METHOD OF REALIZING CONFIGURABLE INTELLIGENT VOICE ROBOT, DEVICE AND STORAGE MEDIUM

METHODE DE REALISATION D'UN ROBOT VOCAL INTELLIGENT CONFIGURABLE, DISPOSITIF ET SUPPORT DE STOCKAGE

English Abstract

Pertaining to field of artificial intelligence, the present application discloses a method of realizing a configurable intelligent voice robot, and device and storage, of which the method comprises: obtaining sample corpora of various conversation scenarios in a plurality of conversation scenarios; generating, with respect to each conversation scenario, a scenario feature of the conversation scenario on the basis of the sample corpus of the conversation scenario, wherein the scenario feature includes feature words of the conversation scenario and a feature word sequence obtained by mapping transformation of the feature words; and configuring the intelligent voice robot on the basis of a preset word vector space model and scenario features of the various conversation scenarios, wherein the word vector space model is employed for the intelligent voice robot to perform word vector similarity calculation on a user's conversation and the scenario features of the various conversation scenarios.

French Abstract

Concernant le domaine de l'intelligence artificielle, il est décrit une méthode de réalisation d'un robot vocal intelligent configurable, ainsi qu'un dispositif et un stockage, ladite méthode comprenant : l'obtention d'échantillon de corpus d'une pluralité diverse de scénarios de conversation; la génération, pour chaque scénario de conversation, d'une caractéristique de scénario du scénario de conversation sur la base de l'échantillon de corpus du scénario de conversation, la caractéristique de scénario comprenant des mots de caractéristiques du scénario de conversation et une séquence de mots de caractéristiques obtenue par transformation de mappage des mots de caractéristiques; et la configuration du robot vocal intelligent sur la base d'un modèle d'espace vectoriel de mots prédéfini et de caractéristiques de scénario des divers scénarios de conversation, le modèle d'espace vectoriel de mots étant employé pour le robot vocal intelligent afin de calculer la similarité de vecteur de mots sur la conversation d'un utilisateur ou d'une utilisatrice et les caractéristiques de scénario des divers scénarios de conversation.

Claims

Claims:
1. A device comprising:
an obtaining module, configured to obtain sample corpora of various
conversation
scenarios in a plurality of conversation scenarios;
a generating module, configured to generate a scenario feature of the
conversation
scenario based on a sample corpus of the conversation scenario, wherein the
scenario
feature includes feature words of the conversation scenario and a feature word
sequence
obtained by mapping transformation of the feature words; and
a configuring module, configured to configure an intelligent voice robot based
on a preset
word vector space model and scenario features of the various conversation
scenarios,
wherein the word vector space model is for the intelligent voice robot to
perfomi word
vector similarity calculation on a user's conversation and the scenario
features of the
various conversation scenarios, to recognize an intended scenario of the
user's
conversation.
2. The device of claim 1, wherein the generating module further comprises:
a representing unit, configured to obtain discrete representation of the
sample corpus of
the conversation scenario based on a preset domain dictionary;
a screening unit, configured to apply a feature selection algorithm to extract
feature
words of the conversation scenario based on the discrete representation of the
sample
corpus of the conversation scenario; and
a generating unit, configured to:
map and transform the feature words of the conversation scenario to a
corresponding dictionary index;
generate the feature word sequence of the conversation scenario, wherein the
feature selection algorithm is a chi-square statistic feature selection
algorithm.
37

3. The device of claim 1, further comprises:
a storing module, configured to store the various conversation scenarios and
the scenario
features of the conversation scenarios in a scenario feature relation table;
4. The device of claim 3, further comprises:
a receiving module, configured to receive a configuration feature word input
with respect
to any conversation scenario; and
a maintaining module, configured to maintain the scenario feature of the
conversation
scenario in the scenario feature relation table based on the configuration
feature word of
the conversation scenario and a configuration feature word sequence obtained
by
mapping transformation of the configuration feature word.
5. The device of claim 4, wherein the receiving module is configured to
receive a
configuration feature word input by a user having feature configuration
permission with
respect to the conversation scenario.
6. The device of claim 4, wherein the maintaining module is configured to:
merge the configuration feature word of the conversation scenario to the
feature words of the conversation scenario in the scenario feature relation
table;
and
add the configuration feature word sequence of the merged configuration
feature word to the feature word sequence of the conversation scenario.
7. The device of claim 1, further comprises a training module configured to
use a domain
corpora of a domain to which the various conversation scenarios pertain to
train a pre-
trained bidirectional encoder representations from transformers (BERT) word
vector space,
and obtain the word vector space model.
8. The device of claim 1, further comprises a state machine configuring
module configured to:
38

receive a state transition graph input by a first user for the conversation
scenario and
receive supplementary information input by a second user for a state
transition graph, to
generate a state transition matrix of the conversation scenario; and
generate a script file for containing state transition logical relation based
on the state
transition matrix of the conversation scenario and generate a finite state
machine (FSM)
based on the script file, to return a corresponding pattern when the intended
scenario of
the user's conversation is recognized.
9. The device of claim 1, further comprises an intended scenario
recognizing module
comprising:
an obtaining unit, configured to:
preprocess the user's conversation to obtain a plurality of segmented terms in
the user's conversation;
perform mapping transformation on the plural segmented terms;
obtain a feature word sequence of the user's conversation;
a constructing unit, configured to implement the word vector space model to
construct a
feature vector of the user's conversation and scenario feature vectors of the
various
conversation scenarios based on the feature word sequence of the user's
conversation and
feature word sequences of the various conversation scenarios;
a matching unit, configured to:
perform similarity calculation on the feature vector of the user's
conversation
and the scenario feature vectors of the various conversation scenarios; and
recognize intention of the user's conversation based on a similarity
calculation
result, to return a pattern to which the intention corresponds.
39

10. The device of any one of claims 1 to 9, wherein the plurality of
conversation scenarios are
contained in a preset conversation scenario list, wherein the conversation
scenario list is to
record one or more conversation scenarios of a specific business field.
11. The device of any one of claims 1 to 10, wherein the sample corpora of the
various
conversation scenarios are obtained by classifying and marking specific domain
corpora
according to classes of the conversation scenarios.
12. The device of any one of claims 1 to 11, wherein the conversation
scenarios are obtained by
performing scenario abstraction on the specific domain corpora, and the
scenario abstraction
is a process from data to information and then to knowledge.
13. The device of any one of claims 1 to 12, wherein each conversation
scenario is abstracted as
a conversation state, and dialogue process is abstracted as a transition
between conversation
states.
14. The device of any one of claims 1 to 13, wherein the conversation state is
taken as a node,
an oriented line between the conversation states is a process in which one
state is transferred
to another state, wherein entire dialogue process is abstracted as a graph
consisting of nodes
and oriented lines.
15. The device of any one of claims 1 to 14, wherein content transformation
based on a word of
bag (WOB) model is performed on the sample corpus of each conversation
scenario,
wherein discrete representation of the sample corpus of each conversation
scenario is
obtained, wherein the feature selection algorithm is to extract feature words
of each
conversation scenario on the basis of the discrete representation of the
sample corpus of
each conversation scenario.
16. The device of any one of claims 1 to 15, wherein the WOB model divides a
corpus text into
separate terms, and it is imaged that all terms are placed in a bag, such
elements as their
word orders, grammars and syntaxes are ignored, wherein the bag is only
regarded as a
collection of plural vocabularies, wherein each term as it appears in the
corpus text is
independent.

17. The device of any one of claims 1 to 16, wherein the WOB model includes a
one-hot, TF-
EDF and N-gram model.
18. The device of any one of claims 1 to 17, wherein name, the feature words
and the feature
word sequence of each conversation scenario are correspondingly stored in the
scenario
feature relation table.
19. The device of any one of claims 1 to 18, wherein the scenario feature
relation table stores
correspondence relations between conversation scenarios and scenario features,
including
feature words and feature word sequences.
20. The device of any one of claims 1 to 19, wherein the scenario feature
relation table is stored
in a server, maintained offline by backstage algorithm technical personnel
based on
periodical text data mining operations, and isolated from frontstage business
personnel.
21. The device of any one of claims 1 to 20, wherein when the intelligent
voice robot is
conversing with the user, obtain user's conversation text by recognizing and
transforming
user's conversation voice via an automatic speech recognition (ASR) technique,
and extract
feature information out of the user's conversation text.
22. The device of any one of claims 1 to 21, wherein pre-trained embedding is
obtained by
introducing large-scale BERT word vector space, pretrained in google bert
serving.
23. The device of any one of claims 1 to 22, wherein the BERT word vector
space is retrained
by bringing in own business customer service corpora, and calibration of the
BERT word
vector is realized, to conform to specific business scenarios.
24. The device of any one of claims 1 to 23, wherein total segmentation
corpora where stop-
words are removed are to construct the preset domain dictionary, wherein the
preset domain
dictionary includes total effective vocabularies appearing in the corpora, and
the preset
domain dictionary is to perform content transformation on all sample corpora
of a target
conversation scenario based on the WOB model to obtain the discrete
representation.
41

25. The device of any one of claims 1 to 24, wherein the chi-square statistic
(CHI) technique is
to extract feature words of the target conversation scenario.
26. The device of any one of claims 1 to 25, wherein CHI calculation formula
comprises:
<IMG>
wherein c is a certain class, namely a "conversation scenario", t is a certain
term, and N is
a total number of texts in a training corpora.
27. The device of claim 26, wherein the x2 is for chi square hypothesis
verification in statistics,
to judge uniformity or goodness of fit between actual distribution and
theoretical
distribution, wherein a null hypothesis 1/0 being "no marked difference
between observed
frequency and desired frequency".
28. The device of any one of claims 26 to 27, wherein less the chi-square
statistic is, the closer
the observed frequency to the desired frequency, and higher relevancy between.
29. The device of any one of claims 26 to 28, wherein X2 is a measure of
distance between an
observed object and a desired object, wherein less the distance is, higher the
relevancy
between.
30. The device of any one of claims 26 to 29, wherein the observed object is
the term, and the
desired object is the conversation scenario, wherein if the term and the
conversation
scenario are highly relevant, statistic distributions of the two are close to
each other in the
entire samples.
31. The device of any one of claims 26 to 30, wherein through x2 statistic,
relevancies between
all vocabularies in the domain dictionary and the various classes are
calculated quickly and
accurately based on quantities of corpora, and a preset number of terms with
the x2 selected
according to relevancy sorting result to serve as a feature set of
conversation scenarios, to
complete feature mapping of various scenarios and the various classes in the
conversation
scenario list.
42

32. The device of any one of claims 1 to 31, wherein the configuration feature
word input by the
user with respect to the target conversion scenario through a system frontend.
33. The device of any one of claims 1 to 32, wherein the system frontend
provides business
personnel with a feature relation extension function to maintain the business
field.
34. The device of any one of claims 1 to 33, wherein on receiving the
configuration feature
word, updates it into an externally input feature set of the target conversion
scenario.
35. The device of any one of claims 1 to 34, wherein the externally input
configuration feature
word is not contained in the domain dictionary, the configuration feature word
is ignored.
36. The device of any one of claims 1 to 35, wherein the first user is
business personnel, the
second user is an algorithm developing technical personnel.
37. The device of any one of claims 1 to 36, wherein the FSM is a mathematical
model that
expresses a finite number of states and behaviors as transitions and actions
amongst these
states.
38. The device of any one of claims 1 to 37, wherein the FSM is to describe
state sequences of
an object experienced within the objects life cycle, and how to respond to
various events
coming from outside to transition between the states.
39. The device of any one of claims 1 to 38, wherein the FSM includes current
state, condition,
action, and next state.
40. The device of any one of claims 1 to 39, wherein the current state is a
state currently being
in, wherein the condition is referred to as event, wherein the condition is
satisfied, an action
is triggered or one round of state transition is performed;
41. The device of any one of claims 1 to 40, wherein the action indicates the
action executed
after the condition has been satisfied, wherein the action has been executed
to completion, a
new state is transitioned, or the original state is maintained.
43

42. The device of any one of claims 1 to 41, wherein the action is not
indispensable, after the
condition has been satisfied, not execute any action and directly transition
to the new state;
43. The device of any one of claims 1 to 42, wherein the next state is the new
state to be
transitioned after the condition has been satisfied.
44. The device of any one of claims 1 to 43, wherein the next state is
relative to the current
state, once the next state is activated, it becomes a new current state.
45. The device of any one of claims 1 to 44, wherein FSM model is abstracted
comprising:
START event'END.
46. The device of any one of claims 1 to 45, wherein the FSM state transition
torque completely
repaired is automatically translated to the script file of j son format.
47. The device of any one of claims 1 to 46, wherein the json script file is
read and input by the
program to a state machine object when a finite state machine instance is
generated, to
validate logic.
48. The device of any one of claims 1 to 47, wherein the finite state machine
instance as
generated is stored in Redis as an index value according to uuid transferred
by the frontend,
to facilitate subsequent program access when the interactive voice response
(IVR) service
starts.
49. The device of any one of claims 1 to 48, wherein the user performs
persistent operation on
the FSM.
50. The device of any one of claims 1 to 49, wherein the user selects a task
type as a single task
at the frontend, the finite state machine instance of the single task as
stored in Redis is
cleared away as an invalidated object within a preset time period after the
IVR marketing
service has been triggered.
44

51. The device of any one of claims 1 to 50, wherein the intelligent voice
robot converses with
the user, wherein the user's conversation is text content recognized and
transcribed from
user's conversation speech through ASR technology.
52. The device of any one of claims 1 to 51, wherein the text content is word-
segmented to
obtain the plurality of segmented terms, and word-segmentation process
includes character
purification, rectification, word segmentation, and removal of stop-words.
53. The device of any one of claims 1 to 52, wherein plural segmented terms
are mapped and
transformed through index of the domain dictionary to form of expression of
the
conversation scenario in the scenario feature relation table.
54. The device of any one of claims 1 to 53, wherein each element in the
feature word sequence
of the user's conversation is mapped to the BERT word vector space to obtain
the feature
vector of 768 dimensions, and all elements are summated and averaged to obtain
a 1x768
vector to serve as feature expression input by the user's conversation.
55. The device of any one of claims 1 to 54, wherein the feature word
sequences of the various
conversation scenarios in the scenario feature relation table are operated is
converted to
1x768 feature vectors.
56. The device of any one of claims 1 to 55, wherein a cosine similarity
between the feature
vector input by the user's conversation and the scenario feature vector of the
conversation
scenario is calculated, wherein the greater cosine similarity calculation
result, the higher
similarity, and the higher cosine similarity calculation result is the
relevancy between the
user's conversation and the conversation scenario.
57. The device of any one of claims 1 to 56, wherein by arranging all the
conversation scenarios
in a descending order according to the cosine similarity calculation result,
the conversation
scenario with the highest cosine similarity calculation result is returned to
serve as a
judgement result of the intended scenario of current input by the user, and a
corresponding
response pattern under the intended scenario is returned according to the
state in which the
FSM is currently.

58. The device of any one of claims 1 to 57, wherein the response pattern is
transformed to
speech content through Text-To-Speech (TTS) technology is broadcast to the
user.
59. A system comprising:
an obtaining module, configured to obtain sample corpora of various
conversation
scenarios in a plurality of conversation scenarios;
a generating module, configured to generate a scenario feature of the
conversation
scenario based on a sample corpus of the conversation scenario, wherein the
scenario
feature includes feature words of the conversation scenario and a feature word
sequence
obtained by mapping transformation of the feature words; and
a configuring module, configured to configure an intelligent voice robot based
on a preset
word vector space model and scenario features of the various conversation
scenarios,
wherein the word vector space model is for the intelligent voice robot to
perfomi word
vector similarity calculation on a user's conversation and the scenario
features of the
various conversation scenarios, to recognize an intended scenario of the
user's
conversation.
60. The system of claim 59, wherein the generating module further comprises:
a representing unit, configured to obtain discrete representation of the
sample corpus of
the conversation scenario based on a preset domain dictionary;
a screening unit, configured to apply a feature selection algorithm to extract
feature
words of the conversation scenario based on the discrete representation of the
sample
corpus of the conversation scenario; and
a generating unit, configured to:
map and transform the feature words of the conversation scenario to a
corresponding dictionary index;
46

generate the feature word sequence of the conversation scenario, wherein the
feature selection algorithm is a chi-square statistic feature selection
algorithm.
61. The system of claim 59, further comprises:
a storing module, configured to store the various conversation scenarios and
the scenario
features of the conversation scenarios in a scenario feature relation table;
62. The system of claim 61, further comprises:
a receiving module, configured to receive a configuration feature word input
with respect
to any conversation scenario; and
a maintaining module, configured to maintain the scenario feature of the
conversation
scenario in the scenario feature relation table based on the configuration
feature word of
the conversation scenario and a configuration feature word sequence obtained
by
mapping transformation of the configuration feature word.
63. The system of claim 62, wherein the receiving module is configured to
receive a
configuration feature word input by a user having feature configuration
permission with
respect to the conversation scenario.
64. The system of claim 62, wherein the maintaining module is configured to:
merge the configuration feature word of the conversation scenario to the
feature words of the conversation scenario in the scenario feature relation
table;
and
add the configuration feature word sequence of the merged configuration
feature word to the feature word sequence of the conversation scenario.
65. The system of claim 59, further comprises a training module configured to
use a domain
corpora of a domain to which the various conversation scenarios pertain to
train a pre-
trained bidirectional encoder representations from transformers (BERT) word
vector space,
and obtain the word vector space model.
47

66. The system of claim 59, further comprises a state machine configuring
module configured
to:
receive a state transition graph input by a first user for the conversation
scenario and
receive supplementary information input by a second user for a state
transition graph, to
generate a state transition matrix of the conversation scenario; and
generate a script file for containing state transition logical relation based
on the state
transition matrix of the conversation scenario and generate a finite state
machine (FSM)
based on the script file, to return a corresponding pattern when the intended
scenario of
the user's conversation is recognized.
67. The system of claim 59, further comprises an intended scenario recognizing
module
comprising:
an obtaining unit, configured to:
preprocess the user's conversation to obtain a plurality of segmented terms in
the user's conversation;
perform mapping transformation on the plural segmented terms;
obtain a feature word sequence of the user's conversation;
a constructing unit, configured to implement the word vector space model to
construct a
feature vector of the user's conversation and scenario feature vectors of the
various
conversation scenarios based on the feature word sequence of the user's
conversation and
feature word sequences of the various conversation scenarios;
a matching unit, configured to:
perform similarity calculation on the feature vector of the user's
conversation
and the scenario feature vectors of the various conversation scenarios; and
48

recognize intention of the user's conversation based on a similarity
calculation
result, to return a pattern to which the intention corresponds.
68. The system of any one of claims 59 to 67, wherein the plurality of
conversation scenarios
are contained in a preset conversation scenario list, wherein the conversation
scenario list is
to record one or more conversation scenarios of a specific business field.
69. The system of any one of claims 59 to 68, wherein the sample corpora of
the various
conversation scenarios are obtained by classifying and marking specific domain
corpora
according to classes of the conversation scenarios.
70. The system of any one of claims 59 to 69, wherein the conversation
scenarios are obtained
by performing scenario abstraction on the specific domain corpora, and the
scenario
abstraction is a process from data to information and then to knowledge.
71. The system of any one of claims 59 to 70, wherein each conversation
scenario is abstracted
as a conversation state, and dialogue process is abstracted as a transition
between
conversation states.
72. The system of any one of claims 59 to 71, wherein the conversation state
is taken as anode,
an oriented line between the conversation states is a process in which one
state is transferred
to another state, wherein entire dialogue process is abstracted as a graph
consisting of nodes
and oriented lines.
73. The system of any one of claims 59 to 72, wherein content transformation
based on a word
of bag (WOB) model is performed on the sample corpus of each conversation
scenario,
wherein discrete representation of the sample corpus of each conversation
scenario is
obtained, wherein the feature selection algorithm is to extract feature words
of each
conversation scenario on the basis of the discrete representation of the
sample corpus of
each conversation scenario.
49

74. The system of any one of claims 59 to 73, wherein the WOB model divides a
corpus text
into separate terms, and it is imaged that all terms are placed in a bag, such
elements as their
word orders, grammars and syntaxes are ignored, wherein the bag is only
regarded as a
collection of plural vocabularies, wherein each term as it appears in the
corpus text is
independent.
75. The system of any one of claims 59 to 74, wherein the WOB model includes a
one-hot, TF-
EDF and N-gram model.
76. The system of any one of claims 59 to 75, wherein name, the feature words
and the feature
word sequence of each conversation scenario are correspondingly stored in the
scenario
feature relation table.
77. The system of any one of claims 59 to 76, wherein the scenario feature
relation table stores
correspondence relations between conversation scenarios and scenario features,
including
feature words and feature word sequences.
78. The system of any one of claims 59 to 77, wherein the scenario feature
relation table is
stored in a server, maintained offline by backstage algorithm technical
personnel based on
periodical text data mining operations, and isolated from frontstage business
personnel.
79. The system of any one of claims 59 to 78, wherein when the intelligent
voice robot is
conversing with the user, obtain user's conversation text by recognizing and
transforming
user's conversation voice via an automatic speech recognition (ASR) technique,
and extract
feature information out of the user's conversation text.
80. The system of any one of claims 59 to 79, wherein pre-trained embedding is
obtained by
introducing large-scale BERT word vector space, pretrained in google bert
serving.
81. The system of any one of claims 59 to 80, wherein the BERT word vector
space is retrained
by bringing in own business customer service corpora, and calibration of the
BERT word
vector is realized, to conform to specific business scenarios.

82. The system of any one of claims 59 to 81, wherein total segmentation
corpora where stop-
words are removed are to construct the preset domain dictionary, wherein the
preset domain
dictionary includes total effective vocabularies appearing in the corpora, and
the preset
domain dictionary is to perfomi content transformation on all sample corpora
of a target
conversation scenario based on the WOB model to obtain the discrete
representation.
83. The system of any one of claims 59 to 82, wherein the chi-square statistic
(CBI) technique
is to extract feature words of the target conversation scenario.
84. The system of any one of claims 59 to 83, wherein CHI calculation formula
comprises:
<IMG>
wherein c is a certain class, namely a "conversation scenario", t is a certain
term, and N is
a total number of texts in a training corpora.
85. The system of claim 84, wherein the x2 is for chi square hypothesis
verification in statistics,
to judge uniformity or goodness of fit between actual distribution and
theoretical
distribution, wherein a null hypothesis 1/0 being "no marked difference
between observed
frequency and desired frequency".
86. The system of any one of claims 84 to 85, wherein less the chi-square
statistic is, the closer
the observed frequency to the desired frequency, and higher relevancy between.
87. The system of any one of claims 84 to 86, wherein x' is a measure of
distance between an
observed object and a desired object, wherein less the distance is, higher the
relevancy
between.
88. The system of any one of claims 84 to 87, wherein the observed object is
the term, and the
desired object is the conversation scenario, wherein if the term and the
conversation
scenario are highly relevant, statistic distributions of the two are close to
each other in the
entire samples.
51

89. The system of any one of claims 84 to 88, wherein through e statistic,
relevancies between
all vocabularies in the domain dictionary and the various classes are
calculated quickly and
accurately based on quantities of corpora, and a preset number of terms with
the x2 selected
according to relevancy sorting result to serve as a feature set of
conversation scenarios, to
complete feature mapping of various scenarios and the various classes in the
conversation
scenario list.
90. The system of any one of claims 59 to 89, wherein the configuration
feature word input by
the user with respect to the target conversion scenario through a system
frontend.
91. The system of any one of claims 59 to 90, wherein the system frontend
provides business
personnel with a feature relation extension function to maintain the business
field.
92. The system of any one of claims 59 to 91, wherein on receiving the
configuration feature
word, updates it into an externally input feature set of the target conversion
scenario.
93. The system of any one of claims 59 to 92, wherein the externally input
configuration feature
word is not contained in the domain dictionary, the configuration feature word
is ignored.
94. The system of any one of claims 59 to 93, wherein the first user is
business personnel, the
second user is an algorithm developing technical personnel.
95. The system of any one of claims 59 to 94, wherein the FSIV1 is a
mathematical model that
expresses a finite number of states and behaviors as transitions and actions
amongst these
states.
96. The system of any one of claims 59 to 95, wherein the FSIV1 is to describe
state sequences of
an object experienced within the objects life cycle, and how to respond to
various events
coming from outside to transition between the states.
97. The system of any one of claims 59 to 96, wherein the FSIV1 includes
current state,
condition, action, and next state.
52

98. The system of any one of claims 59 to 97, wherein the current state is a
state currently being
in, wherein the condition is referred to as event, wherein the condition is
satisfied, an action
is triggered or one round of state transition is performed;
99. The system of any one of claims 59 to 98, wherein the action indicates the
action executed
after the condition has been satisfied, wherein the action has been executed
to completion, a
new state is transitioned, or the original state is maintained.
100. The system of any one of claims 59 to 99, wherein the action is not
indispensable, after the
condition has been satisfied, not execute any action and directly transition
to the new state;
101. The system of any one of claims 59 to 100, wherein the next state is the
new state to be
transitioned after the condition has been satisfied.
102. The system of any one of claims 59 to 101, wherein the next state is
relative to the current
state, once the next state is activated, it becomes a new current state.
103. The system of any one of claims 59 to 102, wherein FSM model is
abstracted comprising:
<IMG>
104. The system of any one of claims 59 to 103, wherein the FSM state
transition torque
completely repaired is automatically translated to the script file of json
format.
105. The system of any one of claims 59 to 104, wherein the json script file
is read and input by
the program to a state machine object when a finite state machine instance is
generated, to
validate logic.
106. The system of any one of claims 59 to 105, wherein the finite state
machine instance as
generated is stored in Redis as an index value according to uuid transferred
by the frontend,
to facilitate subsequent program access when the interactive voice response
(IVR) service
starts.
107. The system of any one of claims 59 to 106, wherein the user performs
persistent operation
on the FSM.
53

108. The system of any one of claims 59 to 107, wherein the user selects a
task type as a single
task at the frontend, the finite state machine instance of the single task as
stored in Redis is
cleared away as an invalidated object within a preset time period after the
IVR marketing
service has been triggered.
109. The system of any one of claims 59 to 108, wherein the intelligent voice
robot converses
with the user, wherein the user's conversation is text content recognized and
transcribed
from user's conversation speech through ASR technology.
110. The system of any one of claims 59 to 109, wherein the text content is
word-segmented to
obtain the plurality of segmented terms, and word-segmentation process
includes character
purification, rectification, word segmentation, and removal of stop-words.
111. The system of any one of claims 59 to 110, wherein plural segmented terms
are mapped and
transformed through index of the domain dictionary to form of expression of
the
conversation scenario in the scenario feature relation table.
112. The system of any one of claims 59 to 111, wherein each element in the
feature word
sequence of the user's conversation is mapped to the BERT word vector space to
obtain the
feature vector of 768 dimensions, and all elements are summated and averaged
to obtain a
1x768 vector to serve as feature expression input by the user's conversation.
113. The system of any one of claims 59 to 112, wherein the feature word
sequences of the
various conversation scenarios in the scenario feature relation table are
operated is
converted to 1x768 feature vectors.
114. The system of any one of claims 59 to 113, wherein a cosine similarity
between the feature
vector input by the user's conversation and the scenario feature vector of the
conversation
scenario is calculated, wherein the greater cosine similarity calculation
result, the higher
similarity, and the higher cosine similarity calculation result is the
relevancy between the
user's conversation and the conversation scenario.
54

115. The system of any one of claims 59 to 114, wherein by arranging all the
conversation
scenarios in a descending order according to the cosine similarity calculation
result, the
conversation scenario with the highest cosine similarity calculation result is
returned to
serve as a judgement result of the intended scenario of current input by the
user, and a
corresponding response pattern under the intended scenario is returned
according to the state
in which the FSM is currently.
116. The system of any one of claims 59 to 115, wherein the response pattern
is transformed to
speech content through Text-To-Speech (TTS) technology is broadcast to the
user.
117.A method comprising:
obtaining sample corpora of various conversation scenarios in a plurality of
conversation
scenarios;
generating a scenario feature of the conversation scenario based on a sample
corpus of
the conversation scenario, wherein the scenario feature includes feature words
of the
conversation scenario and a feature word sequence obtained by mapping
transformation
of the feature words; and
configuring an intelligent voice robot with a preset word vector space model
and the
scenario features of the various conversation scenarios, wherein the word
vector space
model is for the intelligent voice robot to perform word vector similarity
calculation on a
user's conversation and the scenario features of the various conversation
scenarios, to
recognize an intended scenario of the user's conversation.
118. The method of claim 117, wherein generating the scenario feature of the
conversation
scenario based on the sample corpus of the conversation scenario comprises:
obtaining discrete representation of the sample corpus of the conversation
scenario based
on a preset domain dictionary;

employing a feature selection algorithm to extract feature words of the
conversation
scenario based on the discrete representation of the sample corpus of the
conversation
scenario;
mapping and transforming the feature words of the conversation scenario to a
corresponding dictionary index; and
generating the feature word sequence of the conversation scenario, wherein the
feature
selection algorithm is a chi-square statistic feature selection algorithm.
119. The method of claim 117, further comprises:
storing the various conversation scenarios and the scenario features of the
conversation
scenarios in a scenario feature relation table.
120. The method of claim 117, further comprises:
receiving a configuration feature word input with respect to the conversation
scenario;
and
maintaining the scenario feature of the conversation scenario in the scenario
feature
relation table based on the configuration feature word of the conversation
scenario and a
configuration feature word sequence obtained by mapping transformation of the
configuration feature word.
121. The method of claim 120, further comprises receiving the configuration
feature word input
by the user having feature configuration permission with respect to the
conversation
scenario.
122. The method of claim 120, wherein maintaining the scenario feature of the
conversation
scenario in the scenario feature relation table based on the configuration
feature word of the
conversation scenario and the configuration feature word sequence obtained by
mapping
transformation of the configuration feature word comprises:
56

merging the configuration feature word of the conversation scenario to the
feature words
of the conversation scenario in the scenario feature relation table; and
adding the configuration feature word sequence of a merged configuration
feature word
to the feature word sequence of the conversation scenario.
123. The method of claim 117, wherein the word vector space model is trained
and obtained
comprises:
applying a domain corpora of a domain to the various conversation scenarios
pertaining
to train a pre-trained bidirectional encoder representations from transfonners
(BERT)
word vector space; and
obtaining the word vector space model.
124. The method of claim 117, further comprises:
receiving a state transition graph input by a first user for the conversation
scenario and
receiving supplementary information input by a second user for a state
transition graph,
to generate a state transition matrix of the conversation scenario; and
generating a script file for containing state transition logical relation
based on the state
transition matrix of the conversation scenario and generating a finite state
machine (FSM)
based on the script file, to return a corresponding pattern when the intended
scenario of
the user's conversation is recognized.
125. The method of any one of claims 117 to 124, further comprises:
preprocessing the user's conversation to obtain a plurality of segmented terms
in the
user's conversation;
performing mapping transformation on the segmented terms;
obtaining a feature word sequence of the user's conversation;
57

using the word vector space model to construct a feature vector of the user's
conversation
and scenario feature vectors of the various conversation scenarios based on
the feature
word sequence of the user's conversation and feature word sequences of the
various
conversation scenarios;
performing similarity calculation on the feature vector of the user's
conversation and the
scenario feature vectors of the various conversation scenarios; and
recognizing intention of the user's conversation based on a similarity
calculation result,
to return a pattern to which the intention corresponds.
126. The method of any one of claims 117 to 125, wherein the plurality of
conversation scenarios
are contained in a preset conversation scenario list, wherein the conversation
scenario list is
to record one or more conversation scenarios of a specific business field.
127. The method of any one of claims 117 to 126, wherein the sample corpora of
the various
conversation scenarios are obtained by classifying and marking specific domain
corpora
according to classes of the conversation scenarios.
128. The method of any one of claims 117 to 127, wherein the conversation
scenarios are
obtained by perfonning scenario abstraction on the specific domain corpora,
and the
scenario abstraction is a process from data to information and then to
knowledge.
129. The method of any one of claims 117 to 128, wherein each conversation
scenario is
abstracted as a conversation state, and dialogue process is abstracted as a
transition between
conversation states.
130. The method of any one of claims 117 to 129, wherein the conversation
state is taken as a
node, an oriented line between the conversation states is a process in which
one state is
transferred to another state, wherein entire dialogue process is abstracted as
a graph
consisting of nodes and oriented lines.
58

131. The method of any one of claims 117 to 130, wherein content
transformation based on a
word of bag (WOB) model is performed on the sample corpus of each conversation
scenario, wherein discrete representation of the sample corpus of each
conversation scenario
is obtained, wherein the feature selection algorithm is to extract feature
words of each
conversation scenario on the basis of the discrete representation of the
sample corpus of
each conversation scenario.
132. The method of any one of claims 117 to 131, wherein the WOB model divides
a corpus text
into separate terms, and it is imaged that all terms are placed in a bag, such
elements as their
word orders, grammars and syntaxes are ignored, wherein the bag is only
regarded as a
collection of plural vocabularies, wherein each term as it appears in the
corpus text is
independent.
133. The method of any one of claims 117 to 132, wherein the WOB model
includes a one-hot,
`11, -IDF and N-gram model.
134. The method of any one of claims 117 to 133, wherein name, the feature
words and the
feature word sequence of each conversation scenario are correspondingly stored
in the
scenario feature relation table.
135. The method of any one of claims 117 to 134, wherein the scenario feature
relation table
stores correspondence relations between conversation scenarios and scenario
features,
including feature words and feature word sequences.
136. The method of any one of claims 117 to 135, wherein the scenario feature
relation table is
stored in a server, maintained offline by backstage algorithm technical
personnel based on
periodical text data mining operations, and isolated from frontstage business
personnel.
137. The method of any one of claims 117 to 136, wherein when the intelligent
voice robot is
conversing with the user, obtain user's conversation text by recognizing and
transforming
user's conversation voice via an automatic speech recognition (ASR) technique,
and extract
feature information out of the user's conversation text.
59

138. The method of any one of claims 117 to 137, wherein pre-trained embedding
is obtained by
introducing large-scale BERT word vector space, pretrained in google bert
serving.
139. The method of any one of claims 117 to 138, wherein the BERT word vector
space is
retrained by bringing in own business customer service corpora, and
calibration of the
BERT word vector is realized, to conform to specific business scenarios.
140. The method of any one of claims 117 to 139, wherein total segmentation
corpora where
stop-words are removed are to construct the preset domain dictionary, wherein
the preset
domain dictionary includes total effective vocabularies appearing in the
corpora, and the
preset domain dictionary is to perform content transformation on all sample
corpora of a
target conversation scenario based on the WOB model to obtain the discrete
representation.
141. The method of any one of claims 117 to 140, wherein the chi-square
statistic (CHI)
technique is to extract feature words of the target conversation scenario.
142. The method of any one of claims 117 to 141, wherein CHI calculation
formula comprises:
<IMG>
wherein c is a certain class, namely a "conversation scenario", t is a certain
term, and N is
a total number of texts in a training corpora.
143. The method of claim 142, wherein the X2 is for chi square hypothesis
verification in
statistics, to judge uniformity or goodness of fit between actual distribution
and theoretical
distribution, wherein a null hypothesis 1/0 being "no marked difference
between observed
frequency and desired frequency".
144. The method of any one of claims 141 to 143, wherein less the chi-square
statistic is, the
closer the observed frequency to the desired frequency, and higher relevancy
between.
145. The method of any one of claims 141 to 144, wherein x2 is a measure of
distance between
an observed object and a desired object, wherein less the distance is, higher
the relevancy
between.

146. The method of any one of claims 141 to 143, wherein the observed object
is the tem, and
the desired object is the conversation scenario, wherein if the term and the
conversation
scenario are highly relevant, statistic distributions of the two are close to
each other in the
entire samples.
147. The method of any one of claims 141 to 143, wherein through x2 statistic,
relevancies
between all vocabularies in the domain dictionary and the various classes are
calculated
quickly and accurately based on quantities of corpora, and a preset number of
terms with the
x2 selected according to relevancy sorting result to serve as a feature set of
conversation
scenarios, to complete feature mapping of various scenarios and the various
classes in the
conversation scenario list.
148. The method of any one of claims 117 to 147, wherein the configuration
feature word input
by the user with respect to the target conversion scenario through a system
frontend.
149. The method of any one of claims 117 to 148, wherein the system frontend
provides business
personnel with a feature relation extension function to maintain the business
field.
150. The method of any one of claims 117 to 149, wherein on receiving the
configuration feature
word, updates it into an externally input feature set of the target conversion
scenario.
151. The method of any one of claims 117 to 150, wherein the externally input
configuration
feature word is not contained in the domain dictionary, the configuration
feature word is
ignored.
152. The method of any one of claims 117 to 151, wherein the first user is
business personnel,
the second user is an algorithm developing technical personnel.
153. The method of any one of claims 117 to 152, wherein the FSM is a
mathematical model that
expresses a finite number of states and behaviors as transitions and actions
amongst these
states.
61

154. The method of any one of claims 117 to 153, wherein the FSM is to
describe state sequences
of an object experienced within the objects life cycle, and how to respond to
various events
coming from outside to transition between the states.
155. The method of any one of claims 117 to 154, wherein the FSM includes
current state,
condition, action, and next state.
156. The method of any one of claims 117 to 155, wherein the current state is
a state currently
being in, wherein the condition is referred to as event, wherein the condition
is satisfied, an
action is triggered or one round of state transition is performed;
157. The method of any one of claims 117 to 156, wherein the action indicates
the action
executed after the condition has been satisfied, wherein the action has been
executed to
completion, a new state is transitioned, or the original state is maintained.
158. The method of any one of claims 117 to 157, wherein the action is not
indispensable, after
the condition has been satisfied, not execute any action and directly
transition to the new
state;
159. The method of any one of claims 117 to 158, wherein the next state is the
new state to be
transitioned after the condition has been satisfied.
160. The method of any one of claims 117 to 159, wherein the next state is
relative to the current
state, once the next state is activated, it becomes a new current state.
161. The method of any one of claims 117 to 160, wherein FSM model is
abstracted comprising:
<IMG>
162. The method of any one of claims 117 to 161, wherein the FSM state
transition torque
completely repaired is automatically translated to the script file of json
format.
163. The method of any one of claims 117 to 162, wherein the json script file
is read and input by
the program to a state machine object when a finite state machine instance is
generated, to
validate logic.
62

164. The method of any one of claims 117 to 163, wherein the finite state
machine instance as
generated is stored in Redis as an index value according to uuid transferred
by the frontend,
to facilitate subsequent program access when the interactive voice response
(IVR) service
starts.
165. The method of any one of claims 117 to 164, wherein the user performs
persistent operation
on the FSM.
166. The method of any one of claims 117 to 165, wherein the user selects a
task type as a single
task at the frontend, the finite state machine instance of the single task as
stored in Redis is
cleared away as an invalidated object within a preset time period after the
IVR marketing
service has been triggered.
167. The method of any one of claims 117 to 166, wherein the intelligent voice
robot converses
with the user, wherein the user's conversation is text content recognized and
transcribed
from user's conversation speech through ASR technology.
168. The method of any one of claims 117 to 167, wherein the text content is
word-segmented to
obtain the plurality of segmented terms, and word-segmentation process
includes character
purification, rectification, word segmentation, and removal of stop-words.
169. The method of any one of claims 117 to 168, wherein plural segmented
terms are mapped
and transformed through index of the domain dictionary to form of expression
of the
conversation scenario in the scenario feature relation table.
170. The method of any one of claims 117 to 169, wherein each element in the
feature word
sequence of the user's conversation is mapped to the BERT word vector space to
obtain the
feature vector of 768 dimensions, and all elements are summated and averaged
to obtain a
1x768 vector to serve as feature expression input by the user's conversation.
171. The method of any one of claims 117 to 170, wherein the feature word
sequences of the
various conversation scenarios in the scenario feature relation table are
operated is
converted to 1x768 feature vectors.
63

172. The method of any one of claims 117 to 171, wherein a cosine similarity
between the
feature vector input by the user's conversation and the scenario feature
vector of the
conversation scenario is calculated, wherein the greater cosine similarity
calculation result,
the higher similarity, and the higher cosine similarity calculation result is
the relevancy
between the user's conversation and the conversation scenario.
173. The method of any one of claims 117 to 172, wherein by arranging all the
conversation
scenarios in a descending order according to the cosine similarity calculation
result, the
conversation scenario with the highest cosine similarity calculation result is
returned to
serve as a judgement result of the intended scenario of current input by the
user, and a
corresponding response pattern under the intended scenario is returned
according to the state
in which the FSM is currently.
174. The method of any one of claims 117 to 173, wherein the response pattern
is transformed to
speech content through Text-To-Speech (TTS) technology is broadcast to the
user.
175.A computer equipment comprising:
a memory, including a nonvolatile storage medium and an internal memory,
wherein the
nonvolatile storage medium stores therein an operating system, and a computer
program,
wherein the internal memory provides environment for running of the operating
system
and the computer program in the nonvolatile storage medium.
a processor, configured to provide computing and controlling capabilities;
a network interface connected to each other via a system bus configured to
connect to
other equipment via network for communication;
a computer program stored on the memory and operable on the processor, wherein
the
processor executes the computer program configured to:
obtain sample corpora of various conversation scenarios in a plurality of
conversation scenarios;
64

generate a scenario feature of the conversation scenario based on a sample
corpus of the conversation scenario, wherein the scenario feature includes
feature words of the conversation scenario and a feature word sequence
obtained by mapping transformation of the feature words; and
configure an intelligent voice robot with a preset word vector space model and
the scenario features of the various conversation scenarios, wherein the word
vector space model is for the intelligent voice robot to perform word vector
similarity calculation on a user's conversation and the scenario features of
the
various conversation scenarios, to recognize an intended scenario of the
user's
conversation.
176. The equipment of claim 175, wherein generating the scenario feature of
the conversation
scenario based on the sample corpus of the conversation scenario comprises:
obtaining discrete representation of the sample corpus of the conversation
scenario based
on a preset domain dictionary;
employing a feature selection algorithm to extract feature words of the
conversation
scenario based on the discrete representation of the sample corpus of the
conversation
scenario;
mapping and transforming the feature words of the conversation scenario to a
corresponding dictionary index; and
generating the feature word sequence of the conversation scenario, wherein the
feature
selection algorithm is a chi-square statistic feature selection algorithm.
177. The equipment of claim 175, further comprises:
storing the various conversation scenarios and the scenario features of the
conversation
scenarios in a scenario feature relation table.
178. The equipment of claim 175, further comprises:

receiving a configuration feature word input with respect to the conversation
scenario;
and
maintaining the scenario feature of the conversation scenario in the scenario
feature
relation table based on the configuration feature word of the conversation
scenario and a
configuration feature word sequence obtained by mapping transformation of the
configuration feature word.
179. The equipment of claim 178, further comprises receiving the configuration
feature word
input by the user having feature configuration permission with respect to the
conversation
scenario.
180. The equipment of claim 178, wherein maintaining the scenario feature of
the conversation
scenario in the scenario feature relation table based on the configuration
feature word of the
conversation scenario and the configuration feature word sequence obtained by
mapping
transformation of the configuration feature word comprises:
merging the configuration feature word of the conversation scenario to the
feature words
of the conversation scenario in the scenario feature relation table; and
adding the configuration feature word sequence of a merged configuration
feature word
to the feature word sequence of the conversation scenario.
181. The equipment of claim 175, wherein the word vector space model is
trained and obtained
comprises:
applying a domain corpora of a domain to the various conversation scenarios
pertaining
to train a pre-trained bidirectional encoder representations from transformers
(BERT)
word vector space; and
obtaining the word vector space model.
182. The equipment of claim 175, further comprises:
66

receiving a state transition graph input by a first user for the conversation
scenario and
receiving supplementary information input by a second user for a state
transition graph,
to generate a state transition matrix of the conversation scenario; and
generating a script file for containing state transition logical relation
based on the state
transition matrix of the conversation scenario and generating a finite state
machine (FSM)
based on the script file, to return a corresponding pattern when the intended
scenario of
the user's conversation is recognized.
183. The equipment of any one of claims 175 to 182, further comprises:
preprocessing the user's conversation to obtain a plurality of segmented terms
in the
user's conversation;
performing mapping transformation on the segmented terms;
obtaining a feature word sequence of the user's conversation;
using the word vector space model to construct a feature vector of the user's
conversation
and scenario feature vectors of the various conversation scenarios based on
the feature
word sequence of the user's conversation and feature word sequences of the
various
conversation scenarios;
performing similarity calculation on the feature vector of the user's
conversation and the
scenario feature vectors of the various conversation scenarios; and
recognizing intention of the user's conversation based on a similarity
calculation result,
to return a pattern to which the intention corresponds.
184. The equipment of any one of claims 175 to 183, wherein the plurality of
conversation
scenarios are contained in a preset conversation scenario list, wherein the
conversation
scenario list is to record one or more conversation scenarios of a specific
business field.
67

185. The equipment of any one of claims 175 to 184, wherein the sample corpora
of the various
conversation scenarios are obtained by classifying and marking specific domain
corpora
according to classes of the conversation scenarios.
186. The equipment of any one of claims 175 to 185, wherein the conversation
scenarios are
obtained by performing scenario abstraction on the specific domain corpora,
and the
scenario abstraction is a process from data to information and then to
knowledge.
187. The equipment of any one of claims 175 to 186, wherein each conversation
scenario is
abstracted as a conversation state, and dialogue process is abstracted as a
transition between
conversation states.
188. The equipment of any one of claims 175 to 187, wherein the conversation
state is taken as a
node, an oriented line between the conversation states is a process in which
one state is
transferred to another state, wherein entire dialogue process is abstracted as
a graph
consisting of nodes and oriented lines.
189. The equipment of any one of claims 175 to 188, wherein content
transformation based on a
word of bag (WOB) model is performed on the sample corpus of each conversation
scenario, wherein discrete representation of the sample corpus of each
conversation scenario
is obtained, wherein the feature selection algorithm is to extract feature
words of each
conversation scenario on the basis of the discrete representation of the
sample corpus of
each conversation scenario.
190. The equipment of any one of claims 175 to 189, wherein the WOB model
divides a corpus
text into separate terrns, and it is imaged that all terms are placed in a
bag, such elements as
their word orders, grammars and syntaxes are ignored, wherein the bag is only
regarded as a
collection of plural vocabularies, wherein each term as it appears in the
corpus text is
independent.
191. The equipment of any one of claims 175 to 190, wherein the WOB model
includes a one-
hot, TF-IDF and N-gram model.
68

192. The equipment of any one of claims 175 to 191, wherein name, the feature
words and the
feature word sequence of each conversation scenario are correspondingly stored
in the
scenario feature relation table.
193. The equipment of any one of claims 175 to 192, wherein the scenario
feature relation table
stores correspondence relations between conversation scenarios and scenario
features,
including feature words and feature word sequences.
194. The equipment of any one of claims 175 to 193, wherein the scenario
feature relation table
is stored in a server, maintained offline by backstage algorithm technical
personnel based on
periodical text data mining operations, and isolated from frontstage business
personnel.
195. The equipment of any one of claims 175 to 194, wherein when the
intelligent voice robot is
conversing with the user, obtain user's conversation text by recognizing and
transforming
user's conversation voice via an automatic speech recognition (ASR) technique,
and extract
feature information out of the user's conversation text.
196. The equipment of any one of claims 175 to 195, wherein pre-trained
embedding is obtained
by introducing large-scale BERT word vector space, pretrained in google bert
serving.
197. The equipment of any one of claims 175 to 196, wherein the BERT word
vector space is
retrained by bringing in own business customer service corpora, and
calibration of the
BERT word vector is realized, to conform to specific business scenarios.
198. The equipment of any one of claims 175 to 197, wherein total segmentation
corpora where
stop-words are removed are to construct the preset domain dictionary, wherein
the preset
domain dictionary includes total effective vocabularies appearing in the
corpora, and the
preset domain dictionary is to perform content transformation on all sample
corpora of a
target conversation scenario based on the WOB model to obtain the discrete
representation.
199. The equipment of any one of claims 175 to 198, wherein the chi-square
statistic (CHI)
technique is to extract feature words of the target conversation scenario.
69

200. The equipment of any one of claims 175 to 199, wherein CHI calculation
formula
comprises:
<IMG>
wherein c is a certain class, namely a "conversation scenario", t is a certain
term, and N is
a total number of texts in a training corpora.
201. The equipment of claim 200, wherein the x2 is for chi square hypothesis
verification in
statistics, to judge uniformity or goodness of fit between actual distribution
and theoretical
distribution, wherein a null hypothesis Ho being "no marked difference between
observed
frequency and desired frequency".
202. The equipment of any one of claims 200 to 201, wherein less the chi-
square statistic is, the
closer the observed frequency to the desired frequency, and higher relevancy
between.
203. The equipment of any one of claims 200 to 202, wherein x2 is a measure of
distance
between an observed object and a desired object, wherein less the distance is,
higher the
relevancy between.
204. The equipment of any one of claims 200 to 203, wherein the observed
object is the term,
and the desired object is the conversation scenario, wherein if the term and
the conversation
scenario are highly relevant, statistic distributions of the two are close to
each other in the
entire samples.
205. The equipment of any one of claims 200 to 204, wherein through x2
statistic, relevancies
between all vocabularies in the domain dictionary and the various classes are
calculated
quickly and accurately based on quantities of corpora, and a preset number of
terms with the
x2 selected according to relevancy sorting result to serve as a feature set of
conversation
scenarios, to complete feature mapping of various scenarios and the various
classes in the
conversation scenario list.

206. The equipment of any one of claims 175 to 205, wherein the configuration
feature word
input by the user with respect to the target conversion scenario through a
system frontend.
207. The equipment of any one of claims 175 to 206, wherein the system
frontend provides
business personnel with a feature relation extension function to maintain the
business field.
208. The equipment of any one of claims 175 to 207, wherein on receiving the
configuration
feature word, updates it into an externally input feature set of the target
conversion scenario.
209. The equipment of any one of claims 175 to 208, wherein the externally
input configuration
feature word is not contained in the domain dictionary, the configuration
feature word is
ignored.
210. The equipment of any one of claims 175 to 209, wherein the first user is
business personnel,
the second user is an algorithm developing technical personnel.
211. The equipment of any one of claims 175 to 210, wherein the FSIVI is a
mathematical model
that expresses a finite number of states and behaviors as transitions and
actions amongst
these states.
212. The equipment of any one of claims 175 to 211, wherein the FSIVI is to
describe state
sequences of an object experienced within the objects life cycle, and how to
respond to
various events coming from outside to transition between the states.
213. The equipment of any one of claims 175 to 212, wherein the FSIVI includes
current state,
condition, action, and next state.
214. The equipment of any one of claims 175 to 213, wherein the current state
is a state currently
being in, wherein the condition is referred to as event, wherein the condition
is satisfied, an
action is triggered or one round of state transition is performed;
215. The equipment of any one of claims 175 to 214, wherein the action
indicates the action
executed after the condition has been satisfied, wherein the action has been
executed to
completion, a new state is transitioned, or the original state is maintained.
71

216. The equipment of any one of claims 175 to 215, wherein the action is not
indispensable,
after the condition has been satisfied, not execute any action and directly
transition to the
new state;
217. The equipment of any one of claims 175 to 216, wherein the next state is
the new state to be
transitioned after the condition has been satisfied.
218. The equipment of any one of claims 175 to 217, wherein the next state is
relative to the
current state, once the next state is activated, it becomes a new current
state.
219. The equipment of any one of claims 175 to 218, wherein FSM model is
abstracted
comprising:
<IMG>
220. The equipment of any one of claims 175 to 219, wherein the FSM state
transition torque
completely repaired is automatically translated to the script file of json
format.
221. The equipment of any one of claims 175 to 220, wherein the json script
file is read and input
by the program to a state machine object when a finite state machine instance
is generated,
to validate logic.
222. The equipment of any one of claims 175 to 221, wherein the finite state
machine instance as
generated is stored in Redis as an index value according to uuid transferred
by the frontend,
to facilitate subsequent program access when the interactive voice response
(IVR) service
starts.
223. The equipment of any one of claims 175 to 222, wherein the user performs
persistent
operation on the FSM.
224. The equipment of any one of claims 175 to 223, wherein the user selects a
task type as a
single task at the frontend, the finite state machine instance of the single
task as stored in
Redis is cleared away as an invalidated object within a preset time period
after the IVR
marketing service has been triggered.
72

225. The equipment of any one of claims 175 to 224, wherein the intelligent
voice robot
converses with the user, wherein the user's conversation is text content
recognized and
transcribed from user's conversation speech through ASR technology.
226. The equipment of any one of claims 175 to 225, wherein the text content
is word-segmented
to obtain the plurality of segmented terms, and word-segmentation process
includes
character purification, rectification, word segmentation, and removal of stop-
words.
227. The equipment of any one of claims 175 to 226, wherein plural segmented
terms are
mapped and transforrned through index of the domain dictionary to form of
expression of
the conversation scenario in the scenario feature relation table.
228. The equipment of any one of claims 175 to 227, wherein each element in
the feature word
sequence of the user's conversation is mapped to the BERT word vector space to
obtain the
feature vector of 768 dimensions, and all elements are summated and averaged
to obtain a
1x768 vector to serve as feature expression input by the user's conversation.
229. The equipment of any one of claims 175 to 228, wherein the feature word
sequences of the
various conversation scenarios in the scenario feature relation table are
operated is
converted to 1x768 feature vectors.
230. The equipment of any one of claims 175 to 229, wherein a cosine
similarity between the
feature vector input by the user's conversation and the scenario feature
vector of the
conversation scenario is calculated, wherein the greater cosine similarity
calculation result,
the higher similarity, and the higher cosine similarity calculation result is
the relevancy
between the user's conversation and the conversation scenario.
231. The equipment of any one of claims 175 to 230, wherein by arranging all
the conversation
scenarios in a descending order according to the cosine similarity calculation
result, the
conversation scenario with the highest cosine similarity calculation result is
returned to
serve as a judgement result of the intended scenario of current input by the
user, and a
corresponding response pattern under the intended scenario is returned
according to the state
in which the FSM is currently.
73

232. The equipment of any one of claims 175 to 231, wherein the response
pattern is transformed
to speech content through Text-To-Speech (TTS) technology is broadcast to the
user.
233.A computer readable physical memory having stored thereon a computer
program executed
by a computer configured to:
obtain sample corpora of various conversation scenarios in a plurality of
conversation
scenarios;
generate a scenario feature of the conversation scenario based on a sample
corpus of the
conversation scenario, wherein the scenario feature includes feature words of
the
conversation scenario and a feature word sequence obtained by mapping
transformation
of the feature words; and
configure an intelligent voice robot with a preset word vector space model and
the
scenario features of the various conversation scenarios, wherein the word
vector space
model is for the intelligent voice robot to perform word vector similarity
calculation on a
user's conversation and the scenario features of the various conversation
scenarios, to
recognize an intended scenario of the user's conversation.
234. The memory of claim 233, wherein generating the scenario feature of the
conversation
scenario based on the sample corpus of the conversation scenario comprises:
obtaining discrete representation of the sample corpus of the conversation
scenario based
on a preset domain dictionary;
employing a feature selection algorithm to extract feature words of the
conversation
scenario based on the discrete representation of the sample corpus of the
conversation
scenario;
mapping and transforming the feature words of the conversation scenario to a
corresponding dictionary index; and
74

generating the feature word sequence of the conversation scenario, wherein the
feature
selection algorithm is a chi-square statistic feature selection algorithm.
235. The memory of claim 233, further comprises:
storing the various conversation scenarios and the scenario features of the
conversation
scenarios in a scenario feature relation table.
236. The memory of claim 233, further comprises:
receiving a configuration feature word input with respect to the conversation
scenario;
and
maintaining the scenario feature of the conversation scenario in the scenario
feature
relation table based on the configuration feature word of the conversation
scenario and a
configuration feature word sequence obtained by mapping transformation of the
configuration feature word.
237. The memory of claim 236, further comprises receiving the configuration
feature word input
by the user having feature configuration permission with respect to the
conversation
scenario.
238. The memory of claim 236, wherein maintaining the scenario feature of the
conversation
scenario in the scenario feature relation table based on the configuration
feature word of the
conversation scenario and the configuration feature word sequence obtained by
mapping
transformation of the configuration feature word comprises:
merging the configuration feature word of the conversation scenario to the
feature words
of the conversation scenario in the scenario feature relation table; and
adding the configuration feature word sequence of a merged configuration
feature word
to the feature word sequence of the conversation scenario.
239. The memory of claim 233, wherein the word vector space model is trained
and obtained
comprises:

applying a domain corpora of a domain to the various conversation scenarios
pertaining
to train a pre-trained bidirectional encoder representations from transformers
(BERT)
word vector space; and
obtaining the word vector space model.
240. The memory of claim 233, further comprises:
receiving a state transition graph input by a first user for the conversation
scenario and
receiving supplementary information input by a second user for a state
transition graph,
to generate a state transition matrix of the conversation scenario; and
generating a script file for containing state transition logical relation
based on the state
transition matrix of the conversation scenario and generating a finite state
machine (FSM)
based on the script file, to return a corresponding pattern when the intended
scenario of
the user's conversation is recognized.
241. The memory of any one of claims 233 to 240, further comprises:
preprocessing the user's conversation to obtain a plurality of segmented terms
in the
user's conversation;
performing mapping transformation on the segmented terms;
obtaining a feature word sequence of the user's conversation;
using the word vector space model to construct a feature vector of the user's
conversation
and scenario feature vectors of the various conversation scenarios based on
the feature
word sequence of the user's conversation and feature word sequences of the
various
conversation scenarios;
performing similarity calculation on the feature vector of the user's
conversation and the
scenario feature vectors of the various conversation scenarios; and
76

recognizing intention of the user's conversation based on a similarity
calculation result,
to return a pattern to which the intention corresponds.
242. The memory of any one of claims 233 to 241, wherein the plurality of
conversation
scenarios are contained in a preset conversation scenario list, wherein the
conversation
scenario list is to record one or more conversation scenarios of a specific
business field.
243. The memory of any one of claims 233 to 242, wherein the sample corpora of
the various
conversation scenarios are obtained by classifying and marking specific domain
corpora
according to classes of the conversation scenarios.
244. The memory of any one of claims 233 to 243, wherein the conversation
scenarios are
obtained by perfonning scenario abstraction on the specific domain corpora,
and the
scenario abstraction is a process from data to information and then to
knowledge.
245. The memory of any one of claims 233 to 244, wherein each conversation
scenario is
abstracted as a conversation state, and dialogue process is abstracted as a
transition between
conversation states.
246. The memory of any one of claims 233 to 245, wherein the conversation
state is taken as a
node, an oriented line between the conversation states is a process in which
one state is
transferred to another state, wherein entire dialogue process is abstracted as
a graph
consisting of nodes and oriented lines.
247. The memory of any one of claims 233 to 246, wherein content
transformation based on a
word of bag (WOB) model is performed on the sample corpus of each conversation
scenario, wherein discrete representation of the sample corpus of each
conversation scenario
is obtained, wherein the feature selection algorithm is to extract feature
words of each
conversation scenario on the basis of the discrete representation of the
sample corpus of
each conversation scenario.
77

248. The memory of any one of claims 233 to 247, wherein the WOB model divides
a corpus text
into separate terms, and it is imaged that all terms are placed in a bag, such
elements as their
word orders, grammars and syntaxes are ignored, wherein the bag is only
regarded as a
collection of plural vocabularies, wherein each term as it appears in the
corpus text is
independent.
249. The memory of any one of claims 233 to 248, wherein the WOB model
includes a one-hot,
`11,-IDF and N-gram model.
250. The memory of any one of claims 233 to 249, wherein name, the feature
words and the
feature word sequence of each conversation scenario are correspondingly stored
in the
scenario feature relation table.
251. The memory of any one of claims 233 to 250, wherein the scenario feature
relation table
stores correspondence relations between conversation scenarios and scenario
features,
including feature words and feature word sequences.
252. The memory of any one of claims 233 to 251, wherein the scenario feature
relation table is
stored in a server, maintained offline by backstage algorithm technical
personnel based on
periodical text data mining operations, and isolated from frontstage business
personnel.
253. The memory of any one of claims 233 to 252, wherein when the intelligent
voice robot is
conversing with the user, obtain user's conversation text by recognizing and
transforming
user's conversation voice via an automatic speech recognition (ASR) technique,
and extract
feature information out of the user's conversation text.
254. The memory of any one of claims 233 to 253, wherein pre-trained embedding
is obtained by
introducing large-scale BERT word vector space, pretrained in google bert
serving.
255. The memory of any one of claims 233 to 254, wherein the BERT word vector
space is
retrained by bringing in own business customer service corpora, and
calibration of the
BERT word vector is realized, to conform to specific business scenarios.
78

256. The memory of any one of claims 233 to 255, wherein total segmentation
corpora where
stop-words are removed are to construct the preset domain dictionary, wherein
the preset
domain dictionary includes total effective vocabularies appearing in the
corpora, and the
preset domain dictionary is to perform content transformation on all sample
corpora of a
target conversation scenario based on the WOB model to obtain the discrete
representation.
257. The memory of any one of claims 233 to 256, wherein the chi-square
statistic (CHI)
technique is to extract feature words of the target conversation scenario.
258. The memory of any one of claims 233 to 257, wherein CHI calculation
formula comprises:
<IMG>
wherein c is a certain class, namely a "conversation scenario", t is a certain
term, and N is
a total number of texts in a training corpora.
259. The memory of claim 258, wherein the x2 is for chi square hypothesis
verification in
statistics, to judge uniformity or goodness of fit between actual distribution
and theoretical
distribution, wherein a null hypothesis H 0 being "no marked difference
between observed
frequency and desired frequency".
260. The memory of any one of claims 258 to 259, wherein less the chi-square
statistic is, the
closer the observed frequency to the desired frequency, and higher relevancy
between.
261. The memory of any one of claims 258 to 260, wherein X2 is a measure of
distance between
an observed object and a desired object, wherein less the distance is, higher
the relevancy
between.
262. The memory of any one of claims 258 to 261, wherein the observed object
is the term, and
the desired object is the conversation scenario, wherein if the term and the
conversation
scenario are highly relevant, statistic distributions of the two are close to
each other in the
entire samples.
79

263. The memory of any one of claims 258 to 262, wherein through X2 statistic,
relevancies
between all vocabularies in the domain dictionary and the various classes are
calculated
quickly and accurately based on quantities of corpora, and a preset number of
temis with the
x2 selected according to relevancy sorting result to serve as a feature set of
conversation
scenarios, to complete feature mapping of various scenarios and the various
classes in the
conversation scenario list.
264. The memory of any one of claims 233 to 263, wherein the configuration
feature word input
by the user with respect to the target conversion scenario through a system
frontend.
265. The memory of any one of claims 233 to 264, wherein the system frontend
provides
business personnel with a feature relation extension function to maintain the
business field.
266. The memory of any one of claims 233 to 265, wherein on receiving the
configuration
feature word, updates it into an externally input feature set of the target
conversion scenario.
267. The memory of any one of claims 233 to 266, wherein the externally input
configuration
feature word is not contained in the domain dictionary, the configuration
feature word is
ignored.
268. The memory of any one of claims 233 to 267, wherein the first user is
business personnel,
the second user is an algorithm developing technical personnel.
269. The memory of any one of claims 233 to 268, wherein the FSM is a
mathematical model
that expresses a finite number of states and behaviors as transitions and
actions amongst
these states.
270. The memory of any one of claims 233 to 269, wherein the FSM is to
describe state
sequences of an object experienced within the objects life cycle, and how to
respond to
various events coming from outside to transition between the states.
271. The memory of any one of claims 233 to 270, wherein the FSM includes
current state,
condition, action, and next state.

272. The memory of any one of claims 233 to 271, wherein the current state is
a state currently
being in, wherein the condition is referred to as event, wherein the condition
is satisfied, an
action is triggered or one round of state transition is performed;
273. The memory of any one of claims 233 to 272, wherein the action indicates
the action
executed after the condition has been satisfied, wherein the action has been
executed to
completion, a new state is transitioned, or the original state is maintained.
274. The memory of any one of claims 233 to 273, wherein the action is not
indispensable, after
the condition has been satisfied, not execute any action and directly
transition to the new
state;
275. The memory of any one of claims 233 to 274, wherein the next state is the
new state to be
transitioned after the condition has been satisfied.
276. The memory of any one of claims 233 to 275, wherein the next state is
relative to the current
state, once the next state is activated, it becomes a new current state.
277. The memory of any one of claims 233 to 276, wherein FSM model is
abstracted comprising:
<IMG>
278. The memory of any one of claims 233 to 277, wherein the FSM state
transition torque
completely repaired is automatically translated to the script file of json
format.
279. The memory of any one of claims 233 to 278, wherein the json script file
is read and input
by the program to a state machine object when a finite state machine instance
is generated,
to validate logic.
280. The memory of any one of claims 233 to 279, wherein the finite state
machine instance as
generated is stored in Redis as an index value according to uuid transferred
by the frontend,
to facilitate subsequent program access when the interactive voice response
(IVR) service
starts.
81

281. The memory of any one of claims 233 to 280, wherein the user performs
persistent
operation on the FSIVI.
282. The memory of any one of claims 233 to 281, wherein the user selects a
task type as a single
task at the frontend, the finite state machine instance of the single task as
stored in Redis is
cleared away as an invalidated object within a preset time period after the
IVR marketing
service has been triggered.
283. The memory of any one of claims 233 to 282, wherein the intelligent voice
robot converses
with the user, wherein the user's conversation is text content recognized and
transcribed
from user's conversation speech through ASR technology.
284. The memory of any one of claims 233 to 283, wherein the text content is
word-segmented to
obtain the plurality of segmented terms, and word-segmentation process
includes character
purification, rectification, word segmentation, and removal of stop-words.
285. The memory of any one of claims 233 to 284, wherein plural segmented
terms are mapped
and transformed through index of the domain dictionary to form of expression
of the
conversation scenario in the scenario feature relation table.
286. The memory of any one of claims 233 to 285, wherein each element in the
feature word
sequence of the user's conversation is mapped to the BERT word vector space to
obtain the
feature vector of 768 dimensions, and all elements are summated and averaged
to obtain a
1x768 vector to serve as feature expression input by the user's conversation.
287. The memory of any one of claims 233 to 286, wherein the feature word
sequences of the
various conversation scenarios in the scenario feature relation table are
operated is
converted to 1x768 feature vectors.
82

288. The memory of any one of claims 233 to 287, wherein a cosine similarity
between the
feature vector input by the user's conversation and the scenario feature
vector of the
conversation scenario is calculated, wherein the greater cosine similarity
calculation result,
the higher similarity, and the higher cosine similarity calculation result is
the relevancy
between the user's conversation and the conversation scenario.
289. The memory of any one of claims 233 to 288, wherein by arranging all the
conversation
scenarios in a descending order according to the cosine similarity calculation
result, the
conversation scenario with the highest cosine similarity calculation result is
returned to
serve as a judgement result of the intended scenario of current input by the
user, and a
corresponding response pattern under the intended scenario is returned
according to the state
in which the FSM is currently.
290. The memory of any one of claims 233 to 289, wherein the response pattern
is transformed to
speech content through Text-To-Speech (TTS) technology is broadcast to the
user.
83

Description

METHOD OF REALIZING CONFIGURABLE INTELLIGENT VOICE ROBOT,
DEVICE AND STORAGE MEDIUM
BACKGROUND OF THE INVENTION
Technical Field
The present application relates to the field of artificial intelligence, and
more particularly to a
method of realizing a configurable intelligent voice robot, and corresponding
device and storage
medium.
Description of Related Art
At present, intelligent voice robots are widely applied in telemarketing and
customer service
systems, and telemarketing on designated clients is completed by robots
through IVR (interactive
voice response) services. Telemarketing includes a simple form and a
complicated form, by the
simple form is meant the "one-sided marketing promotion", such as a broadcast
of marketing
content in one sentence, by the complicated form can be further realized
several rounds of voiced
questioning and answering (hereinafter referred to as "Q&A") interactions with
clients, as well
as judgement and feedback of clients' potential intentions. The intelligent
voice robots can
address problems perplexing the field for long in which the traditional human
telemarketing is
high in recruitment cost, long in training cycle, inconsistent in business
levels, and instable in
service qualities, and so on, as large-scale, repetitive works are now
completed by backstage
robots based on natural language models, and nearly 80% of general outbound
manpower cost is
helped to be reduced thereby for enterprises.
However, as the inventor has found during the process of achieving the present
invention,
currently available intelligent voice robots that can make the complicated
form of scenario Q&A
interactions are generally technically problematic as discussed below.
1
Date Recue/Date Received 2022-04-19

During the process of a scenario Q&A interaction, the intelligent voice robot
can make a
corresponding reply by recognizing the intended scenario in a user's
conversation. What the
traditional intention recognition algorithm, which is based on the text
classification model,
employs is the offline training mode, by which the model judges a sample newly
added in the
future by learning historical label corpora, and classifies the sample in a
certain label class already
learnt. Such an algorithmic model that bases on the Bayesian theory can only
process known
classifications, and would classify any class that never appeared in the
historical corpora still in
the known classifications, thus necessarily leading to classification error.
In order to deal with
this problem in practical application, the developing personnel can generally
only add marked
corpora of new classes to the historical label corpora, and thereafter roll
back again to train the
model. Such a mode is not only low in efficiency and high in cost, but can
also not guarantee
forward convergence of the model, that is to say, the introduction of new
corpora to relearn might
lower the classification precision of the original classes; moreover, since
the algorithm relies on
the conditional probability at the bottom, and discriminant probabilities of
the various classes are
not mutually independent, thusly, when intelligent voice robots are developed
for different
scenario classes, customized developments are respectively required, and
migration and reuse
are impossible therebetween, so that the development cost of scenarios for the
intelligent voice
robots is rendered unduly high.
SUMMARY OF THE INVENTION
In order to overcome the technical problems mentioned in the above Description
of Related Art,
the present application provides a method of realizing a configurable
intelligent voice robot, and
corresponding device and storage medium. The technical solutions of the
present application are
as follows.
According to the first aspect, there is provided a method of realizing a
configurable intelligent
voice robot, and the method comprises:
2
Date Recue/Date Received 2022-04-19

obtaining sample corpora of various conversation scenarios in a plurality of
conversation
scenarios;
generating, with respect to each said conversation scenario, a scenario
feature of the conversation
scenario on the basis of the sample corpus of the conversation scenario,
wherein the scenario
feature includes feature words of the conversation scenario and a feature word
sequence obtained
by mapping transformation of the feature words; and
configuring the intelligent voice robot on the basis of a preset word vector
space model and
scenario features of the various conversation scenarios, wherein the word
vector space model is
employed for the intelligent voice robot to perform word vector similarity
calculation on a user's
conversation and the scenario features of the various conversation scenarios,
so as to recognize
an intended scenario of the user's conversation.
Further, the step of generating, with respect to each said conversation
scenario, a scenario feature
of the conversation scenario on the basis of the sample corpus of the
conversation scenario
includes:
obtaining, with respect to each said conversation scenario, discrete
representation of the sample
corpus of the conversation scenario on the basis of a preset domain
dictionary;
employing a feature selection algorithm to extract feature words of the
conversation scenario on
the basis of the discrete representation of the sample corpus of the
conversation scenario; and
mapping to transform the features words of the conversation scenario to a
corresponding
dictionary index, and generating the feature word sequence of the conversation
scenario;
preferably, the feature selection algorithm is a chi-square statistic feature
selection algorithm.
Moreover, the method further comprises:
storing the various conversation scenarios and the scenario features of the
conversation scenarios
in a scenario feature relation table;
preferably, the method further comprises:
receiving a configuration feature word input with respect to any conversation
scenario; and
maintaining the scenario feature of the conversation scenario in the scenario
feature relation table
3
Date Recue/Date Received 2022-04-19

on the basis of the configuration feature word of the conversation scenario
and a configuration
feature word sequence obtained by mapping transformation of the configuration
feature word.
Preferably, the step of receiving a configuration feature word input with
respect to any
conversation scenario includes:
receiving a configuration feature word input by a user having feature
configuration permission
with respect to the conversation scenario.
Further, the step of maintaining the scenario feature of the conversation
scenario in the scenario
feature relation table on the basis of the configuration feature word of the
conversation scenario
and a configuration feature word sequence obtained by mapping transformation
of the
configuration feature word includes:
merging the configuration feature word of the conversation scenario to the
feature words of the
conversation scenario in the scenario feature relation table, and adding the
configuration feature
word sequence of the merged configuration feature word to the feature word
sequence of the
conversation scenario.
Further, the word vector space model is trained and obtained by:
employing domain corpora of a domain to which the various conversation
scenarios pertain to
train a pre-trained BERT word vector space, and obtaining the word vector
space model.
Moreover, the method further comprises:
receiving, with respect to any conversation scenario, a state transition graph
input by a first user
for the conversation scenario, and receiving supplementary information input
by a second user
for the state transition graph, to generate a state transition matrix of the
conversation scenario;
and
generating a script file for containing state transition logical relation on
the basis of the state
transition matrix of the conversation scenario, and generating a finite state
machine on the basis
of the script file, to return a corresponding pattern when the intended
scenario of the user's
4
Date Recue/Date Received 2022-04-19

conversation is recognized.
Moreover, the method further comprises:
preprocessing the user's conversation to obtain a plurality of segmented terms
in the user's
conversation after the well-configured intelligent voice robot has received
the user's conversation,
performing mapping transformation on the plural segmented terms, and obtaining
a feature word
sequence of the user's conversation;
employing the word vector space model to construct a feature vector of the
user's conversation
and scenario feature vectors of the various conversation scenarios on the
basis of the feature word
sequence of the user's conversation and feature word sequences of the various
conversation
scenarios; and
performing similarity calculation on the feature vector of the user's
conversation and the scenario
feature vectors of the various conversation scenarios, and recognizing
intention of the user's
conversation on the basis of a similarity calculation result, to return a
pattern to which the
intention corresponds.
According to the second aspect, there is provided a device for realizing a
configurable intelligent
voice robot, and the device comprises:
an obtaining module, for obtaining sample corpora of various conversation
scenarios in a
plurality of conversation scenarios;
a generating module, for generating, with respect to each said conversation
scenario, a scenario
feature of the conversation scenario on the basis of the sample corpus of the
conversation scenario,
wherein the scenario feature includes feature words of the conversation
scenario and a feature
word sequence obtained by mapping transformation of the feature words; and
a configuring module, for configuring the intelligent voice robot on the basis
of a preset word
vector space model and scenario features of the various conversation
scenarios, wherein the word
vector space model is employed for the intelligent voice robot to perform word
vector similarity
calculation on a user's conversation and the scenario features of the various
conversation
scenarios, so as to recognize an intended scenario of the user's conversation.
Date Recue/Date Received 2022-04-19

Further, the generating module includes:
a representing unit, for obtaining, with respect to each said conversation
scenario, discrete
representation of the sample corpus of the conversation scenario on the basis
of a preset domain
dictionary;
a screening unit, for employing a feature selection algorithm to extract
feature words of the
conversation scenario on the basis of the discrete representation of the
sample corpus of the
conversation scenario; and
a generating unit, for mapping to transform the features words of the
conversation scenario to a
corresponding dictionary index, and generating the feature word sequence of
the conversation
scenario;
preferably, the feature selection algorithm is a chi-square statistic feature
selection algorithm.
Moreover, the device further comprises:
a storing module, for storing the various conversation scenarios and the
scenario features of the
conversation scenarios in a scenario feature relation table;
preferably, the device further comprises:
a receiving module, for receiving a configuration feature word input with
respect to any
conversation scenario; and
a maintaining module, for maintaining the scenario feature of the conversation
scenario in the
scenario feature relation table on the basis of the configuration feature word
of the conversation
scenario and a configuration feature word sequence obtained by mapping
transformation of the
configuration feature word.
Preferably, the receiving module is employed for receiving a configuration
feature word input by
a user having feature configuration permission with respect to the
conversation scenario.
Further, the maintaining module is employed for merging the configuration
feature word of the
conversation scenario to the feature words of the conversation scenario in the
scenario feature
6
Date Recue/Date Received 2022-04-19

relation table, and adding the configuration feature word sequence of the
merged configuration
feature word to the feature word sequence of the conversation scenario.
Further, the device further comprises a training module for employing domain
corpora of a
domain to which the various conversation scenarios pertain to train a pre-
trained BERT word
vector space, and obtaining the word vector space model.
Moreover, the device further comprises a state machine configuring module for:
receiving, with respect to any conversation scenario, a state transition graph
input by a first user
for the conversation scenario, and receiving supplementary information input
by a second user
for the state transition graph, to generate a state transition matrix of the
conversation scenario;
and
generating a script file for containing state transition logical relation on
the basis of the state
transition matrix of the conversation scenario, and generating a finite state
machine on the basis
of the script file, to return a corresponding pattern when the intended
scenario of the user's
conversation is recognized.
Moreover, the device further comprises an intended scenario recognizing module
that includes:
an obtaining unit, for preprocessing the user's conversation to obtain a
plurality of segmented
terms in the user's conversation after the well-configured intelligent voice
robot has received the
user's conversation, performing mapping transformation on the plural segmented
terms, and
obtaining a feature word sequence of the user's conversation;
a constructing unit, for employing the word vector space model to construct a
feature vector of
the user's conversation and scenario feature vectors of the various
conversation scenarios on the
basis of the feature word sequence of the user's conversation and feature word
sequences of the
various conversation scenarios; and
a matching unit, for performing similarity calculation on the feature vector
of the user's
conversation and the scenario feature vectors of the various conversation
scenarios, and
recognizing intention of the user's conversation on the basis of a similarity
calculation result, to
7
Date Recue/Date Received 2022-04-19

return a pattern to which the intention corresponds.
According to the third aspect, there is provided a computer equipment that
comprises a memory,
a processor and a computer program stored on the memory and operable on the
processor, and
the following operational steps are realized when the processor executes the
computer program:
obtaining sample corpora of various conversation scenarios in a plurality of
conversation
scenarios;
generating, with respect to each said conversation scenario, a scenario
feature of the conversation
scenario on the basis of the sample corpus of the conversation scenario,
wherein the scenario
feature includes feature words of the conversation scenario and a feature word
sequence obtained
by mapping transformation of the feature words; and
configuring the intelligent voice robot on the basis of a preset word vector
space model and
scenario features of the various conversation scenarios, wherein the word
vector space model is
employed for the intelligent voice robot to perform word vector similarity
calculation on a user's
conversation and the scenario features of the various conversation scenarios,
so as to recognize
an intended scenario of the user's conversation.
According to the fourth aspect, there is provided a computer-readable storage
medium storing a
computer program thereon, and the following operational steps are realized
when the computer
program is executed by a processor:
obtaining sample corpora of various conversation scenarios in a plurality of
conversation
scenarios;
generating, with respect to each said conversation scenario, a scenario
feature of the conversation
scenario on the basis of the sample corpus of the conversation scenario,
wherein the scenario
feature includes feature words of the conversation scenario and a feature word
sequence obtained
by mapping transformation of the feature words; and
configuring the intelligent voice robot on the basis of a preset word vector
space model and
scenario features of the various conversation scenarios, wherein the word
vector space model is
employed for the intelligent voice robot to perform word vector similarity
calculation on a user's
8
Date Recue/Date Received 2022-04-19

conversation and the scenario features of the various conversation scenarios,
so as to recognize
an intended scenario of the user's conversation.
The present application provides a method of realizing a configurable
intelligent voice robot, and
corresponding device and storage medium, by obtaining sample corpora of
various conversation
scenarios in a plurality of conversation scenarios, generating, with respect
to each said
conversation scenario, a scenario feature of the conversation scenario on the
basis of the sample
corpus of the conversation scenario, wherein the scenario feature includes
feature words of the
conversation scenario and a feature word sequence obtained by mapping
transformation of the
feature words, and configuring the intelligent voice robot on the basis of a
preset word vector
space model and scenario features of the various conversation scenarios,
wherein the word vector
space model is employed for the intelligent voice robot to perform word vector
similarity
calculation on a user's conversation and the scenario features of the various
conversation
scenarios, an intended scenario of the user's conversation is recognized. By
employing the
technical solution provided by the present application, it is made possible to
process with a set
of general intelligent voice robot solving scheme with respect to any business
scenario of the
same business field, to avoid the embarrassing situation in which it is
required to develop
different voice robots when specific IVR (Interactive Voice Response) scenario
conversations are
carried out for different products or different customer groups of the same
business line, and to
achieve cross-task reuse between common conversation scenarios, whereby the
development
cost of marketing robots and the threshold for customized services are greatly
lowered.
BRIEF DESCRIPTION OF THE DRAWINGS
In order to describe the technical solutions in the embodiments of the present
application more
clearly, accompanying drawings required for use in the description of the
embodiments will be
briefly introduced below. Apparently, the accompanying drawings introduced
below are merely
directed to partial embodiments of the present application, and it is possible
for persons ordinarily
skilled in the art to acquire other drawings based on these drawings without
spending any creative
9
Date Recue/Date Received 2022-04-19

effort in the process.
Fig. 1 is a view illustrating the whole framework for the realization of a
configurable intelligent
voice robot provided by the present application;
Fig. 2 is a flowchart schematically illustrating the method of realizing a
configurable intelligent
voice robot in an embodiment;
Fig. 3 is a flowchart schematically illustrating step 202 in the method shown
in Fig. 2;
Fig. 4 is a flowchart schematically illustrating scenario feature maintenance
in an embodiment;
Fig. 5 is a flowchart schematically illustrating logical entity configuration
in an embodiment;
Fig. 6 is a view illustrating the drawing board effect of frontstage state
transition in an
embodiment;
Fig. 7 is a view schematically illustrating a json file form expressing state
transition torque logical
relation in an embodiment;
Fig. 8 is a flowchart schematically illustrating intended scenario recognition
in an embodiment;
Fig. 9 is a view schematically illustrating the structure of the device for
realizing a configurable
intelligent voice robot in an embodiment; and
Fig. 10 is a view illustrating the internal structure of a computer equipment
provided by an
embodiment of the present application.
Date Recue/Date Received 2022-04-19

DETAILED DESCRIPTION OF THE INVENTION
In order to make more lucid and clear the objectives, technical solutions, and
advantages of the
present application, the technical solutions in the embodiments of the present
application will be
clearly and comprehensively described below with reference to accompanying
drawings in the
embodiments of the present application. Apparently, the embodiments as
described are merely
partial, rather than the entire, embodiments of the present application. All
other embodiments
obtainable by persons ordinarily skilled in the art based on the embodiments
in the present
application shall all be covered by the protection scope of the present
application.
As should be noted, unless definitely required otherwise in the context, such
terms as
"comprising", "including", "containing" and their various forms as used
throughout the
Description and Claims should be understood to mean inclusion rather than
exclusion or
exhaustion, in other words, these terms mean "including, but not limited to.
In addition, as should be noted, such terms as "first" and "second" as used in
the description of
the present application are merely meant for descriptive purposes, and shall
not be understood to
indicate or imply relative importance. In addition, unless otherwise
specified, such terms as
"plural" and "plurality of' as used in the description of the present
application mean "two or
more".
As noted in the Description of Related Art, it is required in prior-art
technology to make
respectively customized developments of intelligent voice robots with respect
to different
scenario classes, migration and reuse therebetween are impossible, so that the
development cost
of scenarios of the intelligent voice robots is rendered unduly high. In view
of this, the present
application creatively proposes a method of realizing a configurable
intelligent voice robot,
whereby unified interface services can be provided and the scenarios are
configurable, the
business is supplied with a visual operation interface through a matched
"intelligent voice
platform", scenario development efficiency and business participation
experience of intelligent
11
Date Recue/Date Received 2022-04-19

voice robots are greatly enhanced, the problem in which robot logical entity
cannot be reused is
solved to certain extent, and a business-friendly and open scenario
configuration interface is
provided, so that the business can directly participate in the training and
generating processes of
the robot entity.
Fig. 1 is a view illustrating the whole framework for the realization of a
configurable intelligent
voice robot provided by the present application. With reference to Fig. 1, the
framework mainly
involves internal basic configuration, external configuration, intention
recognition and FSM
(Finite-state machine) robot logical entity, of which the internal basic
configuration can be
realized by backstage algorithm developing personal through a relevant
algorithm and includes
internal basic features (namely scenario features), internal basic patterns,
and internal basic logic,
and the external configuration can be correspondingly configured by the
frontstage business
personnel through the frontend of the intelligent voice robot backstage
administration system,
and includes externally configurable features, externally configurable
patterns, and conversation
configurable logic. After the intelligent voice robot has been configured and
generated by means
of the technical solution of the present application, the intelligent voice
robot can recognize the
intended scenario on the input content of the framework, namely on text
content recognized and
transcribed from a client's voice through an ASR (automatic speech
recognition) module, and
return corresponding pattern content according to the finite state machine
logical entity, so as to
output voice content transformed from the pattern through TTS (Text-To-Speech)
technique.
The technical solution of the present application is described in detail below
through a plurality
of embodiments.
In one embodiment, there is provided a method of realizing a configurable
intelligent voice robot,
the method can be applied to any computer equipment, such as a server, and the
server can be
embodied as an independent server or a server cluster consisting of a
plurality of servers. As
shown in Fig. 2, the method can comprise the following steps.
12
Date Recue/Date Received 2022-04-19

201 - obtaining sample corpora of various conversation scenarios in a
plurality of conversation
scenarios.
Here, the plural conversation scenarios are contained in a preset conversation
scenario list, and
the conversation scenario list is used to record one or more conversation
scenario(s) of a specific
business field.
Specifically, the sample corpora of the various conversation scenarios can be
obtained by
classifying and marking specific domain corpora according to classes of the
conversation
scenarios, here, the specific domain corpora indicate corpora of a specific
business field, such as
customer service corpora of consumption loan telemarketing. As should be
understood, the
specific obtaining process is not restricted in the embodiments of the present
application.
The aforementioned conversation scenarios can be obtained by performing
scenario abstraction
on the specific domain corpora, and the scenario abstraction is a process from
data to information
and then to knowledge. Exemplarily, with respect to the consumption loan
telemarketing field, it
is possible to sort out common conversation scenarios of telemarketing
activities by analyzing
telemarketing dialogue logs of consumption loan products under the guidance of
business
personnel, such as "problems related to credit", "problems related to limit",
"problems related to
interests" and "operation consultation", etc., and statements in customer
service conversation
logs are marked according to several scenarios that have been summarized and
classified. For
instance, with respect to the application of a consumption loan telemarketing
activity, conversion
scenarios as shown in the following Table 1 can be abstracted:
Table 1: Consumption Loan Telemarketing Conversation Scenarios
Serial Number Conversation Scenario
1 credit related
2 interests related
3 limit related
4 operation consultation
13
Date Recue/Date Received 2022-04-19

contact again
6 affirm
7 negate
8 terminate request
9 unknown
In practical application, each conversation scenario can be abstracted as a
conversation state, and
the dialogue process between customer service and customer can be abstracted
as a transition
between conversation states. If the conversation state is taken as a node, an
oriented line between
conversation states is precisely a process in which one state is transferred
to another state, then
the entire dialogue process can be abstracted as a graph consisting of nodes
and oriented lines.
202 ¨ generating, with respect to each conversation scenario, a scenario
feature of the
conversation scenario on the basis of the sample corpus of the conversation
scenario, wherein
the scenario feature includes feature words of the conversation scenario and a
feature word
sequence obtained by mapping transformation of the feature words.
Specifically, content transformation based on a WOB (Word of Bag) model can be
performed on
the sample corpus of each conversation scenario, discrete representation of
the sample corpus of
each conversation scenario is obtained, a feature selection algorithm is
thereafter employed to
extract feature words of each conversation scenario on the basis of the
discrete representation of
the sample corpus of each conversation scenario, the feature words of each
conversation scenario
are subsequently mapped and transformed to a dictionary index of a preset
domain dictionary,
and a feature word sequence of each conversation scenario is obtained.
The aforementioned WOB model can divide a corpus text into separate terms, and
it is imaged
that all terms are placed in a bag, such elements as their word orders,
grammars and syntaxes are
ignored, the bag is only regarded as a collection of plural vocabularies, each
term as it appears in
the text is independent, and does not rely on whether other terms appear or
not. The WOB model
can be embodied as a one-hot (also referred to as one-hot coding), TF-IDF or N-
gram model.
14
Date Recue/Date Received 2022-04-19

Exemplarily, with respect to five feature words of a "credit related"
conversation scenario,
namely ["credit", "People's Bank of China", "report", "personal credit", "risk
contron, the index
serial numbers as these are mapped and transformed in the corresponding domain
dictionary are
[12, 223, 166, 17, 621, and a feature word sequence of the "credit related"
conversation scenario
is obtained.
Preferably, after the step of generating, with respect to each conversation
scenario, a scenario
feature of the conversation scenario on the basis of the sample corpus of the
conversation scenario,
the method can further comprise:
storing the various conversation scenarios and the scenario features of the
various conversation
scenarios in a scenario feature relation table.
Specifically, the name, feature words and feature word sequence of each
conversation scenario
are correspondingly stored in the scenario feature relation table. The
scenario feature relation
table is used to store correspondence relations between conversation scenarios
and scenario
features (including feature words and feature word sequences).
Exemplarily, a scenario feature relation table of a consumption loan
telemarketing activity can
be as shown in Table 2.
Table 2: Scenario Feature Relation Table of Consumption Loan Telemarketing
Activity
Serial Conversation Feature Word Feature Word
Number Scenario Sequence
1 credit related ['credit', 'People's
Bank of [12, 2331
Chinal
2 interests ['limit', 'loan amount', 'loaned [2, 12, 13,
91
related amount', ' amount '1
3 limit related ['interests',
'interest rate', [3, 5, 91
' amount'l
4 operation ['operate', 'handle', 'set up', [8, 103,
198,
consultation `configure'l 2101
Date Recue/Date Received 2022-04-19

... ... ... ...
In practical application, the above scenario feature relation table is stored
in a server, maintained
offline by backstage algorithm technical personnel based on periodical text
data mining
operations, and isolated from frontstage business personnel.
203 - configuring the intelligent voice robot on the basis of a preset word
vector space model and
scenario features of the various conversation scenarios, wherein the word
vector space model is
employed for the intelligent voice robot to perform word vector similarity
calculation on a user's
conversation and the scenario features of the various conversation scenarios,
so as to recognize
an intended scenario of the user's conversation.
In this embodiment, the intelligent voice robot is configured on the basis of
the word vector space
model and the correspondence relations between the scenario features of the
various conversation
scenarios, and it is thus possible, in practical application, to recognize the
intended scenario of
the user's conversation based on the word vector space model and the scenario
features of the
various conversation scenarios. Specifically speaking, when the intelligent
voice robot is
conversing with the user, it is possible to obtain the user's conversation
text by recognizing and
transforming the user's conversation voice via the automatic speech
recognition (ASR) technique,
extract feature information out of the user's conversation text, hence perform
word vector
similarity calculation on the user's conversation and scenario features of the
various conversation
scenarios according to the word vector space model, and recognize the intended
scenario of the
user's conversation according to the word vector similarity calculation
result.
In an example, the word vector space model can be trained and obtained by:
employing domain corpora of a domain to which the various conversation
scenarios pertain to
train a pre-trained BERT word vector space, and obtaining the word vector
space model.
Here, the domain corpora indicate corpora of a specific business field to
which the various
16
Date Recue/Date Received 2022-04-19

conversation scenarios pertain, such as customer service corpora of
consumption loan
telemarketing.
In this embodiment, pre-trained embedding, which is based on large-scale
corpora and high
computational cost, and which cannot be achieved via current hardware
resources, is obtained by
introducing large-scale BERT (Bidirectional Encoder Representations from
Transformers) word
vector space (768 dimensions) pretrained in google bert serving. On the basis
thereof, the BERT
word vector space is retrained by bringing in own business customer service
corpora, and
calibration of the BERT word vector is realized, so that it more conforms to
specific business
scenarios.
By employing the technical solution of this embodiment, it is made possible to
process with a set
of general intelligent voice robot solving scheme with respect to any business
scenario of the
same business field, to avoid the embarrassing situation in which it is
required to develop
different voice robots when specific IVR (Interactive Voice Response) scenario
conversations are
carried out for different products or different customer groups of the same
business line, to
achieve cross-task reuse between common conversation scenarios, and to make
the business
scenarios extendible, whereby the development cost of marketing robots and the
threshold for
customized services are greatly lowered.
In one embodiment, as shown in Fig. 3, the above step 202 of generating, with
respect to each
conversation scenario, a scenario feature of the conversation scenario on the
basis of the sample
corpus of the conversation scenario can include the following.
301 - obtaining, with respect to each conversation scenario, discrete
representation of the sample
corpus of the conversation scenario on the basis of a preset domain
dictionary.
Specifically, total segmentation corpora from which stop-words are removed can
be based on to
construct the domain dictionary, the domain dictionary includes total
effective vocabularies
17
Date Recue/Date Received 2022-04-19

appearing in the corpora, and the domain dictionary is based on to perform
content transformation
on all sample corpora of a target conversation scenario on the basis of the
WOB model to obtain
the discrete representation, i.e., to transform the corpora to the following
mode of expression:
[index of terms in a dictionary, word frequencies of terms in a document]
For instance, there is a basic dictionary that contains four terms, ["I",
"you", "dislike", "love",
"Nanjing", "hometown"1, then the statement "I / love / my / hometown /
Nanjing" will be
[(0,2),(3,1),(4,1),(5,1)] after transcription by the word of bag.
302 - employing a feature selection algorithm to extract feature words of the
conversation
scenario on the basis of the discrete representation of the sample corpus of
the conversation
scenario.
Specifically, after word of bag transformation, discrete representation of the
sample corpus of the
target conversation scenario is obtained, and the feature selection algorithm
can be used thereafter
to extract feature words of the target conversation scenario.
Preferably, the feature selection algorithm is a chi-square statistic feature
selection algorithm.
In this embodiment, the chi-square statistic (CHI) technique can be based on
to extract feature
words of the target conversation scenario. During the process of specific
implementation, feature
word extraction of the target conversation scenario can be performed according
to the form
shown in the following Table 3 and CHI calculation formula.
Table 3: Classes and Document Ownership
Document Ownership Number of Documents of Number of Documents of
Class C non-Class C
number of t documents a b
included
number oft documents c d
18
Date Recue/Date Received 2022-04-19

not included
CHI calculation formula:
N x (ad ¨ cb)2
x2 (t, c) ¨
(a + c) x (b + d) x (a + b) x (c + d)
where c is a certain class, namely a "conversation scenario", t is a certain
term, and N is the total
number of texts in the training corpora.
The above x2 is generally used for chi square hypothesis verification in
statistics, to judge
uniformity or goodness of fit between actual distribution and theoretical
distribution, with its null
hypothesis Ho being "no marked difference between observed frequency and
desired
frequency". Accordingly, the less the chi square statistic is, the closer will
be the observed
frequency to the desired frequency, and the higher will be the relevancy
between the two.
Therefore, x2 can be regarded as a measure of the distance between an observed
object and a
desired object, the less the distance is, the higher will be the relevancy
therebetween. The
"observed object" in the present application is a term, and the desired object
is a conversation
scenario, if the term and the conversation scenario are highly relevant,
statistic distributions of
the two should be close to each other in the entire samples. Consequently,
through the x2
statistic, relevancies between all vocabularies in a dictionary and the
various classes can be
calculated relatively quickly and accurately on the basis of great quantities
of corpora, and a
preset number of terms (the preset number can be set as 5) with the least x2
is selected according
to the relevancy sorting result to serve as a feature set of conversation
scenarios, to complete
feature mapping of various scenarios/classes in the scenario list.
303 - mapping to transform the features words of the conversation scenario to
a corresponding
dictionary index, and generating the feature word sequence of the conversation
scenario.
Specifically, feature words of the target conversation scenario are mapped and
transformed
19
Date Recue/Date Received 2022-04-19

through dictionary index of the domain dictionary, to obtain the feature word
sequence of the
target conversation scenario.
Exemplarily, suppose that the five feature words extracted from the "credit
related" conversation
scenario in step 302 are ["credit", "People's Bank of China", "report",
"personal credit", "risk
contron, through dictionary index mapping transformation, index serial numbers
can be
obtained in the corresponding dictionary as [12, 223, 166, 17, 621, namely the
feature word
sequence of the target conversation scenario.
In the state of the art, such word vector models as Word2vec, Glove or ELMo
are usually
employed with respect to sample corpora to generate word vectors, but because
abstract
dimensions of terms are still hidden in the word vectors as generated, it is
impossible for the user
to maintain and to extend, whereas in the embodiments of the present
application, with respect
to each conversation scenario, discrete representation of the sample corpus of
the conversation
scenario is obtained on the basis of the preset domain dictionary, and the
feature words of the
target conversation scenario are extracted by means of a feature selection
algorithm and mapped
and transformed to a feature word sequence, since the feature words and the
feature word
sequence of the target conversation scenario can be expressed as specific
terms, it is therefore
made possible to facilitate the user to perform visual maintenance and
extension, so as to provide
the business personnel and the algorithm personnel with the possibility to
cooperate in the
development of scenario features of intelligent voice robots.
In one embodiment, as shown by the flowchart of Fig. 4 schematically
illustrating scenario
feature maintenance, the method further comprises the following.
401 - receiving a configuration feature word input with respect to any
conversation scenario.
Specifically, a configuration feature word input by the user with respect to
the target conversion
scenario through the system frontend is received. The user here indicates the
business personnel,
Date Recue/Date Received 2022-04-19

such as the customer service personnel; during the process of specific
implementation, the
frontend can provide the business personnel with a "feature relation
extension" function to
maintain the business field, whereby the business personnel selects a specific
conversion scenario
option at the frontend, and keys in a vocabulary highly related to the
scenario as summarized and
extracted by the business personnel himself according to business knowledge
and experience at
the input box, to service as the configuration feature word of the target
conversion scenario; on
receiving the configuration feature word, the system updates it into an
externally input feature
set of the target conversion scenario after review at the backstage.
The step of receiving a configuration feature word input with respect to any
conversation scenario
includes:
receiving a configuration feature word input by a user having feature
configuration permission
with respect to the conversation scenario.
In this embodiment, since configuration feature words coming from the outside
might be
problematic in terms of instable qualities due to difference in experiences
and levels of the
business personnel, it is therefore possible to open the external feature
configuration permission
only to those well-experienced, selected business personnel by constructing a
feature
configuration permission list.
402 - maintaining the scenario feature of the conversation scenario in the
scenario feature relation
table on the basis of the configuration feature word of the conversation
scenario and a
configuration feature word sequence obtained by mapping transformation of the
configuration
feature word.
In one example, the process of realizing step 402 may include:
merging the configuration feature word of the conversation scenario to the
feature words of the
conversation scenario in the scenario feature relation table, and adding the
configuration feature
word sequence of the merged configuration feature word to the feature word
sequence of the
21
Date Recue/Date Received 2022-04-19

conversation scenario.
Specifically, after having performed such basic processes as duplicate-
removing and rectifying
the configuration feature word input with respect to the target conversation
scenario, the
configuration feature word is directly added to the feature words of the
target conversation
scenario in the scenario feature relation table, and feature word sequence
mapping is carried out.
The final merging result is as shown in the following Table 4.
Table 4: Scenario Feature Relation Table after Internal and External
Configurations Merging
Serial Conversation Feature Feature Externally Merged Merged
Feature
Number Scenario Word Sequence Input Feature Word Sequence
Vocabulary Word
1 credit related ['credit', [12,223] ['People's
['credit', [12,223]
'People's Bank of 'People's
Bank'] China'] Bank',
'People's
Bank of
China']
2 interests ['interests', [3,5,9] ['rate', ['interests',
[3,5,9,102]
related 'interest 'interest 'interest
rate', rate'] rate',
'amount of 'amount of
money 'I money',
'rate]
3 limit related ['limit', [2,12,13,9] ['loan ['limit',
[2,12,13,9,4,86,321]
'loan limit', 'loan
amount', 'limit', amount',
'loaned 'amounel 'loaned
amount', amount',
'amount of 'amount of
money 'I money',
'loan limit',
'amount'l
4 operation ['operate', [8,103,198,210] [] ['operate',
[8,103,198,210]
consultation 'handle', 'handle',
22
Date Recue/Date Received 2022-04-19

'set up', 'set up',
`configure'l `configure'l
... ... - - - ... ... - - - ...
As should be noted, if an externally input configuration feature word is not
contained in the
domain dictionary, this configuration feature word is ignored. For instance,
the term "People's
Bank of China" in Table 4 is not in the domain dictionary, this term will be
directly ignored when
feature word sequences are merged. In this embodiment, by maintaining the
scenario feature of
the conversation scenario in the scenario feature relation table on the basis
of the configuration
feature word of the conversation scenario and a configuration feature word
sequence obtained by
mapping transformation of the configuration feature word, it is made possible
to overcome the
traditional operation mode of "frontstage and backstage" in which "businesses
define
requirements, and technologies develop applications", and to greatly alleviate
the problem of
"detachment of development from production, and mismatch of requirement and
response"
pending in the traditional robot development pattern by realizing cooperative
development of
marketing robots by business personnel and technical personnel through the
pattern of "common
definition of internal and external configurations".
In one embodiment, such as shown in the flowchart schematically illustrating
logical entity
configuration in Fig. 5, the method further comprises the following.
501 ¨ receiving, with respect to any conversation scenario, a state transition
graph input by a first
user for the conversation scenario, and receiving supplementary information
input by a second
user for the state transition graph, to generate a state transition matrix of
the conversation scenario.
The first user indicates business personnel, the second user indicates
algorithm developing
technical personnel.
502 ¨ generating a script file for containing state transition logical
relation on the basis of the
23
Date Recue/Date Received 2022-04-19

state transition matrix of the conversation scenario, and generating a finite
state machine on the
basis of the script file, to return a corresponding pattern when the intended
scenario of the user's
conversation is recognized.
The finite state machine (FSM), also referred to as a finite state automat and
abbreviated as a
state machine, is a mathematical model that expresses a finite number of
states and such
behaviors as transitions and actions amongst these states. The main function
thereof is to describe
state sequences of an object experienced within its life cycle, and how to
respond to various
events coming from the outside to transition between the states. A state
machine includes four
elements, namely current state, condition, action, and next state. The current
state is a state
currently being in; the condition is also referred to as "event", when a
condition is satisfied, an
action will be triggered or one round of state transition will be performed;
the action indicates an
action executed after the condition has been satisfied, after the action has
been executed to
completion, a new state can be transitioned, and it is also possible to
maintain the original state,
the action is not indispensable, after the condition has been satisfied, it is
also possible not to
execute any action but to directly transition to a new state; the next state
indicates a new state to
be transitioned after the condition has been satisfied. The "next state" is
relative to the "current
state", once a "next state" is activated, it becomes a new "current state".
The logical relation of
the finite state machine can be expressed as a state transition matrix as
shown in the following
Table 5.
Table 5: FSM State Transition Matrix
Next
State
State A State B State C ...
Curren
State
Triggering Triggering Triggering
State A Condition: Event Condition: Event Condition:
Event ...
1.1 1.2 1.3
24
Date Recue/Date Received 2022-04-19

Triggering Triggering
State B Condition: Event Condition: Event Nonexistent ...
2.1 2.2
Triggering Triggering
State C Condition: Event Nonexistent Condition:
Event ...
3.1 3.3
... ... ... ... ...
The "action" here indicates a changing process in which the "current state" is
transformed to the
"next state" after the "triggering condition" has been satisfied.
In practical application, after the business personnel has drawn a basic state
transition graph
through drawing board dragging at the frontend platform, a possibly incomplete
state transition
torque will be generated at the backstage. The drawing board effect of
frontstage state transition
can be as shown with reference to Fig. 6. Considering restriction of the
business technical
capability, it is very difficult to guarantee the basic requirement on
"logical completeness" of the
state transition torque, so it is usual for backstage technical personnel to
modify the newly
generated state transition torques daily at fixed timing, so as to make them
conform to the states
shown in Fig. 5. A finite state machine model can be abstracted as the
following combination of
patterns:
START event END;
When the intended scenario of the user's conversation is recognized, it is
possible to return the
corresponding pattern content through logical entity of the finite state
machine. The
"conversation scenario" is a "condition/event". For instance, the current
state is a "greeting" state,
the condition/event is a "limit related" conversation scenario, and the next
state of "consulting
on limit" is transitioned after the action has been triggered and executed. In
the next round of
dialogue, "consulting on limit" is the current state, the condition/event is
"negate", and the next
state of "transaction not concluded" is transitioned after the action has been
triggered and
executed. A complete customer service dialogue process is thus completed. The
state transition
Date Recue/Date Received 2022-04-19

logic is as follows:
( greeting) limit related' ( consulting on limit) negate' (transaction not
concluded) ;
The FSM state transition torque completely repaired above will be
automatically translated to a
script file of json format by the program.
In practical application, with respect to a consumption loan activity, the
json script file form
expressing its state transition torque logical relation can be as shown with
reference to Fig. 7.
The json script file will be subsequently read and input by the program to a
state machine object
when a finite state machine instance is generated, to validate the logic. The
state machine instance
as generated will be stored in Redis as an index value according to uuid
transferred by the
frontend, so as to facilitate subsequent program access when the IVR service
is started. It is also
possible for the user to perform persistent operation on the state machine as
required, so as to
facilitate long-term stable invoking of the state machine. If the user selects
a task type as "single
task" at the frontend, the finite state machine instance of this task as
stored in Redis will be
cleared away as an invalidated object within a preset time period (such as 24
hours) after the IVR
marketing service has been triggered, so as to economize on storage resources.
Relative to the traditional development mode in which the marketing robot
instances are trained
according to business scenario requirements offline, since the traditional
development mode
relies on solidified configuration files and state transition logics, once
environmental change
occurs at the business side, such as requiring to modify the logic or the
pattern, the reconstruction
cost and risk will be huge at the R&D side, use of the technical solution of
the present
embodiment ensures timely response by the robot to the update of the
configuration information
through the mode of frontend configuration and backstage translation,
moreover, configuration
of file is decoupled from program application, so that update of the core
assembly of the robot,
namely the FSM, is made more simple and handy. Furthermore, the pattern
section also relies on
the frontstage for maintenance, thereby realizing timely validation.
26
Date Recue/Date Received 2022-04-19

In addition, in practical application, the developing personnel completes the
abstraction of
business logic and construction of the state transition graph according to
product requirements
on their own initiative, but there lacks direct participation and supervision
by the business in the
entire process of robot development, while the final application effect of the
robot depends
greatly on the understanding of the business background and the commercial
environment, so
this causes detachment of the development from the production and mismatch
between
requirement and response to a greater extent. However, the technical solution
of the present
embodiment makes it possible to overcome the traditional operation mode of
"frontstage and
backstage" in which "businesses define requirements, and technologies develop
applications",
and to greatly alleviate the problem of "detachment of development from
production, and
mismatch of requirement and response" pending in the traditional robot
development pattern by
realizing cooperative development of intelligent voice robots by business
personnel and technical
personnel through the pattern of "common definition of internal and external
configurations".
In one embodiment, as shown in the flowchart schematically illustrating
intended scenario
recognition in Fig. 8, the method further comprises the following.
801 - preprocessing the user's conversation to obtain a plurality of segmented
terms in the user's
conversation after the well-configured intelligent voice robot has received
the user's conversation,
performing mapping transformation on the plural segmented terms, and obtaining
a feature word
sequence of the user's conversation.
Specifically, after the intelligent voice robot has been configured and
generated, the intelligent
voice robot can converse with the user, the user's conversation can be text
content recognized
and transcribed from user's conversation speech through automatic speech
recognition (ASR)
technology, the text content is word-segmented to obtain a plurality of
segmented terms, and the
word-segmentation process includes character purification, rectification, word
segmentation, and
removal of stop-words. The plural segmented terms are mapped and transformed
through index
of the domain dictionary to the form of expression like the "feature word
sequence" of the
27
Date Recue/Date Received 2022-04-19

conversation scenario in the scenario feature relation table, namely to obtain
a feature word
sequence of the user's conversation.
802 - employing the word vector space model to construct a feature vector of
the user's
conversation and scenario feature vectors of the various conversation
scenarios on the basis of
the feature word sequence of the user's conversation and feature word
sequences of the various
conversation scenarios.
In an example, the process of realizing step 802 can include:
employing the word vector space model to respectively map the feature word
sequence of the
user's conversation and the feature word sequences of the various conversation
scenarios in the
scenario feature relation table, and generating a feature vector of the user's
conversation and
scenario feature vectors of the various conversation scenarios.
Specifically speaking, each element in the feature word sequence of the user's
conversation is
mapped to the BERT word vector space to obtain a feature vector of 768
dimensions, and all
elements are summated and averaged (or chosen with the maximal or medium value
thereof) to
obtain a 1x768 vector to serve as feature expression input by the user's
conversation.
Correspondingly, the feature word sequences of the various conversation
scenarios in the
scenario feature relation table are likewise operated to be respectively
converted to 1x768 feature
vectors.
803 ¨ performing similarity calculation on the feature vector of the user's
conversation and the
scenario feature vectors of the various conversation scenarios, and
recognizing intention of the
user's conversation on the basis of a similarity calculation result, to return
a pattern to which the
intention corresponds.
Specifically, with respect to each conversation scenario, a cosine similarity
between the feature
vector input by the user's conversation and the scenario feature vector of the
conversation
28
Date Recue/Date Received 2022-04-19

scenario is calculated, the greater the value is, the higher will be the
similarity, and the higher
will also be the relevancy between the user's conversation and the
conversation scenario. By
arranging all the conversation scenarios in a descending order according to
the cosine similarity
calculation result, a conversation scenario with the highest similarity is
returned to serve as the
judgement result of the intended scenario of the current input by the user,
and a corresponding
response pattern under the intended scenario is returned according to the
state in which the finite
state machine currently is, wherein the response pattern will be transformed
to speech content
through Text-To-Speech (TTS) technology to be broadcast to the user. Forms of
the response
pattern can be as shown with reference to Table 6.
Table 6: State Pattern Relation Table
Serial State Pattern
Number
1 Greeting Hello, this is Xiao Su of the Customer Service
Center.
2 Consulting This product is loanable for a revolving line of
credit at the
on limit maximum of 300 thousand RMBs, borrow and return as you, the
loan can be completed within 30 minutes at the quickest, direct cash
withdrawal is enabled to Your bank card for use after opening.
3 Consulting Hello, the expense standards are unified at our end,
You can repay
on interests the loan by 5 instalments, 10 instalments or 15 instalments, and
You
may also borrow and return as you go, as ours are more flexible as
compared with banks on the market and other finance companies.
4 Consulting This loan will be brought in the personal credit
investigation system
on credit of the Central Bank, maintaining an excellent credit helps You
enjoy subsequent, other services of the State and the banks.
Transaction You may enquire relevant activities on our applets, I thank You
for
not Your listening. Have a good day! Goodbye.
concluded
... - - - ...
Use of the technical solution of the present embodiment can enhance the
precision of the result
of recognizing the intention of the user's conversation by employing the word
vector space model
to construct a feature vector of the user's conversation and scenario feature
vectors of the various
29
Date Recue/Date Received 2022-04-19

conversation scenarios on the basis of the feature word sequence of the user's
conversation and
feature word sequences of the various conversation scenarios, and by
performing similarity
matching on the feature vector of the user's conversation and the scenario
feature vectors of the
various conversation scenarios, so as to recognize the intended scenario of
the user's conversation.
As should be understood, although the various steps in the aforementioned
flowcharts are
sequentially displayed as indicated by arrows, these steps are not necessarily
executed in the
sequences indicated by arrows. Unless otherwise explicitly noted in this
paper, execution of these
steps is not restricted by any sequence, as these steps can also be executed
in other sequences
(than those indicated in the drawings). Moreover, at least partial steps in
the flowcharts may
include plural sub-steps or multi-phases, these sub-steps or phases are not
necessarily completed
at the same timing, but can be executed at different timings, and these sub-
steps or phases are
also not necessarily sequentially performed, but can be performed in turns or
alternately with
other steps or with at least some of sub-steps or phases of other steps.
In one embodiment, there is provided a device for realizing a configurable
intelligent voice robot,
and the device is employed for executing the method of realizing a
configurable intelligent voice
robot as recited in the above embodiment. As shown in Fig. 9, the realizing
device can comprise:
an obtaining module 901, for obtaining sample corpora of various conversation
scenarios in a
plurality of conversation scenarios;
a generating module 902, for generating, with respect to each conversation
scenario, a scenario
feature of the conversation scenario on the basis of the sample corpus of the
conversation scenario,
wherein the scenario feature includes feature words of the conversation
scenario and a feature
word sequence obtained by mapping transformation of the feature words; and
a configuring module 903, for configuring the intelligent voice robot on the
basis of a preset word
vector space model and scenario features of the various conversation
scenarios, wherein the word
vector space model is employed for the intelligent voice robot to perform word
vector similarity
calculation on a user's conversation and the scenario features of the various
conversation
scenarios, so as to recognize an intended scenario of the user's conversation.
Date Recue/Date Received 2022-04-19

In one embodiment, the generating module 902 can include:
a representing unit, for obtaining, with respect to each conversation
scenario, discrete
representation of the sample corpus of the conversation scenario on the basis
of a preset domain
dictionary;
a screening unit, for employing a feature selection algorithm to extract
feature words of the
conversation scenario on the basis of the discrete representation of the
sample corpus of the
conversation scenario; and
a generating unit, for mapping to transform the features words of the
conversation scenario to a
corresponding dictionary index, and generating the feature word sequence of
the conversation
scenario;
preferably, the feature selection algorithm is a chi-square statistic feature
selection algorithm.
In one embodiment, the device can further comprise:
a storing module 904, for storing the various conversation scenarios and the
scenario features of
the conversation scenarios in a scenario feature relation table;
a receiving module 905, for receiving a configuration feature word input with
respect to any
conversation scenario; and
a maintaining module 906, for maintaining the scenario feature of the
conversation scenario in
the scenario feature relation table on the basis of the configuration feature
word of the
conversation scenario and a configuration feature word sequence obtained by
mapping
transformation of the configuration feature word.
Preferably, the receiving module 905 is employed for receiving a configuration
feature word
input by a user having feature configuration permission with respect to the
conversation scenario.
In one embodiment, the maintaining module 906 is employed for merging the
configuration
feature word of the conversation scenario to the feature words of the
conversation scenario in the
scenario feature relation table, and adding the configuration feature word
sequence of the merged
31
Date Recue/Date Received 2022-04-19

configuration feature word to the feature word sequence of the conversation
scenario.
In one embodiment, the device can further comprise a training module for
employing domain
corpora of a domain to which the various conversation scenarios pertain to
train a pre-trained
BERT word vector space, and obtaining the word vector space model.
In one embodiment, the device further comprises a state machine configuring
module 907 for:
receiving, with respect to any conversation scenario, a state transition graph
input by a first user
for the conversation scenario, and receiving supplementary information input
by a second user
for the state transition graph, to generate a state transition matrix of the
conversation scenario;
and
generating a script file for containing state transition logical relation on
the basis of the state
transition matrix of the conversation scenario, and generating a finite state
machine on the basis
of the script file, to return a corresponding pattern when the intended
scenario of the user's
conversation is recognized.
In one embodiment, the device further comprises an intended scenario
recognizing module 908
that includes:
an obtaining unit, for preprocessing the user's conversation to obtain a
plurality of segmented
terms in the user's conversation after the well-configured intelligent voice
robot has received the
user's conversation, performing mapping transformation on the plural segmented
terms, and
obtaining a feature word sequence of the user's conversation;
a constructing unit, for employing the word vector space model to construct a
feature vector of
the user's conversation and scenario feature vectors of the various
conversation scenarios on the
basis of the feature word sequence of the user's conversation and feature word
sequences of the
various conversation scenarios; and
a matching unit, for performing similarity calculation on the feature vector
of the user's
conversation and the scenario feature vectors of the various conversation
scenarios, and
recognizing intention of the user's conversation on the basis of a similarity
calculation result, to
32
Date Recue/Date Received 2022-04-19

return a pattern to which the intention corresponds.
Specific definitions relevant to the device for realizing an intelligent voice
robot may be inferred
from the aforementioned definitions to the method of realizing an intelligent
voice robot, while
no repetition is made in this context. The various modules in the
aforementioned device for
realizing an intelligent voice robot can be wholly or partly realized via
software, hardware, and
a combination of software with hardware. The various modules can be embedded
in the form of
hardware in a processor in a computer equipment or independent of any computer
equipment,
and can also be stored in the form of software in a memory in a computer
equipment, so as to
facilitate the processor to invoke and perform operations corresponding to the
aforementioned
various modules.
In one embodiment, a computer equipment is provided, the computer equipment
can be a server,
and its internal structure can be as shown in Fig. 10. The computer equipment
comprises a
processor, a memory, and a network interface connected to each other via a
system bus. The
processor of the computer equipment is employed to provide computing and
controlling
capabilities. The memory of the computer equipment includes a nonvolatile
storage medium and
an internal memory. The nonvolatile storage medium stores therein an operating
system, and a
computer program. The internal memory provides environment for the running of
the operating
system and the computer program in the nonvolatile storage medium. The network
interface of
the computer equipment is employed to connect to other equipments via network
for
communication. The computer program realizes a method of realizing an
intelligent voice robot
when it is executed by a processor.
As understandable to persons skilled in the art, the structure illustrated in
Fig. 10 is merely a
block diagram of partial structure relevant to the solution of the present
application, and does not
constitute any restriction to the computer equipment on which the solution of
the present
application is applied, as the specific computer equipment may comprise
component parts that
are more than or less than those illustrated in Fig. 10, or may combine
certain component parts,
33
Date Recue/Date Received 2022-04-19

or may have different layout of component parts.
In one embodiment, there is provided a computer equipment that comprises a
memory, a
processor and a computer program stored on the memory and operable on the
processor, and the
following steps are realized when the processor executes the computer program:
generating, with respect to each conversation scenario, a scenario feature of
the conversation
scenario on the basis of the sample corpus of the conversation scenario,
wherein the scenario
feature includes feature words of the conversation scenario and a feature word
sequence obtained
by mapping transformation of the feature words; and
configuring the intelligent voice robot on the basis of a preset word vector
space model and
scenario features of the various conversation scenarios, wherein the word
vector space model is
employed for the intelligent voice robot to perform word vector similarity
calculation on a user's
conversation and the scenario features of the various conversation scenarios,
so as to recognize
an intended scenario of the user's conversation.
In one embodiment, there is provided a computer-readable storage medium
storing thereon a
computer program, and the following steps are realized when the computer
program is executed
by a processor:
generating, with respect to each conversation scenario, a scenario feature of
the conversation
scenario on the basis of the sample corpus of the conversation scenario,
wherein the scenario
feature includes feature words of the conversation scenario and a feature word
sequence obtained
by mapping transformation of the feature words; and
configuring the intelligent voice robot on the basis of a preset word vector
space model and
scenario features of the various conversation scenarios, wherein the word
vector space model is
employed for the intelligent voice robot to perform word vector similarity
calculation on a user's
conversation and the scenario features of the various conversation scenarios,
so as to recognize
an intended scenario of the user's conversation.
In summary, as compared with prior-art technology, the technical solutions
according to the
34
Date Recue/Date Received 2022-04-19

embodiments of the present application can achieve the following technical
effects.
1. Use of the configurable intelligent voice robot according to the present
application can save
R&D cost for approximately 60% by comprehensive consideration. In particular
as regards
individual business fields in which marketing scenarios are relatively
concentrated, the cost saved
is positively correlated with the number of scenarios in the present
application as compared with
the traditional pattern.
2. The frontend configuration validation pattern as employed by the present
application enables
the robot to be updated securely, stably, and handily according to change in
the business
environment, and validation is made possible at the next day, whereby great
convenience is
brought to the business side, and running stability is enhanced at the same
time at the technical
side.
3. In conjunction with the frontend platform, the business personnel can
directly participate in
the R&D process of core assemblies of the marketing robot, whereby
satisfaction and the sense
of participation are greatly enhanced for the business personnel.
4. In cooperation with the frontend voice servicing platform, the robot
provides the business
side with a whole set of closed-loop solving scheme from "requirement
submission" to
"intelligent voice robot generation" to "IVR telemarketing initiation" and
then to "feedback of
adjustment requirement", thus enabling one-step response to customer
intelligent marketing IVR
requirements.
As comprehensible to persons ordinarily skilled in the art, the entire or
partial flows in the
methods according to the aforementioned embodiments can be completed via a
computer
program instructing relevant hardware, the computer program can be stored in a
nonvolatile
computer-readable storage medium, and the computer program can include the
flows as
embodied in the aforementioned various methods when executed. Any reference to
the memory,
Date Recue/Date Received 2022-04-19

storage, database or other media used in the various embodiments provided by
the present
application can all include nonvolatile and/or volatile memory/memories. The
nonvolatile
memory can include a read-only memory (ROM), a programmable ROM (PROM), an
electrically programmable ROM (EPROM), an electrically erasable and
programmable ROM
(EEPROM) or a flash memory. The volatile memory can include a random access
memory (RAM)
or an external cache memory. To serve as explanation rather than restriction,
the RAM is
obtainable in many forms, such as static RAM (SRAM), dynamic RAM (DRAM),
synchronous
DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM),
synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM
(RDRAM), direct Rambus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM),
etc.
Technical features of the aforementioned embodiments are randomly combinable,
while all
possible combinations of the technical features in the aforementioned
embodiments are not
exhausted for the sake of brevity, but all these should be considered to fall
within the scope
recorded in the Description as long as such combinations of the technical
features are not
mutually contradictory.
The foregoing embodiments are merely directed to several modes of execution of
the present
application, and their descriptions are relatively specific and detailed, but
they should not be
hence misunderstood as restrictions to the inventive patent scope. As should
be pointed out,
persons with ordinary skill in the art may further make various modifications
and improvements
without departing from the conception of the present application, and all
these should pertain to
the protection scope of the present application. Accordingly, the patent
protection scope of the
present application shall be based on the attached Claims.
36
Date Recue/Date Received 2022-04-19

Representative Drawing