes. This selection is made in view of many
33
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
factors, including the size of the typable device and the
acceptable level of ambiguity. These factors and their
interplay are best explained in reference to concrete
examples; such examples are taken up later in this
disclosure.
Step 2003, selecting the ergonomic criteria with
respect to which the code should be substantially
optimal. An essential aspect of this invention is the
discovery and definition of several ergonomic criteria
which determine the quality of a typable device based on
ambiguous codes. These criteria include strong touch
typability, lookup rate, query rate, anatomic fidelity,
physiological fidelity, conservation of convention,
partition structure, cross-platform compatibility,
regularity of layout, and scan rate. Depending on the
application, one or more of these criteria may be
relevant to the design of a typable device.
Step 2004, weighting the importance of ergonomic
criteria selected in step 2003. When more than one
ergonomic criterion is relevant to the design of a
typable device, some weighting of the importance of these
criteria must be decided upon. It is rarely the case that
the same optimum with respect to a given ergonomic
criterion can be achieved when that ergonomic criterion
is considered in isolation as when it is also necessary
to optimize relative to another ergonomic criterion.
Step 2005, selecting an optimization method. Two
optimization methods will be discussed in more detail
below, random selection, and directed random walk. Of the
34
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
two, random selection is typically easier to implement,
yet directed random walk produces better codes. These
two methods are representative of a large class of
methods which might be appropriate for the design of a
given typable device. In some cases, for instance the
first chording/ ambiguous code device considered below,
the number of codes to be examined is small enough that
all of them can be checked exhaustively.
Step 2006, applying the optimization method selected
in step 2005, whereby substantially optimal ambiguous
codes are produced. Regardless of the optimization
method selected in step 2005, some skill must be used
when applying the method to produce substantially optimal
ambiguous codes. In particular, when an optimum is
required with respect to several ergonomic criteria at
once, it is preferable to consider each ergonomic
criterion first in isolation, whereby an estimate can be
made of the code quality ultimately achievable. This
estimate can be invaluable for fine-tuning the
optimization process, as will be discussed in more detail
below, once two optimization methods have been fully
introduced.
Random Search The basic method for finding a code with
good properties is to choose codes at random, test their
properties, and select those which have the best
properties. Exhaustive enumeration, in which all'codes in
the candidate set are tested, is typically not a viable
option since the number of codes is too large to be
tested in any reasonable amount of computer time.
CA 02353862 2001-06-08
WO 00/35091 PCTIUS99/29343
Random search provides a benchmark by which the
utility of other methods of code selection can be
measured. Suppose that a set of ergonomic criteria, and
a weighting on those criteria is given. One can estimate
the substantial optimality of a first ambiguous code with
respect to those criteria and those weightings by
generating additional ambiguous codes at random. If in a
small number of random trials it is possible to find a
code with equal or better values with respect to the
given ergonomic criteria than the first code, then that
first code is not substantially optimal.
If, on the other hand, it can be shown that a
substantially large number of random trials is required
to produce a code with values better than or equal to the
first code, or that on better code exists, then the first
code is substantially optimal.
With reference to figure 4 we state in detail a
method for rejecting the hypothesis that a candidate
ambiguous code is substantially optimal. In summary, the
steps are:
= 3000 determining a set of relevant constraints
which define an appropriate set of codes
which contains the candidate code.
= 3001 determining the set of ergonomic criteria
with respect to which the candidate code
may be substantially optimal.
36
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
= 3002 randomly selecting a subset of codes from
the set determined in step 3001.
= 3003 evaluating each of the codes selected in
step 3002 with respect to each of the
ergonomic criteria determined in
step 3001.
= 3004 comparing the values of the candidate code
with respect to the ergonomic criteria
selected in step 3001 with the values
found in step 3003. If any of the values
found in step 3003 are more optimal than
the values of the candidate code, then the
hypothesis that the candidate code is
substantially optimal can be rejected.
Details concerning these steps:
Step 3000, determining a set of relevant constraints
which define an appropriate set of codes which contains
the candidate code. The set with respect to which the
substantial optimality of a candidate code is to be
evaluated must be appropriately defined. Some of the
potentially relevant constraints are: number of encoding
symbols, partition structure, and admission of a
specified ordering, such as alphabetic ordering. Each of
these constraints limits the set of codes to which the
candidate code is appropriately compared to.
Step 3001, determining the set of ergonomic criteria
with respect to which the candidate code may be
substantially optimal. Some of the criteria which might
37
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
be relevant for the analysis of the candidate code are:
lookup error rate, query rate, admission of a
specified ordering, such as alphabetic ordering,
admission of a regular layout, and anatomic fidelity.
Once steps 3000 and 3001 are performed, a
distribution of code properties over a set is defined,
and this distribution can be sampled randomly. Figure 5
presents an example in which sets of codes are defined as
having 1) an even-as-possible partition, and 2) a
specified number of encoding symbols, where the specified
number is 7, 9, 11, and 13. This definition completes
step 3000. Then it is determined that lookup error is the
sole relevant ergonomic criterion. This determination
completes step 3001. Together, these steps determine a
distribution, the shape of which can be determined by
random sampling, steps 3002 and 3003. In the figure 5,
5000 codes from each distribution are selected,
completing step 3002, and the lookup error of each is
measured, completing step 3003. The data are presented as
percent lookup error (the reciprocal of lookup error
rate) vs. the number of codes with the given percent
lookup error. It is seen that the distributions become
increasing strongly peaked as the number of keys
increases. If the process is repeated but with query
probability replacing lookup probability, we obtain the
data shown in figure 6.
To illustrate step 3004, a candidate code whose
substantial optimality to be tested is selected. This
code is the 14-key code pn gt cr zk wj a e hi so ud xf ym
vl qb proposed by [1]. The lookup error of this code
38
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
relative to our reference statistics is 105 words/lookup
error. Proceeding as above, we determine that 14-key code
with an even-as-possible partition with lookup error
equal to or better than that of the candidate code can be
found in 7 random trials on average. If we repeat the
process, except for using query rate instead of lookup
error rate as the relevant ergonomic criterion, we find
that a code with better query rate than the candidate
code (4 words/query) will be found in 3 out of 4 random
trials on average. Thus the ambiguous code of [1]
substantially optimal either with respect to lookup error
rate, or with respect to query rate. Indeed, with respect
to query rate, most codes are better than the given code.
As a rule of thumb, if a code has not been
explicitly optimized with respect to an ergonomic
criterion, then it is likely that it is not substantially
optimal with respect to that criterion, as measured with
respect to any reasonable set of language statistics.
Directed Random Walk
Directed random walk is an iterative optimization
method wherein, at each step, a previously best code is
used as a seed for generating new codes, one or more of
which may be better than the best previously found. As
the process is iterated, better and better codes are thus
found. The procedure will first be explained
intuitively, and then more formally.
In the present context, optimization of ambiguous
codes with respect to one or more ergonomic criteria, we
39
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
assume no detailed knowledge of the structure of the
space to be searched. In the absence of such knowledge,
one is blind of foresight as to which direction to move
to best continue the search. Thus, the safest procedure
to take small as possible steps in as many as possible
directions, and refrain from moving until steps in each
of these as many as possible directions have been
evaluated and compared. As an accumulation of small
steps may lead a searcher into a cul-de-sac, any such
search should be augmented with a"restart" procedure
which always the search step back out of unpromising or
blocked avenues.
More formally, the problem is to take minimal steps
though the space of ambiguous codes and direct these
steps though that space toward the desired codes.
According to the teachings of this invention,
substantially minimal steps in space of ambiguous codes
correspond to single pairwise permutations of assignments
of decoding symbols to encoding symbols. At each step of
the optimization method, it is desirable to test as many
pairwise permutations as possible, preferably all
possible pairwise permutations. The step is completed by
choosing the pairwise permutation which gives the largest
improvement in the property to be thereby optimized. If
there is no largest improvement, then one of the pairwise
permutations is chosen at random.
In reference to figure 7, the steps of the method
are as follows:
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
= 4000 choosing a starting code from the set of
candidate codes.
= 4001 generating new codes from the starting
code by perturbation of the starting code,
preferably, by pairwise permutations of
the assignment of symbols to keys,
preferably all possible pairwise
permutations.
= 4002 measuring the properties of the codes thus
generated.
= 4003 checking if a stopping criterion has been
reached, such as a criterion of limited
further improvement.
= 4004 outputting the best code, if the stopping
criterion has been reached.
= 4005 if the stopping criterion has not been
reached, checking to see if the set of
codes generated by perturbation of the
current starting code contain a code
better than the current best code.
= 4006 if there is a code.in the current set of
codes better than the current best,
selecting.the best of these as the new
current best.
41
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
= 4007 selecting a new starting code. If step
4005 yields YES, selecting as new starting
code the best code from the current set,
else, select a new starting code from the
current set at random. Upon completion,
return to step 4001.
When there is but one criterion to be optimized,
selection of the best code from the set of candidate
codes is a simple matter of choosing the code in the set
which has the most optimal value of the criterion.
However, when there are several criteria to be
simultaneously optimized there is but a partial ordering
on the values of the criteria, and it is not obvious how
to select among these values in order to best advance the
optimization procedure.
One way to perform simultaneous optimization is to
optimize with respect to each variable independently.
Then, in cases of conflict, where the simultaneous
optimum can not be achieved, this being the generic case,
some weighting of the importance of each criterion must
be established. And that relative weighting is part of
the design constraints.
Construction of Strongly Touch-typable Codes In order to
facilitate the explanation of how to make and use
strongly touch-typable ambiguous codes, we will fix three
increasingly strict levels of strong touch typability.
= Level A This level of touch typability is
exemplified by a casual, tolerant typist
42
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
characterized in that he or she 1) types 20
words per minute and accepts distractions every
15 seconds, that is, a query rate of one query
every 5 words on average, and 2) accepts a 2
percent lookup error, that is, a lookup error
rate of one error every 50 words, or two and
minutes of typing.
= Level B This level of touch typability is
exemplified by a less casual, less tolerant
typist characterized in that he or she 1) types
words per minute and accepts distractions
every 30 seconds, that is, a query rate of one
query every 10 words on average, and 2) accepts
15 a 1 percent lookup error, that is, a lookup
error rate of one error every 100 words, or 5
minutes of typing.
~ Level C This level of touch typability is
exemplified by a skilled typist characterized
20 in that he or she 1) types 40 words per minute
and accepts distractions every 30 seconds, that
is, a query every 20 words on average, and 2)
accepts a 0.5 percent lookup error, that is, a
lookup error every 200 words, or 5 minutes of
typing.
With reference to figure 8, we point out that the
method to construct strongly touch touchable codes
comprises the following steps:
43
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
= 5000 Determining quantitative values of
acceptable lookup error rate and query
rate.
= 5001 Selecting an ambiguous code optimization
method.
= 5002 Determining the minimal number of keys
required such that using said number of
keys and the optimization method selected
in step 5001 it is possible to achieve the
values of lookup error rate and query rate
determined in step 5000.
= 5003 Determining the maximal number of keys
allowable given the design of the target
typable device.
= 5004 Deciding if the design criteria are
compatible. If the number determined in
step 5003 is greater than or equal to the
number determined in step 5002, then the
design criteria are compatible, otherwise
they are not.
= 5005 If the design criteria are compatible, as
determined in step 5004, apply the
optimization procedure selected in step
5001 to construct an appropriate strongly
touch-typable ambiguous code. If they are
not compatible then the procedure fails.
44
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
Details of this method are as follows:
Step 5000, Determining quantitative values of
acceptable lookup error rate and query rate. This could
be done by testing of an individual or group of
individuals, or by simply preselecting desired values for
lookup error rate and query rate, for instance, by
selecting a level of strong touch-typability as described
above.
Step 5001, Selecting an ambiguous code optimization
method. In reference to the construction of
substantially optimal ambiguous codes above, two
optimization methods were discussed: random search and
directed random walk. Random search is less powerful than
directed random walk, but may suffice if the number of
allowed keys is high enough, and the level of desired
strong touch-typability is low enough. An ever weaker
method, selection of a code in a single random trial,
could be sufficient in some circumstances. To see this
in more quantitative detail, some experimental results
are discussed in reference to figures 9, 10 and 11.
In this experiment, 5000 ambiguous codes with an
even-as-possible partition were selected at random from
each of the sets of ambiguous codes for 2-20 keys. In
addition, for each number of keys 2-20, an optimization
run was performed using directed random walk, in each of
three conditions, 1) optimization for lookup error rate
only, 2) optimization for query rate only, and 3)
optimization for both lookup error and query rate, using
a target value method. From the values of lookup error
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
rate and query rate calculated for the randomly selected
codes, the following statistics were computed: best
value, worst value, average value, and median value. All
these statistics are plotted in figure 9 for the lookup
error rate, and in figure 10 for the query rate, together
with the results of an optimization run in which lookup
error rate and query rate, respectively, was the sole
ergonomic criterion optimized for. The results from
optimization runs in which lookup error rate and query
rate were simultaneously optimized are shown in
figure 11. From all these data, a decision can be
reached as to which optimization method to use. While it
is always preferable to use the most powerful method at
one's disposal, it may be that a less powerful method may
be sufficient, for instance, any single, randomly
selected code could meet specified criteria, if these
criteria are sufficiently lax. This will be discussed
further below.
Step 5002, Determining the minimal number of keys
required such that using said number of keys and the
optimization method selected in step 5001 it is possible
to achieve the values of lookup error rate and query rate
determined in step 5000.
With reference to the experimental results described
above, and the selected levels of strong touch typability
described above, one can construct a table giving the
minimum number of keys required for each of the three
levels of strong touch typability, and reference to the
three types of optimization mentioned. This table is
presented in figure 12.
46
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
Step 5003, Determining the maximal number of keys
allowable given the design of the target typable device.
Ambiguous codes will be typically used in small devices,
and the number of keys will generally be a compromise
between key size and total typable device size. in some
cases, convention may enforce a key number, such as the
convention of using 12 keys for a telephone keypad.
Step 5004, Deciding if the design criteria are
compatible. If the number determined in step 5003 is
greater than or equal to the number determined in step
5002, then the design criteria are compatible, otherwise
they are not.
As will be seen more clearly in the detailed
specification of the device embodiments presented below,
the number of keys permissible in a typable device can
depend on many factors, and can be more or less rigidly
determined by these factors.
Step 5005, If the design criteria are compatible, as
determined in step 5004, apply the optimization procedure
selected in step 5001 to construct an appropriate
strongly touch-typable ambiguous code. If they are not
compatible then the procedure fails.
If the procedure fails then at least one of the
following things must happen:
= A stronger optimization method is chosen.
47
- ~--------.-_- --
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
= The device design is modified to allow for more
keys.
= A lower value of strong touch typability is
accepted.
= The device is abandoned.
Smart Card on 9 to 16 letter keys Smart cards are
substantially credit-card sized devices containing
computer components such as a processor and a memory.
They are currently used in applications such as security
and banking, but have many other possible uses. This
embodiment shows how it is possible to equip a smart-
card-sized device with a touch-typable keyboard, and thus
vastly expand the range of applications which these
devices can serve. As a simple example, in banking and
security applications, smart card users must currently
remember a string of digits which is the password for the
device. However, with a typable smart card, easy-to-
remember, though relatively long, pass phrase in natural
language could be used in the place of a difficult-to-
remember, albeit short, numeric password. Examples of
smart-card sized devices to which the teachings of this
embodiment could be applied include the personal digital
assistants manufactured by the Franklin Corporation and
sold under the trademark REX.
Given present technology, the small size of smart
cards substantially forbids complex and power-consuming
communication components for transmission of data entered
on the keyboard on the card. Thus, this smart-card
48
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
embodiment teaches a low-cost machine for ergonomically
and efficiently sending messages using standard touch
tones, and standard touch-tone generators.
Most telephones have but 12 keys, each associated
with a touch tone, in the sense that activating each key
causes a distinguished touch tone to be emitted by the
phone. However, the universal DTMF standard provides for
16 touch tones, and the DTMF tone generator installed in
most telephones is capable of generating all 16 of these
tones. By exploiting the additional tones, each of up to
16 keys can be assigned to a touch tone, and used to
encode alphanumeric symbol sequences. Other things being
equal, the larger the number of keys, the lower the
ambiguity of codes associated to these keys. The teaching
of this embodiment is thus to use substantially all of
the 16 touch tones to encode alphanumeric sequences. In
this way, machinery for the communication of information
with low-ambiguity codes can be produced using readily
available, low-cost components.
This embodiment has the further objects of
= providing a touch-typable keyboard for a smart-
card-sized device.
= providing a method for simulating a set of
encoding symbols which is larger than the set
of encoding symbols which can be physically
generated by a device
49
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
= providing an example of a device in which
lookup error is the dominant ergonomic
criterion.
= providing a keyboard/visual display device
geometry which is adapted for smart-card-sized
devices.
= providing a disambiguation mechanism which is
operable using very limited computer memory.
= providing an example of a system wherein more
than one disambiguation mechanism may be
operable, each adapted to local computational
capabilities. In this case, a first
disambiguation mechanism is used to provide
feedback to the user, at the sending end of a
communication, while a second disambiguation
mechanism is used at the receiving end of the
communication.
We will now discuss in detail the manner in which
these objects are achieved by the present embodiment.
Strongly touch-typable keyboard Smart card devices are
small, so only a small number of substantially full-
sized keys can be located on them. If some of the area of
the card is to be reserved for a visual display device,
then the area available for keys is further reduced. The
preferable compromise in number of keys between the
requirement of substantially full-sized keys for touch-
typability, and the requirement of a large number of keys
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
to allow for low-ambiguity codes, is in the range of 9-16
keys. Two possible layouts for devices with a number of
keys in this range are shown in figure 16 and 15. The
arrangement and functionality of the keys and their
relationship to other components of the smart card will
be discussed in detail below.
Synthesizing encoding symbols using encoding sequences
with meaningless decodings With reference to figure 13,
let us divide the set of decoding symbols into two
subsets: 1) a core set, consisting of symbols to be
associated encoding symbols bearing a one-to-one
relationship with physical input means, and 2) an
auxiliary set consisting of symbols to be associated with
encoding symbols bearing a many-to-one relationship with
physical input means (step 100). Then, the method of
synthesizing encoding symbols further comprises the steps
of
= 101 establishing a first, potentially
ambiguous, code which associates subsets of the
core set with encoding symbols, said encoding
symbols bearing a one-to-one relationship with
physical signals which can be generated by the
typable device for physically representing
encoding symbols,
= 102 identifying short sequences of encoding
symbols which are such that no possible
decoding of the sequence of encoding symbols
forms part of a meaningful decoding.
51
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
= 103 establishing a second, potentially
ambiguous, code as a relationship between
subsets of the auxiliary set of decoding
symbols with the short sequences of encoding
symbols identified in step 102.
For example, let us associate 16 encoding symbols
with the 16 DTMF tones, so that said tones physically
represent the encoding symbols. The tones will be
labelled (0,1,2,3,4,5,6,7,8,9, *,#,A,B,C,D). We will
take as the core set of symbols the letters [a-z], and
associate them with the physically representable encoding
symbols via a first ambiguous.code as follow
(0,aw), (1,bi) , (2,cx) , (3,d) , (4,eJ) , (5,fo), (6,g) , (7,hv) ,
(8,ky) . (9,1) , (*,mu) , (#,n) . (A,pZ) . (B,qr). (C,s) . (D,t) .
where the first element of each pair gives the encoding
symbol, and the second element of each pair gives the
decoding symbols associated to the encoding symbol. The
auxiliary set of decoding symbols will be a singleton'set
consisting of the symbol "space". Thus we synthesize one
encoding symbol to represent the one decoding symbol
"space". A candidate sequence is A8A, which corresponds
to the following decoding sequences (pkp pkz zkp zkz pyp
pyz zyp zyz). None of these decoding sequences form part
of any word in our reference list of meaningful
sequences, thus the encoding sequence A8A is a suitable
sequence for representing an element of the auxiliary
set, and we form the pair (ABA,"space") to represent the
"space" symbol. The "space" symbol can then be associated
with an input means, which input means will cause the
sequence of tones associated to A8A to be emitted each
time said input means is activated. On the receiving end,
52
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
a decoding means will transform the sequence A8A to the
"space" symbol. Whether a given input means is associated
to a single physically representable encoding symbol, or
a synthesized encoding symbol, is entirely transparent to
the user. An arbitrarily large auxiliary set of decoding
symbols can be represented in this way. It will be
appreciated that given the above specification, a
programmer of ordinary skill would be easily capable of
creating software to automatically generate any desired
number of synthesized encoding symbols, given a set of
reference statistics, and a first (ambiguous) code for
the core set of encoding symbols.
Providing a disambiguation mechanism which is operable
using very limited computer memory. The limited
processing power and memory capacity of present-
generation smart cards puts a substantial premium on low
ambiguity in ambiguous code designs; little computing
power is available in the card to be devoted to
disambiguation machinery.
For any ambiguous code, most of the disambiguation
effort, on the part of the computing hardware and
software as well as on the part of the user, is incurred
in selecting which of the alternative decodings to an
ambiguous encoding should be selected. In view of the
limited computing capability of the smart card, querying
for alternative decodings can be eliminated entirely. In
the absence of querying, only the most probable decoding
for each encoding need be stored, for only the most-
probable decoding sequence will be output by the
disambiguation mechanism when each encoding sequence is
53
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
received. With this simplification, a particularly
compact form of the database can be obtained, e.g. in the
form of a simple suffix tree. Since querying is
eliminated sufficient visual feedback can be provided
using only a simple, low-power requirement display, for
example, a single-line traveling banner display related
to the type of display used in pocket calculators or
digital watches.
This method of disambiguation, in which only the
most-probable decoding sequence is stored and output,
we will refer to as simple lookup disambiguation. Simple
lookup disambiguation is only effectively operable with
ambiguous codes which are sufficiently strongly touch
typable. Thus a surprising consequence of strongly touch
typable codes is that they permit effective operation of
very simplified disambiguation mechanisms.
An example 16-key substantially optimal ambiguous
code suitable for application in the present embodiment
is the code aw bi cx d ej to g hv ky 1 mu n pz qr s t
with lookup error rate 4043 words/lookup error, and a
query rate of 68 words/query, this code is shown is an
example layout 51 on a 16-letter-key smart card in
figure 15. This figure also reveals a display means 50
for displaying decodings of encoding symbols entered via
the keyboard, and an auxiliary input means 51, which is
thumb-activatable, and could be used to encode a variety
of additional symbols and mode changes, as will be
discussed more fully in reference to other embodiments.
Here it is to be especially noted that the display means
50 is preferably placed in such a way that 1) both the
54
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
letter input means 51 and the thumb-actuated auxiliary
input means 52 are in a comfortable position to be
actuated by one hand (in this figure, the right hand),
while, at the same time, allowing the keys to be as large
as possible, given the small size of the smart card, and
also allowing the screen to be comfortably and fully
viewed in the frame formed by the thumb and the index
finger of the hand actuating the input means 51 and 52.
This unique and preferred arrangement solves the problem
of allowing a touch-typable keyboard and a large-as-
possible display means to functionally co-exist on a
smart card.
Since queries are not permitted at the sending end
in this embodiment, this code was chosen by directed-
random-walk optimization using lookup error as the only
criterion of optimality. It is to be noted that lookup
errors using this code will occur on average only once
every roughly 16 pages of typed text. Thus this code is
suitable for accurate communication of substantially long
messages, even in the absence of a querying mechanism. If
it were desirable to sacrifice some lookup optimality at
the sending end in order to reduce processing of queries
at the receiving end, an example alternate code optimized
for both query and lookup error is aw bu cx d eV pz go hv
im ky 1 NQ p r s t with a lookup error rate of 2670
words/lookup and a query rate of 101 words/query. Choice
of a code optimized both with respect to lookup error
rate and query rate would be appropriate in at least two
circumstances, 1) if the smart card were in fact
sufficiently powerful to support a query mechanism,
and/or 2) a query-based disambiguation mechanism would be
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
used at the receiving end of communications initiated at
the smart card. This latter might be the case, for
instance, if the user composed messages using a smart
card, sent them to another computer over a phone line,
and at some later moment performed a second
disambiguation pass using a more powerful disambiguation
mechanism. Indeed, the second disambiguation pass need
not be performed by the person who composed the message,
but could be performed by a second person, for instance
the first person's secretary.
In any case, the lookup error rate of this second
code is still extremely low, as compared, for instance,
to the rate at which even very skilled typists make
typing errors, approximately 1 error every 100 words. By
any reasonable measure, both of these 16-letter-key codes
must be considered strongly touch typable, as a typist
typing 20 words per minute will only need to answer a
query once every three minutes for the first 16-letter-
key-code, and once every five minutes for the second 16-
letter-key-code. Performing the same optimizations for
ambiguous codes for 9 letter keys, we find, for example,
the code akw bnq cly dhx epv fim gr jot suz optimized
only for lookup error rate, and with lookup error and
query rates of 116 words/lookup error and 4.4
words/query respectively. This code is shown in an
example layout on a 9-letter-key smart card in figure 16.
Optimizing for both lookup error rate and query rate sing
the directed random walk method, one can also construct
codes such as am bnz cfi dhx evw gjr kos luy pqt with
lookup error and query rates of 109 words/lookup error
and 6.2 words/query respectively. It is to be noted
56
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
that since no querying can be performed using simple
lookup disambiguation, strong touch typability can only
be discussed in terms of the lookup error rate. An
evaluation must be made as to whether lookup errors occur
at a sufficiently low rate as to produce acceptable text.
Even for these 9-key codes, the lookup error rate is
comparable to the rate at which skilled typists produce
typing errors, hence these codes can be considered
strongly touch typable in this context. Further, since
the smart card would typically be used for short
messaging, composition of electronic mail, communications
to pagers, and the like, standards for text accuracy may
be lower than standards for transcription of final-copy
text, for instance. It is from these considerations that
we arrive at the definition of 9 to 16 keys as the
preferred range for this embodiment. More than 16 keys
are difficult to fit into a smart card format while
retaining the advantages of substantially full-sized
keys. On the other hand, ambiguous codes on fewer than 9
keys may not be strongly touch typable with respect to
the simple disambiguation mechanisms compatible with the
smart card's limited computing power.
Feedback to the typist A smart card equipped with simple
lookup disambiguation could be operated by a person who
is a competent touch typist, needing no feedback from the
card as to the progress of the communication, and/or
getting feedback from the machinery at the receiving end
of the communication, potentially over a phone line and
potentially in the form of speech synthesized on the
basis of the symbols input by the typist on the smart
card. It is desirable, however, to provide feedback
57
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
directly from the card whenever there are sufficient
computing resources built into the card to supply that
feedback.
Here it is pointed out that the computing resources
needed to provide useful feedback are even less than the
computing resources needed for simple lookup
disambiguation. Even if no disambiguation database and
software is present on the card, and using only
rudimentary electronic circuitry, well known to those
skilled in the art, a unique character can be sent
directly to the display in response to each key press,
where said character is the most probable (according to
the reference statistics) letter associated with the key.
Using for example the code aw bi cx d ej fo g hv ky 1 mu
n pz qr a t described above, the resulting text would
typically be quite readable by a human. For example, the
first line of the Gettysburg address, rendered using 1-
block (single-letter) statistics reads as follows:
oour score and sehen kears ago our oathers irought
oorth on this continent, a nea nation, conceihed in
liiertk, and dedicated to the proposition that all
uen are created erual.
This level of accuracy is already enough provide a
rough guide to the typist as to the text he or she is in
the process of entering on the smart card. This example
shows that disambiguation can be accomplished with
extremely small amounts of memory; here the only memory
required is that needed to store the 16 characters which
will be displayed in response to activation of the 16
58
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
keys. This approach is scalable in terms of the computing
resources required. With successively more memory,
2-,3-, and higher-block probabilities could be stored
and used as the basis of well-known block-based
disambiguation, and thus render the text with increasing
accuracy for display to the user.
While block-based disambiguation is well known in
the art, it has heretofore proved to be not practical.
This example shows the reason for this: block-based
disambiguation is not powerful enough to effectively
disambiguate codes which are too ambiguous. In this
example block-based disambiguation is coupled with a
code which is sufficiently strongly touch-typable,
sufficiently disambiguous, that it permits effective
disambiguation with a block-based method. The prior art
has taught away from block-based disambiguation in favor
or word-based disambiguation. However, in application
of the teachings of this invention, block-based
disambiguation is made operable and viable for practical
use.
The example further shows that 1) word-based
disambiguation is not required to practice the teachings
of this invention, 2) a micro-processor is not required
to practice the teachings of this invention, and 3) more
than one, potentially different, disambiguation
mechanisms can be used in the same communication system
based on ambiguous codes. Word-based disambiguation, or
another disambiguation method, could be used at the
receiving end of a communication sent from the smart
card, while locally the smart card is using simple block-
59
WO 00/35091 PCT/US99129343
based disambiguation to provide feedback to the user of
the smart card.
As the size of the blocks used in block-based
disambiguation increases, the accuracy of text rendering
increases as well. However, at some block size the amount
of storage required approaches that of the storage
requirements of word-based disambiguation, and since
word-based disambiguation generally gives better results
than block-based disambiguation, word-based
disambiguation will generally be preferred, if there is
enough memory available to support it.
Further applications If there is memory available on the
smart card beyond that required for storage of the
disambiguation database and software, then the potential
applications of this device are greatly increased. For
instance, with only a few bytes of additional memory, a
user could call an appropriate voice response system for
telephone directory information, request a phone number
by typing in a name and other identifying information on
the smart-card keyboard, and could then store the
retrieved phone number in the user memory for later
downloading to another, more powerful, machine.
Query Error Minimization - a typable device for a vehicle
The present embodiment concerns an instance in which
query error is the dominant constraint on typable device
design. It is generally of value to reduce the query
rate since the effort to answer queries distracts from
the typing task. In some applications, however,
reduction of query rate is of paramount importance.
WO 00/35091 PCT/US99/29343
Queries will be displayed on a visual display
in most practical implementations of typable devices
implementing ambiguous codes. When the vision of the
user is otherwise urgently occupied, for example when the
user is driving a car, then, for safety reasons,
distracting that vision to the evaluation of queries must
be kept to a minimum. Even when queries are made by
auditory means, it is crucial to minimize driver
distraction. Further, while driving a car, both hands of
the user are generally occupied in holding the steering
wheel of the car, and should preferably not be removed
from the steering wheel to operate a typable device.
This object can be achieved by embedding the input means
of the typable device directly in the steering wheel.
Referring now to figure 17, we note that any number
of input means could be embedded in a steering wheel
200. Many steering wheels comprise crenulations on the
inner or rear surface of the wheel to provide better grip
for the fingers. For these steering wheels, it is
natural to associate a first plurality of input means 201
with each of a plurality of these crenulations. When the
driver grasps the steering wheel, four of the first input
means 201 are contacted by the fingers of each hand. The
region of the steering wheel contacted by the drivers
hands may change from moment to moment, for instance when
the driver turns the steering wheel through a large
angle. Which set of 8 keys are contacted by the driver's
hands at any one time can be recognized by a position-
sensing means, such as a combination of pressure
sensitive keys with simple electronic circuitry, which
combination will be evident to those skilled in the art.
61
WO 00/35091 PCTIUS99/29343
A second plurality of input means 202 can be placed along
the outer or upper surface of the steering wheel, whereby
one of said second input means is contacted by the thumb
of each hand while the steering wheel is grasped by the
driver. Again, which of the second input means are
contacted by the thumbs of the driver at any one time can
be detected by an appropriate position-detecting means.
With respect to the steering-wheel embedded keyboard
just sketched, it is natural to select a code on 8 keys,
said keys to be associated with the first input means
contacted by the fingers of both hands, along with two
mode-switching keys to be associated with the second
input means contacted by the thumbs of the driver.
Ambiguous code selection Applying the directed-random-
walk method taught by this invention for selecting
substantially optimal codes, and optimizing only with
respect to the query rate, we construct, for instance,
the following code on 8 letter keys: aksz bcev dfi gmo
hqt jnw luy prx with a lookup error rate of 70.2
words/lookup error, and a query rate of 4.1 words/query.
As throughout this specification, these rates are
calculated with respect to our reference statistics, and
using simple word-based disambiguation as the
disambiguation method. The query rate for this code might
well be too high to consider this code to be strongly
touch-typable. A typist/driver typing 20 words/minute
would be distracted from driving by a query approximately
every 12 second, likely too often to be compatible with
safe driving. On the other hand, even a skilled typist
may not be able to type 20 words per minute while
62
WO 00/35091 PCT/tJS99/29343
driving, potentially bringing the relationship between
query and typing rates into an acceptable range for
strong touch typability.
There are several additional strategies for reducing
query rate, beyond choosing a substantially optimal code,
and these strategies can be used in combination. They
include
= Increasing the total number of keys by
increasing the number of keys activatable by
each finger. It will be appreciated that this
could be done, for instance, by adding another
row of keys on the steering wheel, or,
equivalently, making each key
multipositionable, or using a chording method
by which two or more keys are pressed
simultaneously to encode a different subset of
encoding symbols.
= Eliminating queries when the less-probable
decodings are much less probable than the most-
probable decodings. The parameter controlling
how close the probability must be between most-
probable and less-probable decodings must be to
invoke a query is a parameter whose value could
be selected by the user. Such a mechanism
could be of value in any embodiment in which
query rate is a relevant ergonomic criterion.
= Using a hybrid chording/ambiguous code method,
as described in detail below.
63
WO 00/35091 PCT/US99/29343
= Using a disambiguation method which is more
powerful than simple word-based disambiguation.
Phone keypads compatible with existing phone keypads
In this embodiment, the limitation on the number of keys
is paramount since the keypad must be compatible with
existing telephone equipment which generally have 12
keys. For the purposes of this embodiment, we require
that two of these keys must be reserved for non-letter
symbols such as space, backspace, period, and an end-of-
transmission symbol. Thus, the 26 letters must be
distributed over at most 10 keys. In this embodiment
both minimal lookup error rate and minimal query rate are
desired. We find that using the optimization methods
taught by this invention and using the even-as-possible
partition on 10 keys, codes such as amq be cdu fiy gpx hl
jsv krz nw ot with lookup error rate of 138 words/lookup
error, and query rate of 9.3 words/query can be found,
while simultaneously optimizing for lookup error rate and
query rate. This should be compared to the lookup error
rate of 29 words/lookup error and the query rate 2.2
words/query of the standard ambiguous code, thus, there
is an overall improvement of by more than a factor of 4
over the standard ambiguous code. Said 10-key code
optimized for lookup error rate and query rate is shown
in an example layout on a telephone keypad in figure 18.
We may also compare this 10-key code with two 9-key
codes proposed in United States Patent CITE tegic, and
EPO patent application CITE epo, respectively. The first
of these codes, afg bkn jlo mqr ew dhi sux ptv cyz has a
lookup error rate 86.5 words/lookup error, and a
64
WO 00/35091 PCT/[)S99/29343
queryrate of 3.9 words/query, while the second, rpq adf
nbz olx ewv img cyk thj su has a lookup error rate of 115
words/lookup error, and a query rate of 5.2 words/query.
These codes are both significantly inferior to the
example 10-key code which is herein designed for this
task. As neither the United States Patent CITE tegic,
nor EPO patent application CITE epo are enabling with
respect to the construction of the cited ambiguous codes,
nor are the statistics available with respect to which
these codes are optimized (if they are indeed optimized),
we can draw no further conclusions as to the substantial
optimality of these codes.
Another useful comparison is to 9-key codes which
are optimized according to the teachings of this
invention with respect to both lookup error rate and
query rate. We construct, for example, the code am bnz
cfi dhx evw gjr kos luy pqt with lodkup error rate 109
words/lookup error and query rate of 6.2 words/query.
Comparing these results, we find that the improvements
which result from the teachings of this embodiment are
from two sources 1) using more than 9 keys to permit the
improvement of lookup error rate and query rate, and 2)
optimizing both with respect to lookup error rate and
with respect to query rate. If the approach taught by
this embodiment to is extended to 11- and 12-key codes,
one finds for example the 11-key code, avy bn cl dhx ew
fip gjo kr mu qt sz with lookup error and query rates of
215 words/lookup error and 10.1 words/query
respectively, and the 12-key code aw bn cky dhq ef go ip
jr iz mx sv tu with lookup error and query rates of 313
words/lookup error and 13.2 words/query respectively.
WO 00/35091 PCTIUS99/29343
Thus, by sacrificing the use of the * and # keys for
the encoding of non-letter symbols, we can dramatically
improve lookup error rates, and substantially improve
query rates, bringing the standard-telephone-compatible
keyboards comfortably into the (level B) strongly touch-
typable range. Whether or not these improvements
outweigh the reduction of the ability to encode non-
letter symbols by using * and # keys can only be
decided in reference to the intended uses of the devices
so constructed. It is to be noted that non-letter
symbols could be encoded using sequences of encoding
symbols as described in reference to the smart-card
embodiment specified above. If the * key and # key are
available to encode non-letter symbols, then a
particularly ergonomic scheme, which respects in part the
convention of using the # symbol as an end-of-
transmission symbol is as follows. Let # encode the
space symbol=end-of-word symbol, ## encode. =end-of-
sentence symbol, and ###=end-of- transmission symbol.
In this way the complexity of encoding a symbols varies
inversely with the probability of the symbol. Depending
on the application, sequences of the * symbol could then
be used to encode other non-letter symbols such as
backspace, @ (for electronic mail applications), and/or
be used as a mode changing symbol.
Telephone keypad in alphabetic order This embodiment
presents a solution for a severely constrained keyboard-
design problem in which the number of keys is fixed, the
placement of the keys is fixed, and the ordering of
symbols on these keys is fixed. This problem arises in
the design of a keyboard which 1) preserves as well as
66
WO 00/35091 PCT/US99/29343
possible the familiar alphabetic ordering of the standard
ambiguous code, 2) is compatible with existing, standard
telephone keypads, and yet 3) has improved lookup error
rate and query rate as compared to the standard ambiguous
code. These constraints allow a limited freedom to choose
the number of keys on which to base the ambiguous code.
One can choose, for example, to allow the ambiguous code
for the letters to occupy 10 keys of the telephone
keypad, leaving the * and # keys available for encoding
non-alphabetic symbols. Also, while the standard
ambiguous code uses an even-as-possible partition, one
may choose an alternate partition and still respect the
given constraints.
Given the constraint of alphabetic ordering, each
ordered partition of 26 elements into 10 groups
corresponds to a unique ambiguous code. Given sufficient
computing time, it would be possible to evaluate the
lookup and query errors of each of these codes, and
choose the best. An alternate and more efficient
procedure is to apply the optimization methods taught by
this invention to this constrained optimization problem.
This invention teaches that a minimal elementary step in
the set of possible codes should be defined, in the
absence of information suggesting the use of some more
complex elementary step. In the current context, an
ambiguous code is a ordered list of 10 groups of letters,
such that all letters occur in exactly one group, and
within and across groups, the letters appear in
alphabetic order. An example is ab cd ef gh ij kl xnn opqr
stuv wxyz. There are thus 9 gaps separating the groups.
An elementary step consists of moving one letter across
67
WO 00/35091 PCTIUS99/29343
one gap. For instance, if we choose the second gap, we
can produce in one elementary move either the code abc d
ef gh ij kl mn opqr stuv wxyz, by moving the letter c to
the left, or the code a bcd ef gh ij kl mn opqr stuv
wxyz, by moving the letter b to the right. Given a
specified code, all possible codes which can be obtained
by one elementary move from the specified code can be
simply generated. Given this observation, and the
specification of the directed random walk method given
above, it will be evident to one skilled in the art how
to apply the optimization methods taught by this
invention in the present context. Applying this method we
find, for example, the code ab cd ef gh ijklm no pqr s tu
vwxyz, with a lookup error rate of 65 words/lookup error
and a query rate of 5.8 words/query. This code is shown
in a preferred arrangement on the keypad of a telephone
in figure 19. The error rates of this code should be
compared to those of standard ambiguous code, with a
lookup error rate of 29 words/lookup error, and a query
rate of 2.2 words/query. Thus the improvement in lookup
error rate is more than a factor of 2, and the
improvement in query rate is nearly a factor of 3, with
no sacrifice of easy-to-scan alphabetic ordering, nor
compatibility with existing telephone equipment. It is
to be appreciated that the the discussion above in
reference to the choice of 11- or 12-key encoding of
letter symbols for a telephone embodiment apply to this
embodiment as well; using optimization over partitions,
substantially optimal codes for 11 and 12 keys can be
produced.
68
WO 00/35091 PCT/US99/29343
The method of optimization over partitions is
evidently not limited to this embodiment; it could be
applied, for instance, to the smart card embodiment
previously discussed, to produce optimal codes with
alphabetic ordering on an array of 9-16 letter-symbol
bearing keys.
Qwerty-like Keyboards The approach used in the previous
embodiment to generate a keyboard which at once 1) is
compatible with a standard keyboard, and 2) is optimized
with respect to various ergonomic criteria, can be used
to produce keyboards which are 1) similar to the standard
Qwerty keyboard, and 2) are optimized with respect to
various ergonomic criteria. As in the previous embodiment
we will maintain as well as possible the layout of the
standard keyboard by maintaining the ordering of the
assignment of symbols to keys, and yet optimize over
partitions of those ordered symbols so as to minimize
lookup and query rates as well as possible. This
embodiment requires the further restriction that letters
remain in the same row of keys as given by the Qwerty
arrangement.
There exists a sequence of keyboard layouts which
are Qwerty-like in that they have three rows devoted to
letter keys, and variable numbers of columns, from one up
to 10 columns. It is clear that with but one column, that
is, but three keys, lookup and query rates must be very
high, and there is but one possible ambiguous code which
corresponds to the ordering of symbols of the Qwerty
keyboard. This code is qwertyuiop asdfghjkl zxcvbnm with
a lookup error rate of 2.8 words/lookup error, and a
69
WO 00/35091 PCT/US99/29343
query rate of 1.1 words/query, a code of such poor
quality that it is unlikely to be acceptable for any
serious use. As the number of columns increases, we are
able to find better and better ambiguous codes. At the
same time, as the number of columns increases, the device
size required to contain the keyboard, maintaining
substantially full-sized keys, increases as well. Thus
design of Qwerty-like keyboards must be a compromise
between code ambiguity and keyboard size. If, for
example, we wish to build a keyboard which is Qwerty-like
and substantially the same size as a pocket calculator
yet using full-sized keys, 7 columns can be used, as
shown in figure 20. A strongly touch-typable code,
substantially optimal with respect to lookup error and
query rates is qwe r t yu i o p as d f g hjk l zxc vb n
m with lookup error every 668 words, and query every
35.5 words, clearly strongly touch typable with respect
to a large class of typists and keyboard uses. In
figure 20 this code is shown in a preferred arrangement.
A typable device with a keyboard as described in this
figure would be suitable for note-taking, composing
electronic mail, and the like. It would be readily
typable with no or minimal learning by anyone familiar
with the standard Qwerty keyboard, and, even built with
full-sized keys, it would fit easily into a pocket.
It is to be noted that in terms of lookup and query
rates, the cost of adhering to the Qwerty convention is
high, even when the convention is adhered to but
approximately. In the code given above 17 keys are
devoted to letter symbols. If we now allow for
arbitrary assignments of letters to 17 keys, we find
WO 00/35091 PCT/US99/29343
codes such as w r t bu gi ov p af s d ej ky 1 hz cx n mq
with a lookup error rate of 7483 words/lookup error, and
a query rate of 290 words/query. This is equivalent to
one lookup error every 30 pages of typed text, and less
than one query per page of typed text. It is difficult to
imagine a use with respect to which this keyboard would
not be strongly touch typable.
With reference to figure 21, we see that this code
can be laid out in such a way that 18 letters at or very
close to their Qwerty positions, these letters are
indicated in bold face. In this arrangement, to maximize
the similarity of typing motions between typing on the
Qwerty keyboard and on this optimized Qwerty-like
keyboard, the fingers should be placed on the home row
such that the index finger of the left hand is on the
(space) key and the index finger of the right hand is on
the (ej) key. Notably by bringing both the space key and
the 'e' to the home row, this layout makes a step from
the Qwerty layout in the direction of anatomic fidelity,
in that the weight on the home row is increased relative
to Qwerty, and the weight on the strongest fingers is
increased relative to the Qwerty weight. By suitable
assignments of symbols to keys, any ambiguous code can be
brought into optimal coherence with the Qwerty (or other
conventional) keyboard.
It is to be noted that by allowing some departure
from strict Qwerty ordering, a very substantial gain has
been made in terms of functional Qwerty similarity. When
the rows are slightly displaced relative to one another,
such as they are in the standard Qwerty layout, much or
71
WO 00/35091 PCT/US99/29343
most of the finger motion required to operate the various
keys of this Qwerty-like keyboard is the same or similar
to the finger motion required to operate the standard
Qwerty keyboard. This illustrates a tradeoff between the
ergonomic criteria of preservation of a conventional
ordering, and the ergonomic criterion of preservation of
conventional function.
In view of the many competing criteria which might
be optimized, and the variety of user populations and
their needs, it would be motivated in practical
implementations of such devices to allow users the choice
between the optimized Qwerty-like keyboard and other
corresponding keyboards optimized or optimized as well
with respect to a selection of other ergonomic criteria,
such as lookup and query rates. This choice would be
facilitated if the labelling of the keys could be changed
in software. This object could be attained if each key
were equipped with a display means capable of displaying
at least one symbol at a time. Such display means could
for instance comprise an light-emitting diode array, or a
liquid-crystal display, etc.
It will be appreciated by those skilled in the art
that the keyboard design method used to prepare the
present embodiment could be applied to the preservation
or partial preservation of other conventional keyboard
designs, such as the Azerty keyboard used in France.
Numeric Keypad-like Keyboards This embodiment has the
object of making the advantages of an ambiguous keyboards
available to most computer users, with minimal cost and
72
WO 00/35091 PCT/tJS99/29343
with no change in their existing hardware. These
advantages notably include the advantage of one-hand
typability, and the advantage of potential compatibility
with ambiguous keyboards designed for hand-held devices.
Standard 101-key keyboards for workstations and personal
computers include a numeric keypad, typically to the
right of the part of the keyboard laid out in a Qwerty
arrangement. Typically, there is included a set of arrow
keys, or other input means effective for moving the
cursor, near the numeric keypad.
Referring now to figure 22 we present an ambiguous
code optimized for a common numeric keypad layout 600
taken together with the means for moving the cursor 601
where such is available. Said numeric keypad 600 has in
this example 17 keys of various sizes. Depending on other
design constraints, some or all of these keys could be
used for punctuation, or other symbols, and these other
design constraints could influence the choice of the
number of keys to be assigned to letters, the
distribution of letters and other symbols over modes and
so on. The essential features of this embodiment are:
= assignment of an ambiguous code to a plurality
of keys in the numeric keypad
= optional use of the thumb-actuatable auxiliary
input means to change modes.
It will be appreciated by those skilled in the art
that said assignment of an ambiguous code to a plurality
of keys in the ambiguous keypad can be achieved in
73
WO 00/35091 PCT/US99/29343
software; there is no need for special-purpose hardware.
However, if it is desired to have the keys thus assigned
be labelled with the elements of the ambiguous code, then
some modification of the keyboard labelling may be
required. To give a concrete example of the use of an
ambiguous code in this setting, we choose an ambiguous
code in which letters are assigned to 17 keys in such a
way as to minimize lookup error and query rate, with
respect to our standard corpus. The code illustrated in
figure 22 af bu cx d ej gi hz ky 1 mq n ov p r s t w
has a lookup error rate of 7483 words/lookup error, and a
query rate of 290 words/query. This same code has already
been discussed above. Here the code is laid out in such
a way as to preserve in part an alphabetic ordered. The
given code has not been optimized relative to alphabetic
ordering; it has only been optimized with respect to
lookup error rate and query rate. According to the
teachings of this invention, it would be possible to
simultaneously optimize with respect to lookup error
rate, query rate, alphabetic ordering, and/or other
ergonomic criteria.
To illustrate the use of the thumb-actuated
auxiliary input means for mode changing, we refer again
to figure 22. We assume for the sake of this
illustration that the auxiliary input means is comprised
of 4 keys: up 602, down 603, left 604, and right 605
arrow keys. These functionalities are typically
.implemented with 4 depressible keys, but they are
sometimes implemented as a touchpad, a joystick, or some
other device capable of generating multiple, different
signals as a function of manipulation by the user. It is
74
WO 00/35091 PCT/US99/29343
to be noted that in figure 22 a plurality of keys are
labeled with symbols other than those of the ambiguous
code, in this case digits. These other symbols can be
obtained by depressing a specified one of the four keys
of the auxiliary input means. A possible assignment of
modes to input means in the auxiliary keypad is
= 602(up) Shift key for upper case letters.
= 603 (down) numeric/punctuation mode.
= 604 (left) left symbol on key.
= 605 (right) right symbol on key.
It will be appreciated that 1) other assignments of
modes and/or symbols to the auxiliary keypad are
possible, and in accordance with the teachings of this
invention, and 2) with a more complex set of auxiliary
input means, additional modes and symbols can be assigned
to the auxiliary input means. Assignment of symbols to
modes will be discussed in more detail in reference to
another embodiment. That discussion is applicable to
other embodiments, including this embodiment.
Objects and advantages of 13-letter-key codes Several
related embodiments of the teachings of this invention
exploit the surprising benefits of optimizing
ambiguous codes for keyboards in which the number of keys
devoted to strongly correlated symbols is substantially
half of the number of strongly correlated symbols. In
particular, if we take the set of strongly correlated
WO 00/35091 PCT/US99/29343
symbols as the set of letters [a-z] used in English, then
the preferred number of keys is 13. The surprising
benefits of a 13-letter-key code for English include:
= strong touch-typability,
= ergonomic, touch-typable, unambiguous text
entry.
= ergonomic, touch-typable, querying.
= compatibility with standard keyboard layouts
(Qwerty keypad, numeric keypad, and telephone
keypad),
= providing for conservation of typing skills
from 1- to 2-handed typing,
= providing for an integrated mouse/keyboard.
= providing a mechanism to reduce typing
inj uries .
Further objects and advantages will become clear in
detailed specification below, which takes up in turn each
of the above-listed objects and advantages.
Strong touch typability Referring again to figures 11
and 12, we see that with respect to word-based
disambiguation, ambiguous codes on 13 keys can be found
which are strongly touch typable, even for an
accomplished typist. Using the directed-random-walk
76
WO 00/35091 PCT/US99/29343
method described above, we find, for instance, the code
aw bn ck du ef go hv ip js ly mx qt rz with a lookup
error rate of 515 words/lookup error, and a query rate of
21 words/query, (level C) strong touch typability. This
code is shown in a preferred arrangement in figure 25.
As discussed above, this query rate can be reduced
further by adjustment of a parameter which controls how
important queries must be to require user attention. In
the limit where querying is turned off entirely, the
lookup error rate controls touch typability. With this
code, lookup errors will occur on average approximately
once every two pages of typed text, that is to say, at a
rate much less than the rate of typing errors of even
very skilled typists. Thus, 13-key codes exist which are
suitable for a wide range of touch typing tasks, for a
wide range of users.
Ergonomic, touch-typable unambiguous text entry With
any ambiguous keyboard, and any disambiguation mechanism,
it may be advantageous to provide a mechanism by which it
is possible to enter information in an unambiguous
manner, for instance, to add information to the
disambiguation database. For all uses of ambiguous
keyboards it is advantageous for the unambiguous text
entry mechanism to be as ergonomic as possible, ergonomic
meaning in this case simple to operate. For ambiguous
keyboards meant to be touch typed, it is advantageous if,
further, the keyboard can be operated in unambiguous
text-entry mode in a touch-typable manner.
One strategy often employed in the prior art to
achieve unambiguous text entry using a small number of
77
WO 00/35091 PCT/US99/29343
keys is to use a chording method. In satisfying their
urge to maximally reduce the number of keys, chording
keyboard designers have consistently taught away from
providing a sufficient number of keys that chording
patterns can be kept simple. By an elementary
combinatorial argument, the number of keys cannot be less
than % the number of symbols to be encoded if the
complexity of a chord is never more than 2, that is, if
it is never required to active more than 2 input means
substantially simultaneously in order to unambiguously
encode a symbol. The present invention teaches, by
contrast to the prior art, to provide a number of keys no
less than % the number of symbols to be encoded, if a
simple mechanism for unambiguous encoding of those
.15 symbols is to be provided. In particular, the present
invention teaches to provide at least 13 keys to
represent the letters [a-z], and at least one mode
changing key, which when active in combination with one
of the keys encoding letters serves to uniquely and
unambiguously encoding one of the letters associated to
said.key encoding letters.
In an ambiguous keyboard, some keys represent a
plurality of symbols with a single keystroke. To use the
same keyboard in an unambiguous mode, said single
keystroke must be combined with at least one other
keystroke, perhaps of the same key, to permit each
individual symbol associated to the key to be singled
out. For ergonomics, it is preferable that said
combination of keystrokes be as simple as possible, and
for a touch typability, it is preferable that
substantially the same combination of strokes is used for
78
WO 00/35091 PCT/US99/29343
unambiguous entry of all symbols. These two criteria are
best satisfied if 1) the same number of symbols is
associated with each ambiguous key, and 2) the number of
symbols on each ambiguous key is small. Taken together,
these criteria imply that the preferred number of keys
for an ambiguous code which is to be touch typed is one-
half of the number of symbols to be represented
ambiguously. And, in the case where the symbols to be
represented are the 26 letters in the alphabet, the
criteria taken together imply that 13 keys are preferred.
One way this observation can be exploited to produce
an ergonomic, touch-typable ambiguous keyboard with an
ergonomic, touch-typable unambiguous text entry mode is
described in reference to figure 23, and in reference to
the case of ambiguous representation of the letters of
the English alphabet with 13 keys. In this figure, each
of the subset of keys used to ambiguously represent the
letters 700, encodes exactly two letters. The typable
device further comprises a mode key 701 and a means for
placing the device in either ambiguous or unambiguous
text-entry mode. The means for placing the typable
device in ambiguous or unambiguous text-entry mode could
be a software means which detects, depending on context,
which of these modes is required at any given instant, or
a key devoted to this mode change, or a particular
pattern of input on other input means, such as a double
tap on the mode-changing key 701. When in unambiguous
text-entry mode, activation of the 701 key causes one a
selected symbol from the two symbols associated with a
key in the plurality 700 to be encoded when said key in
the plurality 700 is activated substantially
79
WO 00/35091 PCT/US99/29343
simultaneously with the 701 key. It is preferable to
think of the pair of symbols on each key in the plurality
700 to be composed of a left symbol and a right symbol,
and to label the keys with the left symbol on the left
and the right symbol on the right. Then one of these,
without loss of generality, the left of these, is
associated with activation of the 701 key whereby
unambiguous text entry can be achieved. When a key in the
plurality 700 is activated in combination with the 701
key, the left symbol is selected unambiguously, and if
the same key in the plurality 700 is activated without
the 701 key being substantially simultaneously activated,
then the right symbol is selected unambiguously. If the
keyboard is designed to incorporate additional modes,
then this same method of unambiguous text entry could be
used also in reference to the symbols in the other modes.
Touch-typing oriented querying Even with an optimal
ambiguous code, coupled to unlimited computing power and
ultimate, yet-to-be discovered artificial-intelligence
techniques for disambiguation, some ambiguous sequences
may be generated in the course of typing a text which
require intervention of a human operator to effect full
disambiguation.
True touch-typists type without looking at the
keyboard, keeping their eyes focused on the text being
produced, or the copy being transcribed. For the touch
typists, then, it is preferable to have all querying for
alternate interpretations of ambiguous sequences done in
such a way that 1) their eyes are not diverted from the
WO 00/35091 PCT/US99/29343
text-display screen and 2) queries can be answered in a
simple, stereotypical manner from the keyboard. These
objects can be accomplished by reserving a key for
advancing in a list of candidate words, and highlighting
the ambiguous word on the screen. Users scroll through
the list of candidate words using this scroll key, and as
soon as a key other than the "scroll" key is pressed, the
word in the scroll box is considered selected.
With reference to 23 and 24, we explain in detail
the software underlying touch-typing-oriented querying,
and the visual display it controls. In the first step 800
a query has been detected, that is, the disambiguation
mechanism has discovered that more than one meaningful
decoding sequence in the database corresponds to the
input encoding sequence. Entrance in the query mode
causes a means to draw the users attention to the
decoding under query to be display. These means could be
visual means such as the framebox 702 shown in figure 23.
The possible decodings are then ordered according to
their likelihood (step 802). Then, (step 804) the most-
likely decoding is displayed on the screen, in context
with the text previously entered, in a place indicated by
the attention means. The software is then prepared to
detect either input from the scroll key, or some other
key (step 806). If some input is received from some other
key, then the means to draw attention to the user is
removed (step 808), the decoding is added to the text
previously entered (step 810), and ambiguous text entry
mode is reentered.
81
WO 00/35091 PCT/US99/29343.
On the other hand, if input from the scroll key is
detected at step 806, then a test is performed to see if
there exists any other meaningful decodings in the data
base 812. If there is, the current decoding is replaced
with the next most probable decoding (step 814), and step
806 is returned to. If there is no next-most-probable
decoding, then, preferably, the typable device enters
unambiguous text entry mode as described above (step
816), when the decoding sequence has been unambiguously
entered, the means to draw attention to the user is
removed (step 808), the unambiguously entered decoding is
added to the existing text (step 810) and ambiguous text
entry mode is reentered (step 818).
While this method for in situ presentation of
alternatives has been presented in terms of touch-typing-
oriented querying, it will be appreciated that the same
method could be applied in other contexts, such as when
the "decodings" are words with related meanings, and the
database is a thesaurus, or when the "decodings" are
various possible translations of a word into a foreign
language, and the probabilities of the decodings are
supplied by an automatic translation program.
Conservation of design across platforms as an ergonomic
criterion: a mouse/keyboard Consider a user of a small
typable device equipped with a one-handed touch-typable
keyboard, such as a personal digital assistant. In the
course of a typical day, the user might also type on his
or her desktop computer equipped with a two-handed
keyboard. If both of these devices are to be effectively
touch typable for that user, the motor patterns used in
82
WO 00/35091 PCT/US99/29343
the one- and two-handed keyboards must be as similar as
possible. The shorter the interval between the uses of
the two keyboards, the stronger the requirement for
conservation of typing skills. To sharply draw the object
of conservation of typing skills across platforms, and
the means by which this invention achieves this object,
we now turn to an embodiment in which 1- and 2-handed
keyboards may be used in rapid alteration.
This embodiment concerns a one-handed keyboard
suitable for entry of data into forms, such as
spreadsheets or web-based forms. It will be of use for
interaction with a program, such as a game or a drawing
program, 1) which requires quick alternation of typing
and cursor movement, and/or 2) where the symbol set which
is appropriate for entry may vary depending on where the
cursor is on the screen. Using the standard computing
configuration of a Qwerty keyboard and a mouse, users
must remove their hands from the keyboard to move the
mouse. In tasks which involve both typing and mouse
manipulation in rapid succession, such as labeling a
design presented on the computer screen, or filling in a
form, such as an HTML form, this alternating use of mouse
and keyboard can be quite slow and laborious. In this
embodiment, a one-handed keyboard is mounted in a frame
which may be moved over a surface and thus perform the
functions of a mouse as well as a keyboard. Since users
may prefer to use a two-handed keyboard for predominantly
text-entry tasks, and intermix usage of the one-handed
keyboard with use of the one-handed keyboard, it is
desirable to have the layout of keys on both keyboards be
as similar as possible, so as to permit seamless transfer
83
WO 00/35091 PCT/US99/29343
of touch-typing skills from the two-handed to the one-
handed keyboard.
Referring now to figures 25, 28 and 26, we see how
this object can be attained by the choice of an ambiguous
code which is such that it can be laid out on the one-
handed keyboard so that the movements of the fingers and
the thumb of one hand (in this case the right hand) are
the same, whether that hand is typing on the one-handed
or two-handed keyboard, and that, further, the typing
movements of the other hand are similar to the movements
of the hand chosen for one handed typing.
The design strategy is as follows:
= Choose a 13-key code with substantially minimal
lookup error rate and query rate.
= Choose a physical layout for the 13 keys. A
layout with 5 keys in the top row, 5 keys in
the middle (home) row, and 3 keys in the bottom
row is a preferred arrangement.
= Choose which hand will actuate the one-handed
keyboard.
= Arrange the keys such that, in the one-handed
layout, and relative to the hand chosen in the
previous step: where
- weight on the home row is maximized.
- weight on the strongest fingers is
maximized.
84
WO 00/35091 PCT/US99/29343
- weight on the top row is higher then
weight on the bottom row.
= Then, to obtain a layout for the two-handed
keyboard, pair each key actuated with the left-
S hand to a key actuated by the right hand.
Preferably, the pairing is done so that members
of the pair are laid out symmetrically, with
respect to a symmetry plane which cuts the
keyboard at the center, and runs from the
bottom to the top of the keyboard.
= Of the two symbols associated to each of the 13
original keys, associate one symbol to one of
the keys in each pair chosen in the previous
step. This can be done in a way favoring the
one-handed keyboard, or in a way favoring the
related two-handed keyboard.
- If favoring the one-handed keyboard:
retain the higher-probability letter
on the chosen-hand side of the two-
handed keyboard, and place the lower-
probability letter on the opposite
side of the keyboard.
- If favoring the two-handed keyboard:
for each of the 13 keys, choose to
place either the lower or the higher
probability letter on a given side of
the two-handed keyboard so that the
WO 00/35091 PCT/1.1S99/29343
summed probability on each half of
the keyboard is as equal as possible.
Beginning with the 13-key code selected to
illustrate the teachings of this embodiment, and choosing
to favor the right hand, one constructs the keyboard
layout described in figure 25. When this one-handed
keyboard is expressed as a two-handed keyboard, the
resulting layout is shown in figure 26. On this
keyboard, the right hand types approximately 84 percent
of the letters, while the left hand types approximately
16 percent of the letters. This asymmetry is desirable
in that the large majority of keystrokes will be
performed in exactly the same manner whether the one-
handed or the two-handed keyboard is used.
Conversely, if most typing is done with a two-handed
keyboard, and only occasionally with a one-handed
keyboard, it may be desirable to have as even as possible
weight on the two hands when used to operate the two-
handed keyboard. This object can be achieved with an
alternate layout of the two-handed keyboard as shown in
figure 27. It will be appreciated that substantially 50
percent of the typing motions used in typing on the two-
handed version of this.keyboard are the same as the
typing motions used on the one-handed keyboard, whether
the one-handed keyboard is typed with the right or the
left hand.
It will be appreciated that for a 13-key ambiguous
code there are 213 different ways of pairing homologous
left- and right-hand keys of a two-handed keyboard. This
86
WO 00/35091 PCT/US99/29343
number is small enough that the symmetry of the weights
assigned to the two hands in each set of pairings can be
evaluated, and depending on whether a most-symmetric or
most-asymmetric weighting is desired, or some
intermediate value, the appropriate assignment can be
selected.
Referring now to figure 28, a detailed view of the
one-handed keyboard, we see how the objects of this
embodiment are achieved by equipping the keyboard with a
plurality of keys 300, a thumb-actuatable input means
301, mouse keys 302, a palm grip 303, and a display 304.
The keyboard may be further equipped with a communication
means to enable symbol selections by means of the
keyboard to be communicated to the computer. This
communication means could be simply a wire, or a wireless
communication means such as an infrared communication
means. The keyboard is slidably supported on a support
means, such as a desktop, whereby the keyboard can be
moved over the support means by means of pressure from
the base of the palm of the hand actuating the keyboard.
The keyboard is preferably equipped with a means for palm
grip means for engaging said base of the palm to allow
said pressure to be effective in moving the keyboard.
Said means for engaging said base of palm of the hand
operating the keyboard are preferably formed means where
the form is such that slight pressure is effective to
move the keyboard in any desired direction. For instance,
the form could be an indentation in the body of the
keyboard which securely engages the base of the palm. By
moving the keyboard in this way, the fingers of the hand
actuating the keyboard are free to move in a way
87
WO 00/35091 PCT/US99/29343
effective to actuate the keys, even while the keyboard is
in motion. Thus this keyboard can be used in situations,
such as for playing computer games, in which motion needs
to be simultaneously input along with symbol sequences
such as text.
It is to be noted that while this device performs
the functions of a mouse, it bears little physical
similarity to a mouse. Its form factor is determined by
the anatomy of the hand in a comfortable position for
touch typing. The device must therefore be considerably
larger than a standard mouse, and the means for moving
the device substantially different.
To communicate motion of the keyboard 305 to the
computer, said keyboard is equipped with a motion sensing
means, such as a trackball, familiar to those skilled in
the art. Preferably, the keyboard 305 can be further
equipped with a biasing means, such as springs, to lift
the keyboard away from the support means when the weight
of the hand on the keyboard is lessened, thus
facilitating movement. In contradistinction, when the
substantially full weight of the hand is applied to the
keyboard, the keyboard remains relatively stably fixed to
the support means whereby typing is facilitated.
The two-handed keyboard may thereby be used for long
typing episodes, uninterrupted by the need to move a
mouse, and the one-handed keyboard for quick typing/mouse
movement alternation.
88
WO 00/35091 PCT/US99/29343
Visual Representation of the Keyboard It will be
appreciated that when using a single device which
supports more than one ambiguous codes, or more than one
mode, it may be helpful to the user to have a
representation of the current association of keys with
symbols at any given moment displayed on the displayed
device. While such a representation could be of value
for any typable device, it is of particular utility for
touch-typable devices, since for such devices, some or
all of the keys are hidden from view (typically by the
fingers of the operator) while in use. Thus, any display
means integrated into the keys is of limited utility for
the touch typist. The most useful visual representations
are those in which the physical layout of the keyboard is
represented in the visual display. Such a device 304 is
shown in figure 28, but could be incorporated in many of
the embodiments described in this disclosure.
Reduction of typing injury Typing injuries (repetitive
stress syndrome) afflict many keyboard users. Numerous
keyboards have been designed in an attempt to attenuate
the stress of the repetitive motions involved in typing.
It has long been recognized that the most effective means
to reduce typing injuries is for the typist to take
regular breaks from typing. This is seldom practical,
however, as typists are often under time pressure to
complete their typing work. The one-handed keyboard just
described offers a solution to this problem. While the
one-handed version of the present embodiment has been
described as being used with the right hand, it is
evident that the same design methods lead to a left-
handed one-hand keyboard as well. Each of these
89
WO 00/35091 PCT/US99/29343
keyboards is capable of encoding all of the same symbols.
A therapeutic typable device equipped with both a right-
handed and a left-handed keyboard, such as both a left-
and a right-handed mouse/keyboard as described above,
could be operated with either the left or the right hand
in alteration. With such a pair of keyboards, a user
wishing to reduce repetitive-stress injuries could type
for some period of time, for instance, for 15 minutes,
using one of the keyboards, and then switch for the next
period of time to the other keyboard, whereby the user
gives each hand a resting period, with no decrement to
typing productivity. If desired, the typable device could
be equipped with a locking means which would alternately
lock one or the other of the keyboards, enforcing
alternating use. It is to be appreciated that when
therapy is completed, the user could return to a two-
handed version of the keyboard, with no relearning of
typing skills required.
Foldable PDA We saw in the smart-card embodiment
described above that it is convenient and ergonomic to
place the text screen of a typable device based on
ambiguous codes in such a way that part of the keyboard
is manipulated by the fingers of one or both hands and
the thumb may be used to actuate further input means, in
particular, mode-changing input means, in the region
above the thumb or thumbs and to the side of the finger-
manipulated part of the keyboard. The present embodiment
concerns a typable device at a somewhat larger scale
which uses the same concept in conjunction with a folding
concept in order to design a twice-foldable information
appliance which may ergonomically perform different
WO 00/35091 PCT/US99/29343
functions when it is unfolded, once folded, and twice
folded.
This twice-foldable design is a surprising
consequence of ambiguous codes. It is remarkable that
typable devices built by the methods taught by the
present invention permit keyboards which are at once 1)
effective for the coding of natural language, 2) use
substantially full-size keys, and 3) are small enough to
be placed in a pocket or a small hand bag. This
embodiment is based on building a hand-held computing
device from substantially same-sized elementary units
which are the size of a keyboard designed for an
ambiguous code, said units being configurable in a
variety of ways depending on the instant needs of a user.
1S The elementary units can be foldably and/or detachably
connected to each other in each of a variety of
configurations. In this way the computing device can
alternately play the role of a laptop computer, a
personal digital assistant, a telephone, a gaming device,
and so on.
With reference first to figure 29 we specify in
detail a twice-foldable computer built from four
substantially same-sized parts, each of which is adapted
to perform a specified function and are connected to each
other in a foldable and/or detachable way. Figure 29
shows such a device in the unfolded state. It is thus
revealed that one of the parts 900 has a first surface
which functions as a first visual display, while 901 has
a first surface which functions as a first keyboard and
902 has a first surface which functions as a second
91
WO 00/35091 PCT/US99/29343
keyboard. Preferably, the keyboard layout on said first
and second keyboards is a 13-letter-key keyboard, though
many other choices are possible. The final part 903 has a
first surface which functions as a pair of mode-changing
thumb switches, to be used in combination with the said
first and second keyboards. The first keyboard is meant
to be operated with the right hand. It will be evident
to those skilled in the art that a similar configuration
exists which is typable with the left hand, and that this
similar configuration can be obtained by simple
rearrangement and reattachment of the four parts. Indeed,
a two-handed keyboard can be obtained from by detachment
and rearrangement of the four units as shown in
figure 33.
Figure 30 shows the unfolded twice-foldable computer
in a bottom view. Part 904 is a telephone keypad 905 and
corresponding second visual display 906. Part 907 is a
third visual display, and part 908 a third keyboard.
Parts 904, 905, 906, 907 form the second surfaces of
parts 900, 901, 902, 903 respectively.
Folding the computer along the line 908 shown in
figures 29 and 30 we obtain the configuration shown in
figure 31. In this configuration the third keyboard is
exposed for typing, and the third visual display is used
as the corresponding display. Here the keyboard layout
is a 12-key keyboard, but many other choices are
possible. This configuration might be used when, due to
time or space limitations, the user is unable or
unwilling to open the computer to its full extent. Tt
might also be used to supply a different functionality
92
WO 00/35091 PCT/US99/29343
than the fully unfolded computer, such as a gaming
functionality.
Finally, the computer can be folded along the fold
909 shown in figure 31 to provide the twice-folded
configuration shown in figure 32. This is the
configuration in which the computer would be typically
kept for transportation; in this configuration the device
could be small enough to fit into a pocket. Further, in
this twice-folded configuration the telephone
functionality is exposed for use. For many users, this
will be the most frequently used configuration of the
device. It will be noted that we have shown the telephone
keypad with the ambiguous code of a previous embodiment,
though many other choices are possible.
To recapitulate: thanks to ambiguous codes, we can
design a portable communication and computing device
which functions alternately as a telephone, a personal
digital assistant, and a laptop computer.
It will be appreciated by those skilled in the art
that if each of the elementary units, including the
keyboard units, are built from touch-screens, then the
variety of configurations and uses of this device could
be further increased. What would be lost, however, is the
tactile feedback from the keyboards built from standard,
depressable keys. Many other variants are possible,
consistent with the teachings of the present invention.
Software embodiment for a typable device comprising a
touchscreen This invention permits software as well as
93
WO 00/35091 PCT/US99/29343
hardware embodiments. In particular, the methods of the
present invention can be used to design typable
mechanisms for devices comprising touch screens, such as
the series of personal digital assistants built by the
3Com corporation and sold under the trademark PALM PILOT,
as well as other trademarks, For the sake of
illustration, we will focus on the PALM PILOT class of
devices, which includes hand-held computers capable or
running a variety of applications programs, though the
methods herein described could be applied to any typable
device comprising a touch screen.
In reference to figure 34, we note that devices of
the PALM PILOT class typically comprise: a touch screen
1000, a touch-sensitive region 1001 which is used for
entering characters via hand-writing recognition
software. Said touch-sensitive region may be a subregion
of the touch screen, or may be implemented separately.
One of the essential and surprising features of the
present embodiment is that by use of a touch-typable
keyboard in a device comprising a touch screen, we design
a radically new user interface for information appliances
such that keyboards need not compete for limited screen
space with applications programs. The same touch screen
area can be used both for the application program and for
the keyboard.
The crucial observation is that if the keyboard is
weakly touch typable, then the keyboard does not need to
be displayed to the user to be operative. The user's
fingers "know" where the keys are, without visual
referents. Thus, the keyboard can be used to enter data
94
CA 02353862 2001-06-08
WO 00/35091 PCT/US99/29343
into which ever application program is currently
displayed on the touchscreen. If, further, the keyboard
is strongly touch typable, then the keyboard can be used
to produce high quality text, even if no screen space is
used for querying feedback to the user.
With reference to figures 34 and 35, some of the
particular features of the PALM PILOT class of devices
which are accounted for in the present embodiment are:
= The ability of the touch screen 1000 to easily
present alternate keyboard layouts.
= The ability of a touch screen to present images
of different levels of intensity and/or in
different colors.
= The placement of the region for entering
characters 1001 at some remove from the touch
screen, or in a non-central region of the touch
screen.
= The use of the personal digital assistant to
run a variety of programs, such as scheduling
programs or address-book programs, which may
compete for space on the touch screen with a
keyboard.
The ability of touch screens to easily present
alternate keyboard layouts is used in this embodiment to
enable a given input means to represent many different
symbols or groups of symbols, depending on the "mode" of
WO 00/35091 PCT/US99/29343
the keyboard at any given time. When a touch screen is
used to embody a keyboard, each input means is associated
with a specified region of the touch screen. The dual
function of the touch screen as a visual display and as a
plurality of mechanically activated input means is
exploited to give each input means different functions
and different labels depending on mode. It will be
appreciated, however, that the same effect could be
obtained with mechanical keys of a traditional,
depressable structure by equipping each mechanical key
with its own display device. In this way, the methods
for mode changing here specified in reference to a
touchscreen comprising device could be applied to devices
comprising mechanical keys, such as many of the other
devices specified in the present disclosure.
Mode Selection One strategy in keyboard design to
increase the number of symbols which can be encoded given
a fixed number of keys is to augment the keyboard with a
number of mode-changing keys. Pressing a mode-changing
key changes the symbol encoded by a plurality of other
keys. The canonical example is the shift key of the
standard typewriter keyboard which changes the symbol
encoded by the letter keys from lower case to upper case.
Upper-case letters could in principle be encoded on a
separate set of keys from the keys encoding lower-case
letters, and if upper-case letters occurred with the
same frequency as lower-case letters in typical
communications, then this could be a defensible choice.
Also in principle, having upper- and lower-case letters
accessible in different modes does not imply that an
upper-case letter must be assigned to the same key as the
96
WO 00/35091 PCT/US99/29343
corresponding lower-case letter. The same-key assignment
is chosen in practice because there are strong
conventional, conceptual and statistical relationships
between lower- and upper-case letters.
Thus there are three principles which guide the
assignment of symbols to modes and of symbols to keys
within modes: adherence to statistical,relationships,
adherence to conventional relationships, and adherence to
conceptual relationships between symbols. In the design
of typable devices employing ambiguous codes the problems
associated with the design of modes are particularly
acute since a plurality of keys must already carry the
burden of encoding more than one letter symbol on each
key, and the number of keys available to encode symbols
is typically sharply limited. However, the same methods
which have been above applied to produce ambiguous codes
for letter symbols can also be applied to non-letter
symbols such as punctuation, provided that these non-
letter symbols are strongly correlated.
The collection of symbols to be encoded by a
keyboard will be divided into subsets corresponding to
modes. Modes can be at least partially ordered according
to how much manipulation is required on the part of the
user to obtain each mode and/or how frequently the
symbols in each mode are used. Thus we can speak of
primary, secondary, tertiary modes and so on in order of
increasing amounts of manipulation required to obtain
each mode and/or decreasing probability of the symbols in
the mode.
97
WO 00/35091 PCT/US99/29343
Letter symbols are preferably placed in the first
mode or modes. The subtle design issues have to do with
methods for assigning non-letter symbols to modes, and
arranging the spatial layout of each mode.
A first statistical measure to be taken into account
is the probability of non-letter symbols. Some of these
non-letter symbols, such as punctuation marks and digits,
are essential to communication and may occur with a
frequency rivalling or exceeding those of letter symbols.
These punctuation marks are candidates for inclusion in
the primary or secondary symbol set in any effective
keyboard design. Next to be considered are correlations
which arise from non-letter acting in concert with other
non-letter symbols. Some non-letter symbols have
conventional and conceptual relationships with other non-
letter symbols, for example, the symbol (left arenthesis)
is related to the symbol) (right parenthesis), as the two
symbols act together to express meaning. The symbol
is related to the symbol , as the two symbols express
similar meanings (end of a phrase or sentence). These
are examples of global relationships, common to most uses
of the language, in these cases including languages such
as English. There are other, more local, relationships
which could be accounted for in keyboard designs for
specialized purposes, such as the relationship between
and / in the expression :/ commonly used in site
addresses on the World Wide Web (URLs or Universal
Resource Locators).
Non-letter symbols may also have statistical,
conventional, and conceptual relationships with letter
98
WO 00/35091 PCT/US99/29343
symbols. For some symbols, it is possible to analyze
their statistical relationships with each other with
reference to a corpus. For others, user studies are
needed, or specialized software to capture these symbols
for statistical analysis, since the symbol may never
appear in a text. Examples here are "backspace", "page
up", and other symbols which are used to edit, examine,
or otherwise manipulate a text.
Given reference statistics constructed in this way,
a next step in the assignment of symbols to modes is to
arrange the symbols in such a way that the statistical,
conventional, and conceptual relationships are best
satisfied. A further constraint which could be taken into
account is the mnemonic potential of the arrangement.
Preferably, all symbols are arranged within and across
modes in a way that "makes sense", that is, that the
symbol pattern is simple, familiar, preferably visually
well structured. Even well-trained touch typists may
revert to a visual scanning mode to find infrequently
used symbols on a keyboard. Thus, mnemonic potential may
be the overriding concern in the arrangement of modes
devoted to less-frequently used symbols. It is to be
.appreciated that mnemonic potential can be quantified
using experimental protocols for memorization tasks well-
known to psychologists.
To illustrate the approach, an example layout of
letter symbols [a-z], digits, and the 32 nonletter
symbols
99
WO 00/35091 PCT/US99/29343
found on a standard keyboard is provided. This
arrangement comprises three mode changing keys, and 3
modes, each mode containing 16 symbol keys, as shown in
figure 36A-C. This layout is designed for PALM PILOT
class machines. It has not been shown to be optimal by
psychological testing.
The first mode (figure 36A) contains an ambiguous
code for the letters, a space/backspace key, a basic
punctuation key, and a key for shifting modes in either
the forward or the backward directions.
The second mode (figure 36B) contains keys for
the digits, and certain punctuation marks, arranged so
that this mode can function either as a telephone, or as
a rudimentary calculator/numeric keypad.
The third mode (figure 36C) contains additional
punctuation marks, arranged so that 1) a shift key
relates symbols with related meanings, such as open and
closed parentheses, or, if there is no related meaning,
related symbol shapes, as an aid to remember the symbol
placements. A convention is applied whereby "hard"
symbols, more angular symbols, are on the left, while
"soft" symbols, more curved symbols, are on the right.
Since all or most keys have exactly two symbols,
ergonomic disambiguation mechanisms as described above
can operate in each mode.
Selectable transparency of the keyboard Currently,
typically implementations of a keyboard displayed on a
touch-screen device in the PALM PILOT class comprise a
100
WO 00/35091 PCT/US99/29343
keyboard occupying one part of the screen, while the rest
of the screen is devoted to an application program which
receives input from the keyboard, such as an address book
application. Since display real estate in such devices
is extremely limited, sharing the display between
keyboard and application program results in both keyboard
and application program being very small. Keyboards used
in present devices are not meant to be practically touch
typable, nor could they be, given their extremely small
size. However, application of the methods of this
invention produces keyboards which are small enough to be
touch typable, even in the limited environment of the
touch screen of a personal digital assistant. The
important observation is that since the keyboard is
touch-typable, it does not have to be displayed to the
user. The user's fingers "know" where the keys are,
without the user having to see them. Thus the keyboard
can be made transparent, occupying the entire touch
screen, while the application program can be opaque, and
also occupy the entire touch screen. Displayed in this
way, the user types directly on the application program
to produce input to it. In figure 35 a keyboard displayed
in this way 1003 is shown with an application program
1002 in this case a drawing program displaying a drawing.
Since it is impossible to draw a transparent keyboard,
the keyboard is indicated in this figure in a shade of
gray, while the application program is shown in black.
Indeed, it would be possible to allow the user to select
the level of transparency of the keyboard, for example,
as a function of his or her level of touch typing skill.
It has already been pointed out that different modes
101
WO 00/35091 PCT/US99/29343
may contain symbols of different levels of familiarity
to the user. To account for these differences, the
transparency of the keyboard could also be adjusted as a
function of mode, becoming increasingly less transparent
as the unfamiliarity of the symbols in the mode
increases. It is to be noted that the same effect of
distinguishing the keyboard for the applications program
could be accomplished by adjusting other visual factors,
such as the color of the image, in addition to its
transparency.
Hybrid Chording/Ambiguous Keyboards The important
observation on which this aspect of the present invention
is based is that chording patterns which require but two
keys to be activated substantially simultaneously, such
as the chording pattern which causes a Qwerty keyboard to
encode capital letters, are readily learnable and can
therefore be adopted by a large user community. However,
the prior art has proven that chording patterns any more
complex than this will not become generally accepted.
It has been noted that there are two main prior-art
approaches toward typable devices with a small number of
keys: chording methods and ambiguous-code methods. One
aspect of the present invention is to teach how to
synergistically combine these two methods.
We can distinguish two kinds of chording methods 1)
methods in which a key or keys are reserved for the
function of forming chords, an example is the familiar
shift key used for keying uppercase letters by a chording
combination of the shift key with a letter key, and we
102
WO 00/35091 PCT/US99/29343
will generally refer to such keys as shift keys, 2)
methods in which chords are formed by activating a
plurality of letter keys substantially simultaneously. In
the present embodiment the first of these methods is
used, in a subsequent embodiment, the second of
these methods is used.
The essential insight of this aspect of this
invention is that substantially simultaneous activation
of a pair of input means is readily unified into a single
gesture by a human user. Thus, a pair of keystrokes is no
or little more difficult to master than a single
keystroke, yet, a pair of keystrokes contains
substantially more information than a single keystroke,
and thus can be used to create easily operable, low-
ambiguity codes and typable devices based on these codes.
Thus, to make the keyboard easy to learn, chords must
require no more than a pair of keys to be activated
substantially simultaneously. In this embodiment, one of
the pair being a key reserved for forming chords, and the
other a key corresponding to at least one decoding
symbol. Rare symbols may still require more than two keys
to be activated substantially simultaneously, and very
frequent symbols could be associated with a single input
means without exceeding the scope of this invention.
It is preferred that at least one of the decoding
symbols is a strongly correlated symbol. In this way,
should a chord be formed improperly, when a shift key is
activated when it should not be or not activated when it
should be, the disambiguation software may be able to
correct for this error.
103
WO 00/35091 PCT/US99/29343
Thus, while chording and ambiguous codes could a
priori be combined in any number of ways, according to
the teachings of this embodiment, preferred combinations
are such that
= no more than 2 input means need be
substantially simultaneously activated in order
to encode any substantially probable symbol.
= lookup error and/or query error is optimized,
= chording is accomplished using a mode-shift
key,
= (preferably) such that the probability of using
the mode-shift key is minimized.
When chording and ambiguous-code methods are
combined according to these teachings, surprising and
synergistic results are obtained as will be
demonstrated by the present embodiment in which a hybrid
chording/ambiguous-code method is applied to a telephone
embodying the standard ambiguous code. We have already
seen that the standard ambiguous code has rather poor
lookup error and query rate. It is therefore quite
extraordinary that using a hybrid method a keyboard based
on the standard ambiguous code can be made (level C)
strongly touch typable. The objects of this embodiment
are to produce a keyboard which is
= strongly touch typable, fully compatible with
standard telephones embodying the standard
ambiguous code,
104
WO 00/35091 PCT/US99/29343
= simple to operate,
= simple to learn,
= and uses a minimal number of keying gestures.
A standard telephone embodying the standard
ambiguous code is shown in figure 38. It is seen that a
plurality of keys 10000 are used to encode letters and
digits, there are eight of these. Two keys 10001, 10002
encode only digits, and two keys 10003,10004 encode the
non-letter symbols * and # respectively. In this
embodiment, one of the keys selected from the group
consisting of 10001, 10002, 10003, 10004 will be used as
a mode-changing key, preferable key 10001 encoding the
digit 1. This selected key will be referred to as the
shift key, for reasons which will become evident. Key
10001 can be conveniently activated by the thumb of the
left hand while the telephone is held in the left hand,
while the right hand is used to activate the other keys.
For an embodiment in which the right thumb is used to
active the shift key while the telephone is held in the
right hand, key 10004 can be used as the shift key.
For each of the keys in the plurality 10000 the
corresponding letters will be divided into two subsets,
which will be referred to as the shift set and the non-
shift set respectively. Letters will be assigned to the
(shift, nonshift) sets so that
= lookup error is minimized,
105
WO 00/35091 PCT/US99/29343
= queries are minimized,
= one of the sets, without loss of generality,
the shift set, contains exactly one letter per
key,
= the probability of activating the shift key is
minimized.
As is usually the case, simultaneous optimization
with respect to lookup error rate, query rate, and other
ergonomic criteria implies compromise. For instance, one
could potentially achieve better lookup error rates and
query rates by removing the restriction that one of the
(shift, nonshift) sets is composed of a single letter per
key and allowing the number of letters in the shift set
to vary from key to key. However, regularity of the
partition into shift and nonshift sets makes the keyboard
easier to learn, an ergonomic criterion which is here,
given high priority. Learnability could be further
improved by selecting the singletons such that the
collection of singletons is easy to remember, perhaps by
mnemonic. This selection, however, could compromise
lookup error rates and query rates.
There are 11664 different pairs of shift/non-shift
sets obeying the constraint that the shift set contains
but one letter per key. This number is small enough that
all possibilities can be tested for their ergonomic
.properties.
106
WO 00/35091 PCT/US99/29343
The results of testing all such codes are shown in
figure 39, where lookup error rate is plotted vs. query
rate, for all 11664 codes, as well as the standard
ambiguous code SAC. It is to be noted that while all of
the codes are better than the standard ambiguous code,
most are better by a small multiple. However, this
distribution is quite large, and the best code CEHLNSTY
with a lookup error rate of 431 words/lookup error and
query rate of 21 words/query is 15 times better than the
standard ambiguous code in terms of lookup error and 10
times better in terms of query error. This best code is
abC dEf gHi jkL mNo pqrS Tuv wxYz where the
elements of the shift set are written with uppercase
letters. It is to be emphasized again that this code is
the best code with respect to our reference statistics,
other statistics may yield other best codes, though it
appears to be among the best for statistics drawn from
many alternate corpora of English. It should be further
appreciated that the same hybrid chording/ambiguous code
method could be applied to arbitrary ambiguous codes in
which the underlying ambiguous code is not constrained to
be the standard ambiguous code, indeed is not constrained
to be in alphabetic ordering, or only 8 keys, or with an
even-as-possible partition. By allowing more freedom to
choose codes, hybrid chording/ambiguous codes can be
found of such high quality in terms of lookup error rate
and query rate that other ergonomic criteria, such as
minimizing use of the shift key, could be profitably
combined with optimization of lookup error rate and query
rate. These optimizations are beyond the scope of the
present embodiment, which seeks full compatibility with
107
WO 00/35091 PCT/US99/29343
existing telephones, they are however, well within the
scope of the present invention.
To further facilitate the learnability and
operability of this keyboard, the letters which form the
shift set are preferably represented on the corresponding
key using an uppercase letter, while the letters in the
nonshift set are represented as a lowercase letter, as
shown in figure 38. Alternately, the two sets could be
indicated by lettering varying in size, color, typeface
etc.
Use of this device is simple. When the text to be
typed contains a letter in the shift set, then the shift
key must be operated substantially simultaneously with
the corresponding letter key, this letter will be
represented unambiguously. By contrast, when a letter in
the nonshift set is required, the corresponding letter
key is operated, and the letter is represented
ambiguously.
In view of the teachings of this invention, it will
be appreciated that the disambiguation mechanism
corresponding to this embodiment could be physically
located within the telephone, at the sending end of the
communication, and or at the receiving end of the
communication, for instance at a central computer which
the user contacts by telephone.
As we have seen, on the standard telephone keypad
there are 4 keys available to encode non-letter
information such as mode changes. Using the method of
108
WO 00/35091 PCT/US99/29343
synthesized encoding symbols described above, the number
of non-letter symbols encodable using a telephone keypad
can be further increased. In particular, added shift keys
can be provided if even lower levels of ambiguity are
required. For instance, with 4 shift keys, each
associated to one letter on each letter key, totally
disambiguous text entry can be achieved, albeit while
increasing the number of keystrokes per letter. In short,
many associations of subsets of decoding symbols with
shift keys can be imagined. One particular class of
assignments is considered in a later section on
internationalization of these teachings, up to now
described in detail mainly with respect to English.
It will be appreciated that using the shift key in
combination with the remaining non-letter keys, in the
preferred arrangement the # and 0 keys, can be used
to encode at least 6 non-letter symbols, such as
punctuation symbols, mode-shift symbols, and the like.
Error correction using the standard ambiguous code
Especially when used by a novice user, the keyboard of
the present embodiment could be operated in such a manner
that the shift key is pressed at times when it should not
be to encode the intended text, and at other times the
shift key will not be pressed when it should be. Often,
such a manipulation will result in a meaningless decoding
if the disambiguation device is expecting correctly typed
encoding sequences in the hybrid chording/ambiguous code.
In these cases, rather than issuing a query, an alternate
disambiguation can be attempted in which the shift key
activation is ignored, and the encoding sequence is
109
WO 00/35091 PCT/US99/29343
interpreted as being an encoding sequence in the standard
ambiguous code. Often, this interpretation will recover
the text intended by the user.
It is to be noted that in the device shown in
figure 38 a strongly correlated symbol (space)
is paired on the same key with a weakly or uncorrelated
symbol (backspace). This pairing potentially allows
disamiguation software to correct errors in which (space)
is meant but (backspace) keyed by the user, or vice
versa.
Touch-typing-oriented querying It is preferable in this
embodiment to use the selected shift key as the scroll
key for querying, when querying is permitted, as
described above. It will be appreciated that whether the
key functions as a shift key or a scroll key at any given
moment can be determined automatically given the
appropriate software. When the device is in querying
mode the key functions as a scroll key, and otherwise
functions as a shift key.
Alternate placement of shift keys Referring again to
figure 38, we note that this embodiment has been designed
to be operable using existing, standard, telephones. If
telephones are manufactured with use of this embodiment
in mind, they are preferably equipped with additional key
or keys 1005 to function as the shift key. It is
preferred to place these additional keys on the sides of
the telephone, where they can be actuated by the thumb of
the hand holding the telephone; such a placement is shown
in figure 38. Additional keys to be actuated by the
110
WO 00/35091 PCT/US99/29343
fingers of the hand holding the telephone (either the
left or the right hand) may also be used 10006.
Culling infrequent words in queries It is often the case
that a very frequent word is ambiguous with a very
infrequent word. For instance, in the case of CEHLNSTY
for English, the very frequent word "for" is ambiguous
with the very infrequent word "fop". By eliminating
these very infrequent words from the dictionary, the
effective query rate can be improved with very little
effect on how well that dictionary represents the
language in question. For instance, the query rate for
CEHLNSTY can be improved to 1 query every 46 words, by
eliminating words whose total probability is less than
one part in 50 thousand. This culling can be achieved
for instance by application of a "gap factor" given by
the ratio between two words in the query, e.g. the most
frequent and the least frequent. If, for instance, the
gap factor is set to 500, the distribution of figure xxx
is obtained, two codes, the standard ambiguous code (SAC)
and the CEHLNSTY code are particularly pointed out in
this drawing.
Internationalization There are two main issues in
internationalization of the present embodiment, both of
which are readily solved by a person skilled in the art
in application of the teachings of this invention. These
are 1) handling accents, and 2) creation of generalized
codes which are applicable to many languages at once.
Both of these issues will be briefly discussed.
111
WO 00/35091 PCT/US99/29343
Handling accents. Many languages are written with
letters which may appear in both an accented and an
unaccented form. For example, in French, "e" may be
written as 'Fe", "e", or "e". It is generally important to
distinguish these accented letters. Without accents, for
instance, the word "eleve" (meaning "student") can be
ambiguous with the word "eleve" (meaning "raised"). One
natural approach is to use another shift key, which we
can call an accent-shift key which has the function of
selecting an accented version of a letter when used in
combination with a key encoding that letter. For
instance, to treat French using CEHLNSTY, we can key the
accent shift key in combination with the "def" key to
encode either "'e" or 'Ile", and then rely on a
disambiguation mechanism to decide which of these two
accents is appropriate for a given word. Using this
approach, we find that with some set of word-frequency
statistics for French, a (lookup, query) rate of (38,3)
for CEHLNSTY without an accent-shift key, but (584,24)
with an accent shift key. For the telephone keypad, any
of the keys on the bottom row could be used as an accent
shift key, for example. In specially built keypads, an
additional key could be provided to supply the accent-
shift function. For ergonomic reasons, it is preferable
that this accent-shift key be operated in a manner
similar to the way the usual shift key is operated, e.g.
such that thumb motion in one direction is used to encode
a regular shift operation, while thumb motion in an
opposite direction is used to encode the accent-shift
operation.
112
WO 00/35091 PCT/US99/29343
Multi-Language Ambiguous Codes Since the statistics of
one language are typically different from the statistics
of another language, a code which is.substantially
optimal with respect to one language may not be
substantially optimal with respect to another language. A
code which is strongly touch typable with respect to one
language may not be strongly touch typable with respect
to another language._
For example, CEHLNSTY, optimized for English,
performs less well for French than a code specifically
chosen for its optimality with respect to French. In this
particular example, CEHLNSTY remains strongly touch
typable with respect French, though this will not be the
case with respect to all languages.
In order to gain economies of scale, manufacturers
may wish to produce a single machine operable in many
local linguistic environments. Since a typable device,
for instance, a mobile phone, may preferably have keys
labeled with the ambiguous code for which it is designed,
it is useful to have a single code which applies to many
languages, so that this labelling can be done in the same
way for all machines in a production run, regardless of
the target linguistic community.
Applying exactly the same techniques as have been
already described, it is possible to produce ambiguous
codes which are simultaneously optimized with respect to
several different languages. In a multi-language
optimization method, the step of weighting ergonomic
criteria with respect to each other can include as a
113
WO 00/35091 PCT/US99/29343
sub-step the step of weighting multiple languages with
respect to each other. Different weighting schemes are
appropriate in different circumstances. For instance, one
may choose to simultaneously optimize with respect to
statistics of English and German, and yet weight
performance of the code with respect to English as more
important than performance with respect to German.
A preferred weighting method is one in which the
minimum performance is maximized, such a procedure will
be referred to as a mini-max procedure.
Consider optimizing with respect to a set of
languages 11, 12,..., ln, and a set of ergonomic
criteria, el, e2, ... , em. Given two ambiguous codes
c1, c2, and for each ergonomic criterion em, we rate cl
as better than c2 if for ci the minimum of em over the
languages ln is greater than the minimum of em over the
languages ln for c2. When several ergonomic criteria must
be optimized against, it may happen that one code is
better than another in the mini-max sense with respect to
one ergonomic criterion, but worse with respect to
another ergonomic criterion. In this case the ergonomic
criteria must be weighted with respect to each other as
has already been described in detail.
As an example of these teachings, consider
optimizing with respect to lookup error and query error
for a set of languages. In this example will we allow
non-alphabetic orderings, and use 8 regular input means,
and a auxiliary input means and an accent-shift auxiliary
input means, so that the 8 regular input means can be
114
WO 00/35091 PCT/US99/29343
used in combination with one of the auxiliary input means
in a hybrid chording/ambiguous code embodiment as has
been previously described.
First consider optimizing with respect to the set
of languages consisting of French, Italian, Portuguese
and Spanish, each language represented by a set of
reference statistics.
Using the directed random walk method, one readily
finds codes such as joz m bhx a kn r pw d iy 1 gq t ev c
fu s with (lookup error rate, query rate) of (3250,265),
(11400,3800), (4720,505), and (6280,400) with respect to
French, Italian, Portuguese and Spanish respectively.
This code has relatively poor performance on Dutch,
English, and German, with values of (65,4.8),(93,10), and
(360,13) respectively.
Using a similar amount of computing time, but
optimizing now with respect to Dutch, English, and
German, one can find codes such as cjk r biy 1 fv e mo a
sz p hx g tu d qw n which yields (1220,44), (816,44),
(480,47) on these languages respectively. This same code
yields (253,20),(306,50),(525,36), (4236,272) for French,
Italian, Portuguese and Spanish respectively. And while
these results are respectable for the out of sample
languages, they are not nearly as good as the results
obtained when ambiguity with respect to these languages
is explicitly optimized. These results suggest that the
more different the languages are from each other, the
less performance on one language generalizes to
performance on another language.
115
WO 00/35091 PCT/US99/29343
In practical circumstances, the decision as to which
languages to include in a multi-language optimization
scheme are more commercial then conceptual. The
important inventive concept taught by this disclosure is
that even the minimal performance over the selected
;languages should be such that the code should be strongly
touch typable. In the cases examined above, optimization
with respect to French, Italian, Portuguese and Spanish
discovered a code which is level C strongly touch typable
with respect to these languages, but the minimal
performance is barely level A strongly touch typable
with respect to Dutch, English, and German.
Strongly touch typable handheld device which is typable
using one hand. In the hybrid chording/ambiguous code
embodiment described above, it was shown how a
distinguished input means can be used to form chords with
input means encoding ambiguously coded symbols in order
to reduce the ambiguity of the overall system. This
present embodiment shows how the same input means can be
used both for chord formation and for encoding
ambiguously encoded symbols. In this case the ambiguous
code can be expressed as a multi-level code: a first
sequence of input means manipulations serves to select a
first subset of decoding symbols, a second sequence of
input means manipulations serves to select a second
subset of decoding symbols, and so on. Preferably, the
second subset is a subset of the first subset, the third
subset a subset of the second subset (and thus a subset
of the first subset), and so on. This is a "divide and
conquer" approach, as such well known to those in the
art. However, it is not been heretofore understood that
116
WO 00/35091 PCT/US99/29343
a) the number of successive subdivisions of the set of
symbols can be limited by making the smallest subsets
contain more than one symbol, and thus represent an
ambiguous code, nor b) that the manner of subdivision can
be chosen so as to minimize the ambiguity of the final
ambiguous code, nor c) that ambiguity reduction can be
optimized while simultaneously optimizing other ergonomic
criteria, such as adherence to convention, nor d) that
the transition between levels in the hierarchy can be
accomplished with chords consisting of pairs of key
presses only.
A concrete manifestation of these discoveries will
now be described in detail, with reference to figures 39
to 47. It will be understood that this is but one of an
infinite number of devices that can be built according to
the teachings of this invention: any typable device which
relies on a divide and conquer approach to code
construction, while minimizing ambiguity and/or some
other ergonomic criterion such as adherence to
convention, fits well within the scope of this
embodiment.
This embodiment is a device which is strongly
touch-typable with a single hand. It has the additional
desirable characteristics of
1) permitting a fully unambiguous text-entry mode.
2) permitting a substantially optimal, strongly
touch typable, ambiguous text-entry mode.
117
WO 00/35091 PCT/US99/29343
3) permitting a minimal-keystroke mode for
data-retrieval.
4) is such that the above three modes are
maximally ergonomically compatible.
Preferably, this device can be configured such that,
in addition to the above-stated ergonomic criteria, a new
ergonomic criterion, scan time, is also optimized.
Optimization of scan time will be discussed in,a section
below.
An overview of the method for constructing a
typable device based on multi-level ambiguous codes is
described in reference to figure 39.
In the first step 150, a set of second-level
decoding symbols are selected. These are the symbols that
are to be represented by the ambiguous code, and might
include, for example for English, the letters a
through z.
In the next step 151, an ergonomic criterion is
selected for the over-all multi-level code. This
ergonomic criterion could be, for example, strong touch
typability or lookup error. In general, many ergonomic
criteria of the multi-level code could be simultaneously
selected. In the next step 152, the second-level decoding
symbols are divided into subsets. An encoding symbol is
assigned to each second-level subset in such a way that
the overall code is optimized with respect to the
selected ergonomic criterion. Up to this point, the
118
WO 00/35091 PCT/US99/29343
construction is not different from the construction of
any optimized ambiguous code. However, there may be
additional constraints, for instance on the allowed
number of encoding symbols, such that the next step of
construction can be executed. In this next step 153, the
second-level encoding symbols are collected into groups.
These groups are treated as decoding symbols for a
first-level ambiguous code. Otherwise said, the encoding
symbols of the second-level code become decoding symbols
for the first-level code. Hence, a first-level encoding
symbol is assigned to each group, forming a first-level
ambiguous code. Additional optimization of ergonomic
criteria can be performed in the assignment of
second-level symbols into groups. In general, each level
in a multi-level code can be optimized with respect to
different ergonomic criteria. These criteria may be the
same, or may be different from, the ergonomic criteria
with respect to which the over-all multi-level code is
optimized. In the final step 154, the multi-level code
thus constructed is embodied in a typable device.
in this description of the construction of a
multi-level code, the construction of the second-level
code was described as proceeding the construction of the
first-level code. In using a device embodying a
multi-level code this order is reversed: first an element
of the first-level code is selected by manipulation of
input means, then an element of the second-level code is
selected by further manipulation of input means. This is
the essence of the divide-and-conquer approach. It will
be evident to one skilled in the art that this
119
WO 00/35091 PCT/US99/29343
construction could be continued in the same way to
include third- and higher-level codes.
In practice, properties of each level of the
multi-level code must be simultaneously optimized to
achieve desired properties of the over-all multi-level
code. The present embodiment is presented to concretely
illustrate how this simultaneous optimization can be
planned and executed.
The overview of the method of construction of this
embodiment is shown in figure 40. To help exhibit the
procedure in full generality, three ergonomic criterion
are selected for the over-all multi-level code, two
criteria for the first-level code and three for the
second-level code which comprise the multi-level code. In
this embodiment the three ergonomic criteria applied to
the multi-level code are strong touch typability, query
error, and lookup error. The first-level code is
optimized relative to anatomic fidelity, and alphabetic
ordering, and the second-level code is optimized relative
to evenness of partition, anatomic fidelity, and
substantial alphabetic ordering.
The first step 3100 in the construction of this
embodiment is the selection of (second-level) decoding
symbols. These are the letters a-z. Then strong touch
typability, query error, and lookup error are selected as
ergonomic criteria for the multi-level code in steps
3101, 3102, and 3103 respectively. Then, in step 3104,
anatomic fidelity is selected as a ergonomic criterion.
Since this device is meant to be typable using the
120
WO 00/35091 PCT/US99/29343
fingers of the hand holding the device, anatomic fidelity
is maximized when there are 4 input means and 4
corresponding first-level encoding symbols, one for each
finger.
Anatomic fidelity is chosen as an ergonomic
criterion for the second-level code in step 3105. Each
encoding symbols in the first-level code will correspond
to several encoding symbols for the second-level code.
Anatomic fidelity of the second-level code is maximized
if each of the 4 first-level symbols corresponds to 4
second-level encoding symbols, so the number of
second-level encoding symbols should be 16 for anatomic
fidelity to be maximized. 16 second-level encoding
symbols can be associated with second-level decoding
symbols such that evenness of partition is maximized if
the 26 second-level decoding symbols are distributed over
the second-level encoding symbols such that either 1 or 2
second-level decoding symbols are associated with each of
the 16 second-level encoding symbols. This distribution
implies, in turn, that between 4 and 8 second-level
decoding symbols will be ultimately associated with each
of the 4 first-level decoding symbols.
Next, in step 3106, alphabetic ordering is selected
as an ergonomic criterion for the first-level code.
Optimizing with respect to this criterion requires
simultaneous optimization of both the first- and
second-level codes. What is required is that the letters
a-z must be displayable in alphabetic order on the
displays corresponding to the input means associated with
each finger respectively. Since these displays are
121
WO 00/35091 PCT/US99/29343
arranged in order of the fingers, this implies in turn
that letters from the first part of the alphabet must be
associated to second-level decoding symbols which are in
turn associated to the first-level encoding symbol which
is associated to input means associated to the first
finger. In the same way, a second group of letters,
following the first group in alphabetic order, must be
assigned to second-level encoding symbols associated to
the first-order encoding symbol associated with the input
means associated to the next finger, and so on for the
other two first-level encoding symbols. Optimizing with
respect to alphabetic ordering thus corresponds to
choosing an ordered partition of the 26 letters, in the
same way as has been discussed for other embodiments of
this invention. This time, each of the 4 elements of the
ordered partition must have between 4 and 8 subelements,
so that all of the ergonomic criteria listed can
simultaneously optimized. As will be shown in the
detailed description of the best mode for this
embodiment, codes with even-as-possible partitions, even
for the first-level code, can be found, while optimizing
as well with respect to all of the other ergonomic
criteria considered.
Finally, in step 3107, substantial alphabetic
ordering is chosen as an ergonomic criterion of the
second-level ambiguous code. This means that it should be
possible to lay out the letters as well as possible in
alphabet ordering, given all of the other constraints on
the assignment of letters to second-level encoding
symbols. Divergences from strict alphabetic ordering can
be measured in any number of ways, for instance by the
122
WO 00/35091 PCT/US99/29343
number of pairwise permutations required to bring a given
ordering into strict alphabetic ordering.
In reference now to figures 41-47, we describe a
strongly touch typable handheld device, typable using one
hand, encoding at least the letters [a-z], and embodying
a code constructed according to the above described
method. In order for a device built according to the
present embodiment to be touch typable, the division of
the symbols into subsets, subsets of subsets, and so on
must be fixed, that is, not changing depending for
instance on which symbols have been previously entered.
This fixedness pertains only to the requirement of touch
typability, and the teachings of this invention could be
applied in a broader context. For instance, allowing
word completion mechanisms can significantly reduce the
number of keystrokes. But since the behavior of the
word-completion mechanism is complex and difficult to
predict, a machine with word-completion is not in any
strict sense touch-typable. Nonetheless, the same
optimizations that lead to strongly touch typable codes
lead to effective word completion mechanisms, since the
lower the ambiguity, the better word completion can be
effected. Thus, augmenting a strongly touch typable
code with word-completion mechanism does not transport
the device beyond the scope of the present invention.
For the present embodiment we make the further
limitation that the symbol-input typable part of the
device must be able to be held in one hand, and be
typable by the hand holding the device, and that hand
only. To limit the requirement for digit motion, most of
123
WO 00/35091 PCT/US99/29343
the symbols can be input though sequences of manipulation
of but 5 input means: 4 input means operable by the
fingers of the hand holding the device 2100-2103 and 1
input means operated by the thumb of the hand holding the
device 2104. The device shown in figure 41 is meant to
be held in the left hand; it is evident that the
symmetric device designed to be held in the right hand or
an ambidextrous device operable by either hand could also
be constructed.
Preferably, associated with each of the input means
2100-2103 is a visual display 2106-2109 showing the
elements of the subset currently associated with the
given input means. Operating the input means selects the
corresponding subset. The input means 2104 can be used to
further refine the subset selection and/or be used to
select other subsets of symbols. For example, the single
symbol "space" can be associated with the input means
2104; this or other symbols associated with input means
2104 can be preferably displayed on display means 2110.
The letters [a-z] can be distributed over the 4 input
means 2100-2103. The distribution of letters over the
input means is preferably chosen so as to minimize the
ambiguity (lookup error rate and/or query rate) of the
resulting code, while simultaneously adhering to the
convention of alphabetic ordering. This adherence aids
the novice user in finding a needed letter by simply
scanning the candidate letters.
Figure 42 shows an arrangement of the letters [a-z]
in which the letters [a-f] are associated with the first
input means 2100 ,[g-1] with the second input means
124
WO 00/35091 PCT/US99/29343
2101, [m-r] with the third input means 2102, and
finally [s-z] with the fourth input means 2103. These
associations constitute first-level subsets in a
first-level code. In general, it will be preferred to
associate 4-8 letters with each of these 4 input means:
thereby the subset of letters associated to each input
means can be further subdivided into 4 subsets, each of
which contains no more than two letters. The utility of
this limitation will become evident shortly, and it will
be evident to one skilled in the art how to extend the
teachings of this embodiment to languages with a
different number of symbols, and a different number of
input means.
An example set of second-level subsets which divide
the first-level subsets shown in figure 42 is shown in
figure 43. Figure 43 is a table of four columns and four
rows. The columns are labeled by the input means
activated at the first step, the rows by the symbols
associated to each input means at the second step. Thus,
for instance, if input means 2100 is first activated,
then at the second step, the symbols ac will be
associated to input means 2100, be to the input means
2101, etc. This assignment is chosen to minimize lookup
error and query rates, given the constraints on subset
size described above. The lookup error and query rates
for this code are (1100,69) using our reference
statistics. It is to be noted very carefully that in
this example, the letters in the first level subsets are
arrangeable in alphabetic ordering, but the letters in
the second-level subsets are only partially arrangeable
in alphabetic ordering. It was decided for this example
125
WO 00/35091 PCT/US99/29343
to loosen the alphabet ordering constraint at the second
level in order to permit better query and lookup rates,
and to produce a code which is as strongly touch typable
as possible. This shows that alphabetic ordering can be
optimized, or not, just like any other ergonomic
criterion, and that weighting of optimization properties
can be different at different levels in a multi-level
ambiguous code. The advantage, again, of alphabetic
ordering is that it reduces scan time, especially for
novice users. Since the number of symbols displayed at
the second level is small, scan time is in any case
small, and can be further reduced by mechanisms to be
discussed presently.
To type a given desired letter, the user first
activates one of the input means 2100-2103 corresponding
to first subset containing the desired letter. The user
then selects one of the second-level subsets by again
activating one of the input means 2100-2103
corresponding to a set of letters containing the desired
letter. Figure 44 shows an example operation of the
device in which the user types the letter e. With
reference to figure 42 we see that e is associated with
input means 2100 by the first level code. The user
activates this input means, and the display becomes that
shown in figure 44. Now the letter e is associated with
the input means 2101. When this input means is
manipulated, the letter e is output. The same sequence
of manipulations of input means serves also to select the
letter b, so the code is ambiguous. As in the other
embodiments, which of the letters b or e is intended is
determined from context by a disambiguation mechanism.
126
WO 00/35091 PCT/US99/29343
Words are entered by successively selecting the required
letters in this manner, and terminating the word by
activation of the input means 2104 associated with the
thumb, two-hand operation. It should be noted that since
a two-stroke method is used to encode each letter, this
input method can form the basis of a one-hand/two-hand
embodiment. More explicitly, if one hand is used to
indicate the first stroke of each letter, and the second
hand is used to indicate the second stroke of each
letter, then first and second-stroke information could be
input simultaneously. There are numerous physical
embodiments which could be based on this remark. For
instance, the "fingering" mechanism of [4] could be the
physical substrate upon which a one-hand/two-hand
embodiment could be based. The code proposed by [43 is
based on motion sensors capable of sensing several
positions per finger, to encode each letter
unambiguously. This requires relatively sophisticated
sensors. However, using a two-handed variant of the
present embodiment, simpler sensors could be used. These
sensors would need only record binary (up/down)
information for each finger. Both software and hardware
complexity could be reduced in this manner. in addition,
a machine build according to the teachings of this
invention would be simpler for the user to learn and
operate.
Visual Cache Scan time is the time it takes to visually
locate a desired letter from a set of letters. The
hunt-and-peck typist visually scans the keyboard to find
a next letter and then presses the corresponding key.
Scan time is determined by a number of factors, including
127
WO 00/35091 PCT/US99/29343
the user's familiarity with the layout of a keyboard. The
hunt-and-peck typist may know basically where the
desired key is, and is using visual scanning only for
confirmation, or precise localization. It is to improve
scan time, through the familiarity of typical users with
an alphabetic ordering, that alphabetic ordering was
chosen for the first-level code of this embodiment.
In a variant of alphabetic ordering, certain letters are
selected from the group of letters on a given key for
display in a distinguished, selected area of the visual
display associated with the key. These letters are the
most likely-to-be-selected letters at any given moment,
and placing them in a distinguished position makes them,
easier to find. The principle is analogous to the cache
used in some computer processors to store recently-used
data in registers where they can be gotten at quickly,
under the hypothesis that data recently used is more
likely to be used again. Here, letters are placed in
cache not on the basis of their recent use, but on the
basis of their likelihood to be used next, given the
statistics of the language. Still, the term "visual
cache" seems appropriate.
One embodiment of a visual cache will now be
described in the context of the current embodiment. It
will be appreciated that this invention permits a wide
variety of modifications without modification of its
essential quality, for instance, modifications in the
size and location of the cache, how the cache is
organized, how it is labelled, and the like.
128
WO 00/35091 PCT/US99/29343
From analysis of our standard statistics, we find
that of the letters [a-f] associated to input means 2100
by the first-level code, "a" is the most likely to be the
first letter of a word. Similarly, of the letters fg-l]
associated to input means 2101 2101, "i." is the most
likely to be the first letter of a word, "o" the most
likely from [m-r] associated with input means 2102, and
"t" the most likely from the letters [s-z] associated to
input means 2103.
By placing the letters a,i,o,t in a distinguished
portion of the display, for instance the upper-left hand
corner of the display area associated with each input
means. This makes these letters the first-encountered
letters in a standard left-to-right, top-to-bottom visual
scan of each associated display. Preferably, other than
this selection of a single letter out of alphabetic
order, alphabetic ordering is maintained for the other
letters in the subset. The distinction between the
letter in cache and the rest of the letters can be
further marked by selecting a different color, size,
style, etc. of font for the cached letter than for the
rest of the letters.
Referring now to figures 45 and 46, we see how this
observation can be exploited to reduce scan time.
Figure 45 shows how the word "think" is entered without
the use of a visual cache, and figure 46 shows the same
word entered using a visual cache. Thus, in figure 45,
the letter "t" is entered by first activating the first
input means 2103 corresponding to the letter "t". Before
the first input means is activated, the display as shown
129
WO 00/35091 PCT/US99/29343
in the second column of the figure. Once 2103 is
activated, the display changes to that shown in the third
column. When the input means 2101 is then activated, the
letter "t" is output. The displays change similarly as
the other letters of the word "think" are entered.
In figure 46, the distinction between letters in and
not in visual cache is marked by writing letters in the
visual cache in upper case letters, while letters not in
visual cache are displayed using lower-case letters. For
our reference statistics, we find that 42 percent of
first letters of words are either a,i,o, or t. Thus 42
percent of the time a user beginning to enter a word will
find the required letter immediately in cache.
As a word is entered the most-likely next letter
changes as context is created by the entering of the
word. Thus, the letter selected to be cached should
change as a word is entered, and will depend on which
word is being entered.
In the case of the word "think" The letter t" is
found in visual cache both before activation of the first
input means, and before activation of the second input
means, as shown in the first four rows of figure 46, each
row corresponding to an input means display. Once "t" has
been selected. The letters in visual cache are a,h,o,w,
as shown in the second set of 4 rows in figure 46. After
the first input means (input means 2101) is selected to
begin entering the letter "h", there are only two
possible letters which form parts of words according to
the reference statistics, these are the letters "h and
130
WO 00/35091 PCT/US99/29343
"i", both of these appear in visual cache in the second
display. Continuing in this way for the letters, i,n;k,
we find that the desired letter is always in visual cache
for this word. Indeed, after the first input means
manipulation for entering the letter "i'l, there is only
one possible letter which could have been intended if the
user is in fact entering a word in the database. In this
case, then, a second input means manipulation is
redundant, and the letter "i" could be output immediately
after the first input means manipulation.
Explicit disambiguation and inputting of additional
symbols. As has already been pointed out, it is
generally desirable to provide a completely unambiguous
method for inputting symbols in a typable device based on
ambiguous codes. In the present embodiment, one simple
way to provide unambiguous input is by provision of an
additional unambiguous input means 2105 shown in figure
41 in a position where it is easily activatable by the
thumb, which is the preferred position. Other positions,
however, could be chosen.
In the present embodiment, it was chosen to limit
the size of second-level subsets to at most 2 symbols.
Thus the disambiguation mechanism will always either
choose the desired symbol correctly, or it will chose the
other, incorrect, symbol to which it is paired. Any
disambiguation software could generate a signal to
indicate which of the two symbols it would choose were
the corresponding input means to be selected by the user.
This signal could be used to provide feedback to the
user, for instance by highlighting the letter to be
131
WO 00/35091 PCT/US99/29343
chosen. If the letter to be chosen is not the desired
letter, the user has the option of activating the
explicit disambiguation input means 2105 shown in
figure 41 to force the choice of the other, non-
highlighted symbol.
An example use of this unambiguous text entry
mechanism is shown in figure 47. As in figures 45 and
46, this figure shows how the word "think" is entered.
Here the fourth column gives the letter that would be
output were the unambiguous input means 2105 to be
activated after activation of the first and second input
means used to enter the letters of the word "think". For
example, if input means 2103 and then input means 2101
had been activated to enter the letter "t", then further
activation of input means 2105 would select the letter
"u". Then 'fu" would become the first letter of the word.
All possible letters can be entered unambiguously in this
way. When the second-level subsets contain but one
letter, this one letter will be entered unambiguously
even without activation of the input means 2105, and thus
input means 2105 is inapplicable. For the given code,
the letters d,f,h,l,n, and p are always entered
unambiguously.
Strong Touch Typability: Measurement and Thresholds
Strong touch typablity is a new, inventive concept
describing a definite class of machines. The breadth of
this concept has been pointed out through a variety of
embodiments placed near the boundaries of this class, and
thus indicating its extent.
132
..~,.~ _
WO 00/35091 PCT/US99/29343
To further add definiteness to the disclosure of
strong touch typability, this section will present an
alternative numerical characterization of strong touch
typability which will allow the strong touch typability
of any ambiguous code to be measured, and to thus decide
if that code falls within the scope of this invention or
not.
Language Statistics It has already been mentioned that
representing a language in terms of a corpus of text is a
topic of research among linguists. For the sake of
numerical definiteness, define a representative corpus
as a collection of at least 10 million words drawn at
random from a general-interest newspaper in the target
language.
Key Number We need to define four kinds of key numbers:
physical key number, chording key number, effective key
number, and combined effective key number. Physical key
number: the number of inputs used to encode symbols. A
minimal qwerty keyboard has 26 keys labeled with a
letter, a shift key and a space key, it thus has a
physical key number of 28. Chording key number: the
number of distinct combinations of keys which encode
symbols. For the minimal qwerty keyboard, the shift key
can be combined with any of the letter keys to form a
capital letter, therefore this keyboard has a chording
key number of 28+26-1=53, since the shift key alone does
not encode any symbols. Otherwise said, a keyboard which
is fully equivalent to the minimal qwerty keyboard
could be build with 53 physical keys, each one encoding a
single symbol, either an upper case or a lower case
133
WO 00/35091 PCT/US99/29343
letter. Indeed, some early typewriters were of this
structure.
Effective key number: Given a set of symbols to
represent in an ambiguous code, a set of language
statistics, and a number of physical keys, P, there
exists an optimal ambiguous code which has the best
possible lookup and query rates, given that these are the
only constraints on the code. Let us call these rates P1
and Pq respectively. Any ambiguous code on any number of
physical keys will have an effective key number of P if
its lookup rates and query rates are equal to Pl and Pq.
It is impossible for a keyboard with a physical number of
keys less than P to support an ambiguous code with an
effective key number equal to or greater than P. It is
perfectly possible, and usually the case, that an
ambiguous code on a physical number of keys P has an
effective key number less than P. Combined effective key
number: it is an experimental observation that the lookup
error rates and query rates of substantially optimal
ambiguous codes are substantially related by a power law,
as shown for example, by the experimental results of
figure 11. Here it is shown that for English, the log
of the substantially optimal query rate is linearly
related to the log of the substantially optimal lookup
rate.
This observation allows us to define a single number
relating the lookup error rate and query rate of a code:
the projection of the point (lookup error rate, query
rate) of the code onto the best-fit line in a log-log
plot.
134
WO 00/35091 i'CT/US99/29343
Consider, for example, the standard ambiguous code.
This code has (lookup error rate, query rate) of
(29,2.2). Projecting this point onto the best-fit line of
figure 11 and linearly interpolating, we find the value
5.96, which is thus the combined effective key number of
the standard ambiguous code. Though the standard
ambiguous code is defined on 8 physical keys (and with a
chording key number also of 8, since no chording is
involved), it is equivalent in ambiguity to a substantial
optimal code on 5.96 physical (or chording) keys. Of
course fractional physical keys are not possible in
practice, but these results indicate that a substantially
optimal code on 6 keys could be found which has lookup
error rates and query rates better than the standard
ambiguous code.
These considerations allow us to define a precise,
albeit arbitrary, numerical threshold for substantial
optimality of combined lookup error rate and query rates:
a code will be said to be substantial optimal with
respect to these rates if its combined effective key
number is within 0.01 of its chording key number, if
there are no other ergonomic constraints on the system.
We can also define a precise, albeit arbitrary, threshold
for strong touch typability: an ambiguous code for
English can be defined as strongly touch typable if its
combined effective key number is at least 10. We can
extend this definition to other languages by requiring
that for a code to be strongly touch typable, it must
have a lookup error rate and a query rate greater than or
equal to those of a strongly touch typable code for
English. Since the combined effective key number of the
135
~~.~
__. ,~.. . ~._...-..---
WO 00/35091 PCT/US99/29343
standard ambiguous code is less than 10, it is not
strongly touch typable by the considerations of this
section.
By measuring the combined effective key number,
any ambiguous code can be screened for possession of the
strong touch typability property. For instance, the
hybrid chording/ambiguous code: ab c df e gi h jk 1 mo n
pqr s uv t wxz y discussed above has 9 physical keys: the
8 letter keys of the standard telephone keypad plus 1
shift key. It has a chording key number of 16; it is
equivalent to an ambiguous code on 16 independent keys
with no shift key. Without application of a gap factor,
its (lookup,query) rates are (431, 21) corresponding
to a combined effective key number of 12.8. With
a gap factor of 500, this improves to (440,46),
corresponding to a combined effective key number of
13.75. With or without a gap factor, this code is
strongly touch typable. It should be noted that the
combined effective key number is less than the chording
key number, still, this code is substantially optimal
given the additional constraint of alphabetic ordering.
In the same way, the one-handed hybrid chording/
ambiguous code embodiment has a code with (lookup, query)
rates of (1100,69), yielding a combined effective key
number of 15, though its physical key number is 4 and its
chording key number is 16. It is a strongly touch typable
code, and the difference between the chording key number
and combined effective key number is due to the
additional ergonomic constraint of alphabetic ordering
of the first-level code. Taking into account this
additional constraint, the code is substantially optimal.
136
WO 00/35091 PCT/US99/29343
By contrast, the 14-physical key code of Fujitsu, pn gt
cr zk wj a e hi so ud xf ym vl qb, has (lookup, query)
rates of (105,4), and a combined effective key number of
8.47. This code is neither substantially optimal
nor strongly touch typable, despite the fact that its
physical number of keys is greater than 10.
Although illustrative embodiments of the present
invention have been described herein with reference to
the accompanying drawings, it is to be understood that
the invention is not limited to those precise
embodiments, and that various other changes and
modifications may be effected therein by one skilled in
the art without departing from the scope of the
invention.
Literature Cited
[1] http:/www.fujitsu.co.jp/hypertext/news/1996/Jun/ms-
txt.html
http:/www.fujitsu.co.jp/hypertext/news/1998/May/27-e.html
http:/www.fujitsu.co.jp/hypertext/news/1996/Jul/text.htm
[2] USP 5,818,437 Reduced keyboard disambiguating
computer. US5818437, Oct 6, 1998
[3] Reduced keyboard disambiguationg system,
PCT/US98/01307; W098/33111
[4] "Body Coupled Fingering": Wireless Wearable
Keyboard, CHI 97 Electronic Publications: Papers, by
FUKUMOTO, Masaaki and TONOMURA, Yoshinobu NTT Human
Interface Laboratories 1-1 Hikari-no-oka, Yokosuka-shi,
137
WO 00/35091 PCT/US99/29343
Kanagawa-ken, 239 JAPAN web reference:
http://www.acm.org/turing/sigs/sigchi/chi97/
proceedings/paper/fkm.htm#U21.
138
