Note: Descriptions are shown in the official language in which they were submitted.
CA 03166556 2022-06-29
METHOD AND DEVICE FOR GENERATING TARGET ADVERTORIAL BASED ON
DEEP LEARNING
BACKGROUND OF THE INVENTION
Technical Field
[0001] The present invention relates to the field of natural language
processing technology, and
more particularly to a method of and a device for generating a target
advertorial based on
deep learning.
Description of Related Art
[0002] Marketing advertorials are frequently used in the promotion of new
products to the market,
and a marketing advertorial usually consists of three sections, namely a
title, an
introduction, and a marketing text. Vivid and concise language is employed in
the title to
describe the marketed product as attractive to the people, the introduction
exerts an
introducing function to guide the direction of consumption and lead to the
following
marketing text, while the marketing text describes the product and recommends
purchase
thereof.
[0003] At present, marketing advertorials are mostly written manually by
operating personnel of
the merchants or automatically generated by templates regardless of titles,
introductions,
and marketing texts. However, both of these methods are more or less
defective.
[0004] As regards manual writing, it is required for related personnel to
organize vivid language
to manually write marketing advertorials according to the categories to be
marketed, once
it is needed to output great quantities of advertorials or to expand to
relatively many
categories in a short time, it is usually problematic in terms of low
production efficiency.
[0005] As regards template generation, although batch generation is possible
in a short time, the
statements as generated are problematic in terms of fixed patterns,
stereotyped styles, and
insufficient diversities, etc.
SUMMARY OF THE INVENTION
[0006] In order to deal with problems pending in the state of the art,
embodiments of the present
1
Date Regue/Date Received 2022-06-29
CA 03166556 2022-06-29
invention provide a method of and a device for generating a target advertorial
based on
deep learning, so as to solve such prior-art problems as low production
efficiency in
manual writing of target advertorials, and fixed patterns, stereotyped styles,
and
insufficient diversities of statements in template generation of target
advertorials.
[0007] In order to solve one or more technical problems mentioned above, the
present invention
employs the following technical solutions.
[0008] According to one aspect, there is provided a method of generating a
target advertorial
based on deep learning, the method comprises the following steps:
[0009] receiving relevant information of a target object, and matching plural
adaptable target
titles out of a title library according to the relevant information, wherein
titles in the title
library are derived by expansion of collected titles by a third generating
model;
[0010] inputting the target titles in a first generating model, and generating
at least one target
introduction;
[0011] generating at least one piece of input information that conforms to a
preset structure
according to the relevant information and a preset rule, inputting the input
information in
a second generating model, and generating at least one target text; and
[0012] assembling the target titles, the target introduction, and the target
text, and obtaining
plural target advertorials.
[0013] Further, the step of generating at least one piece of input information
that conforms to a
preset structure according to the relevant information and a preset rule,
inputting the input
information in a second generating model, and generating at least one target
text includes:
[0014] subjecting the relevant information to a term-segmenting process, and
extracting target
segmented terms that satisfy a preset condition from a first term-segmenting
result as
obtained;
[0015] recombining the target segmented terms, and obtaining at least one
piece of input
information that conforms to a preset structure; and
2
Date Regue/Date Received 2022-06-29
CA 03166556 2022-06-29
[0016] inputting the input information in the second generating model, and
generating at least
one target text.
[0017] Moreover, the method further comprises a process of constructing the
title library,
including:
[0018] subjecting plural collected first sample titles to a term-segmenting
process, and obtaining
a second term-segmenting result;
[0019] employing a preset first keyword extracting method to extract a first
keyword from the
first sample titles; and
[0020] inputting the second term-segmenting result and the first keyword in
the third generating
model, and obtaining plural new titles, wherein the title library consists of
the new titles.
[0021] Moreover, the process of constructing the title library further
includes:
[0022] intersecting a first keyword set with the second term-segmenting
result, and obtaining an
input data set; and
[0023] taking data of the input data set as input, taking the target titles as
output, and training out
the third generating model based on a preset algorithm.
[0024] Moreover, the method further comprises a process of constructing the
first generating
model, including:
[0025] subjecting plural collected second sample titles and introductions
corresponding to the
second sample titles to a term-segmenting process;
[0026] employing a preset second keyword extracting method to extract a second
keyword from
the second sample titles;
[0027] intersecting a second keyword set with each term-segmented second
sample title, and
obtaining a target keyword;
[0028] traversing each said second sample title, matching the target keyword
with the totally
term-segmented introductions corresponding to the second sample titles, and
obtaining a
3
Date Regue/Date Received 2022-06-29
CA 03166556 2022-06-29
successfully matched introduction to serve as a new introduction of the
current second
sample title; and
[0029] taking the second sample title as input, taking the introduction
corresponding to the
second sample title and the new introduction as output, and training out the
first
generating model based on a preset algorithm.
[0030] According to another aspect, there is provided a device for generating
a target advertorial
based on deep learning, the device comprises:
[0031] a title matching module, for receiving relevant information of a target
object, and
matching plural adaptable target titles out of a title library according to
the relevant
information, wherein titles in the title library are derived by expansion of
collected titles
by a third generating model;
[0032] an introduction generating module, for inputting the target titles in a
first generating
model, and generating at least one target introduction;
[0033] a text generating module, for generating at least one piece of input
information that
conforms to a preset structure according to the relevant information and a
preset rule,
inputting the input information in a second generating model, and generating
at least one
target text; and
[0034] an information assembling module, for assembling the target titles, the
target introduction,
and the target text, and obtaining plural target advertorials.
[0035] Further, the text generating module includes:
[0036] a first term-segmenting unit, for subjecting the relevant information
to a term-segmenting
process, and extracting target segmented terms that satisfy a preset condition
from a first
term-segmenting result as obtained;
[0037] a segmented terms recombining unit, for recombining the target
segmented terms, and
obtaining at least one piece of input information that conforms to a preset
structure; and
[0038] a text generating unit, for inputting the input information in the
second generating model,
4
Date Regue/Date Received 2022-06-29
CA 03166556 2022-06-29
and generating at least one target text.
[0039] Moreover, the device further comprises a first constructing module that
includes:
[0040] a second term-segmenting unit, for subjecting plural collected first
sample titles to a term-
segmenting process, and obtaining a second term-segmenting result;
[0041] a first extracting unit, for employing a preset first keyword
extracting method to extract
a first keyword from the first sample titles; and
[0042] a title generating unit, for inputting the second term-segmenting
result and the first
keyword in the third generating model, and obtaining plural new titles,
wherein the title
library consists of the new titles.
[0043] Moreover, the first constructing module further includes:
[0044] a first intersecting unit, for intersecting a first keyword set with
the second term-
segmenting result, and obtaining an input data set; and
[0045] a first training unit, for taking data of the input data set as input,
taking the target titles as
output, and training out the third generating model based on a preset
algorithm.
[0046] Moreover, the device further comprises a second constructing module
that includes:
[0047] a third term-segmenting unit, for subjecting plural collected second
sample titles and
introductions corresponding to the second sample titles to a term-segmenting
process;
[0048] a second extracting unit, for employing a preset second keyword
extracting method to
extract a second keyword from the second sample titles;
[0049] a second intersecting unit, for intersecting a second keyword set with
each term-
segmented second sample title, and obtaining a target keyword;
[0050] an introduction expanding unit, for traversing each said second sample
title, matching the
target keyword with the totally term-segmented introductions corresponding to
the
second sample titles, and obtaining a successfully matched introduction to
serve as a new
introduction of the current second sample title; and
Date Regue/Date Received 2022-06-29
CA 03166556 2022-06-29
[0051] a second training unit, for taking the second sample title as input,
taking the introduction
corresponding to the second sample title and the new introduction as output,
and training
out the first generating model based on a preset algorithm.
[0052] The technical solutions provided by the embodiments of the present
invention bring about
the following advantageous effects.
[0053] 1. In the method of and device for generating a target advertorial
based on deep learning
as provided by the embodiments of the present invention, by receiving relevant
information of a target object, matching plural adaptable target titles out of
a title library
according to the relevant information, wherein titles in the title library are
derived by
expansion of collected titles by a third generating model, inputting the
target titles in a
first generating model, generating at least one target introduction,
generating at least one
piece of input information that conforms to a preset structure according to
the relevant
information and a preset rule, inputting the input information in a second
generating
model, generating at least one target text, assembling the target titles, the
target
introduction, and the target text, and obtaining plural target advertorials,
by means of the
deep learning and natural language processing technologies, it is made
possible to realize
automatic, intellectualized, and diversified generation of marketing
advertorials, save
input of operating personnel, enhance production efficiency of marketing
advertorials,
effectively avoid the problem concerning low efficiency by manual writing, and
avoid
the problem concerning the stereotyped style of template generation at the
same time.
[0054] 2. In the method of and device for generating a target advertorial
based on deep learning
as provided by the embodiments of the present invention, by subjecting plural
collected
first sample titles to a term-segmenting process, obtaining a second term-
segmenting
result, employing a preset first keyword extracting method to extract a first
keyword from
the first sample titles, inputting the second term-segmenting result and the
first keyword
in the third generating model, and obtaining plural new titles, currently
available and
limited titles are used to expand the number of titles in the title library.
[0055] 3. In the method of and device for generating a target advertorial
based on deep learning
6
Date Regue/Date Received 2022-06-29
CA 03166556 2022-06-29
as provided by the embodiments of the present invention, by subjecting plural
collected
second sample titles and introductions corresponding to the second sample
titles to a term-
segmenting process, employing a preset second keyword extracting method to
extract a
second keyword from the second sample titles, intersecting a second keyword
set with
each term-segmented second sample title, obtaining a target keyword,
traversing each
said second sample title, matching the target keyword with the totally term-
segmented
introductions corresponding to the second sample titles, obtaining a
successfully matched
introduction to serve as a new introduction of the current second sample
title, taking the
second sample title as input, taking the introduction corresponding to the
second sample
title and the new introduction as output, and training out the first
generating model based
on a preset algorithm, training data of an introduction generating model is
expanded, and
such problems as overfitting and inferior generation effects easily causable
by insufficient
training data are avoided.
BRIEF DESCRIPTION OF THE DRAWINGS
[0056] In order to describe the technical solutions in the embodiments of the
present invention
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
invention,
and it is possible for persons ordinarily skilled in the art to acquire other
drawings based
on these drawings without spending any creative effort in the process.
[0057] Fig. 1 is a flowchart illustrating the method of generating a target
advertorial based on
deep learning according to an exemplary embodiment;
[0058] Fig. 2 is a flowchart illustrating the step of generating at least one
piece of input
information that conforms to a preset structure according to the relevant
information and
a preset rule, inputting the input information in a second generating model,
and generating
at least one target text according to an exemplary embodiment;
[0059] Fig. 3 is a flowchart illustrating the process of constructing a title
library according to an
7
Date Regue/Date Received 2022-06-29
CA 03166556 2022-06-29
exemplary embodiment;
[0060] Fig. 4 is a flowchart illustrating the process of constructing a title
library according to
another exemplary embodiment;
[0061] Fig. 5 is a flowchart illustrating the process of constructing a first
generating model
according to an exemplary embodiment; and
[0062] Fig. 6 is a view schematically illustrating the structure of a device
for generating a target
advertorial based on deep learning according to an exemplary embodiment.
DETAILED DESCRIPTION OF THE INVENTION
[0063] In order to make more lucid and clear the objectives, technical
solutions, and advantages
of the present invention, the technical solutions in the embodiments of the
present
invention will be clearly and comprehensively described below with reference
to
accompanying drawings in the embodiments of the present invention. Apparently,
the
embodiments as described are merely partial, rather than the entire,
embodiments of the
present invention. All other embodiments obtainable by persons ordinarily
skilled in the
art based on the embodiments in the present invention shall all be covered by
the
protection scope of the present invention.
[0064] The method of generating a target advertorial based on deep learning
provided by the
present invention firstly retrieves adaptable titles from a title library
according to the
relevant information of a target object, then sequentially generates an
introduction and
marketing statements (namely marketing text) according to the matched titles
and the
relevant information, and finally assembles a target title, the introduction
and the
marketing text to output plural marketing advertorials. In the embodiments of
the present
invention, the Seq2Seq algorithm is employed to realize generation of the
introduction
and the marketing text, thus making it possible to effectively avoid the
problem
concerning low efficiency by manual writing, and avoid the problem concerning
the
stereotyped style of template generation at the same time. Seq2Seq is a
generation
framework made up of an encoder and a decoder, generates output sequence Y
according
8
Date Regue/Date Received 2022-06-29
CA 03166556 2022-06-29
to input sequence X, and is widely applied in such tasks as translation, text
automatic
abstraction, and robot automatic Q&A, etc.
[0065] Fig. 1 is a flowchart illustrating the method of generating a target
advertorial based on
deep learning according to an exemplary embodiment, as shown in Fig. 1, the
method
comprises the following steps.
[0066] Si - receiving relevant information of a target object, and matching
plural adaptable target
titles out of a title library according to the relevant information, wherein
titles in the title
library are derived by expansion of collected titles by a third generating
model.
[0067] Specifically, the target advertorial generally includes three sections,
namely a title, an
introduction, and a text. The target advertorial in the embodiments of the
present
invention includes a marketing advertorial, the marketing advertorial is taken
for example
to include three sections, namely a title, an introduction, and a text. The
relevant
information of the target object in the embodiments of the present invention
includes the
title of the product of the target advertorial to be generated, or descriptive
information of
the target object of the target advertorial to be generated. Moreover, in the
embodiments
of the present invention, the received relevant information can be input by a
user, and the
relevant information input by a user can be the title(s) of one or more
product(s) of a
certain category. After the relevant information of the target object input by
the user has
been received, plural target titles adapted to the relevant information are
matched out of
a title library according to a preset matching method (such as character
strings matching
after term segmentation, similarities matching, etc.), wherein titles in the
title library are
derived by expansion of collected titles by a third generating model. As
should be noted
here, the title matching method is not specifically defined in the embodiments
of the
present invention, and it is possible for the user to set the method according
to specific
requirement.
[0068] S2 - inputting the target title in a first generating model, and
generating at least one target
introduction.
9
Date Regue/Date Received 2022-06-29
CA 03166556 2022-06-29
[0069] Specifically, in the embodiments of the present invention, the first
generating model is a
natural language processing model pretrained by means of a preset algorithm
(such as the
5eq25eq algorithm). The input to the model is the aforementioned target title,
and the
output therefrom is the target introduction corresponding to the target title,
wherein the
number of target introduction(s) output from the first generating model can be
one or
more, to which no restriction is made in this context.
[0070] S3 - generating at least one piece of input information that conforms
to a preset structure
according to the relevant information and a preset rule, inputting the input
information in
a second generating model, and generating at least one target text.
[0071] Specifically, the second generating model is also a natural language
processing model
pretrained by means of a preset algorithm (such as the 5eq25eq algorithm). In
the
embodiments of the present invention, in order for the target text output from
the second
generating model to be diversified, this is realized by expanding the input to
the second
generating model. Accordingly, before the target text is generated, at least
one piece of
input information that conforms to a preset structure is firstly generated
according to the
relevant information and a preset rule, the obtained input information is
subsequently
input in the second generating model, and at least one target text is
generated. The "at
least one" means one or more.
[0072] S4 - assembling the target title, the target introduction, and the
target text, and obtaining
plural target advertorials.
[0073] Specifically, the target title as well as the target introduction and
target text obtained
through the foregoing step are finally assembled to obtain plural target
advertorials for
reference and selection by users.
[0074] With reference to Fig. 2, as a preferred mode of execution in the
embodiments of the
present invention, the step of generating at least one piece of input
information that
conforms to a preset structure according to the relevant information and a
preset rule,
inputting the input information in a second generating model, and generating
at least one
Date Regue/Date Received 2022-06-29
CA 03166556 2022-06-29
target text includes the following.
[0075] S101 - subjecting the relevant information to a term-segmenting
process, and extracting
target segmented terms that satisfy a preset condition from a first term-
segmenting result
as obtained.
[0076] Specifically, the relevant information is usually mostly of a structure
as "modifier +
category word", of which the modifier is a term expressing brand, function,
characteristic,
and property, etc. In the embodiments of the present invention, input to the
second
generating model is expanded by the mode of recombining the sequence of the
modifiers,
so that the target text output from the second generating model can be
diversified.
Accordingly, before the target text is generated, the relevant information
should be firstly
subjected to a term-segmenting process to obtain a first term-segmenting
result, and target
segmented terms that satisfy a preset condition are subsequently extracted
from the first
term-segmenting result. Since input to the second generating model is expanded
by the
mode of recombining the sequence of the modifiers, the target segmented terms
that
satisfy a preset condition here are segmented terms that pertain to modifiers
in the first
term-segmenting result.
[0077] S102 - recombining the target segmented terms, and obtaining at least
one piece of input
information that conforms to a preset structure.
[0078] Specifically, in the embodiments of the present invention, it is
possible to preset a
recombining mechanism according to practical requirement, such as to recombine
the
sequence of modifiers after term segmentation. The target segmented terms
obtained in
the foregoing step are thereafter recombined according to the recombining
mechanism to
output plural pieces of input information that conform to a preset structure.
By the same
token, the preset structure can be a structure as "modifier + category word",
whose setup
and adjustment can be made by the user according to practical requirement, and
no
specific restriction is made thereto in this context.
[0079] S103 - inputting the input information in the second generating model,
and generating at
11
Date Regue/Date Received 2022-06-29
CA 03166556 2022-06-29
least one target text.
[0080] Specifically, the input information obtained through the foregoing step
is finally input in
the second generating model, and at least one target text is generated.
[0081] Fig. 3 is a flowchart illustrating the process of constructing a title
library according to an
exemplary embodiment, with reference to Fig. 3, as a preferred mode of
execution in the
embodiments of the present invention, the process of constructing the title
library includes
the following.
[0082] S201 - subjecting plural collected first sample titles to a term-
segmenting process, and
obtaining a second term-segmenting result.
[0083] Specifically, in the embodiments of the present invention, after the
relevant information
of the target object has been received, adaptable target titles are obtained
by the mode of
matching from the title library according to the relevant information,
however, during the
process of constructing the title library, the number of titles actually
collected is rather
limited. To solve such a problem, in the embodiments of the present invention,
the number
of titles in the title library is increased by the mode of expanding the
limited number of
collected titles. During specific expansion of the titles, plural collected
first sample titles
are subjected to a term-segmenting process, and a second term-segmenting
result is
obtained.
[0084] S202 - employing a preset first keyword extracting method to extract a
first keyword from
the first sample titles.
[0085] Specifically, a preset first keyword extracting method is then employed
to extract a first
keyword from the sample titles, wherein it is possible for the user to set the
extraction
proportion of the first keyword(s) (namely the proportion of the first
keyword(s) in the
sample titles) according to practical requirement. As should be noted here, in
the
embodiments of the present invention, the first keyword extracting method is
not
specifically defined, and the user can set the method according to practical
requirement,
for instance, by employing a TS-IDF algorithm, etc.
12
Date Regue/Date Received 2022-06-29
CA 03166556 2022-06-29
[0086] S203 - inputting the second term-segmenting result and the first
keyword in the third
generating model, and obtaining plural new titles, wherein the title library
consists of the
new titles.
[0087] Specifically, the second term-segmenting result and the first keyword
obtained in the
foregoing steps are taken as input in the third generating model, the
resultant outputs (the
outputs are new titles) are expanded titles obtained according to the target
titles, and these
new titles make up the title library provided by the embodiments of the
present invention.
As should be noted here, in the embodiments of the present invention, in the
third
generating module can be employed a beam search decoder, thus making it
possible to
generate great quantities of titles.
[0088] Fig. 4 is a flowchart illustrating the process of constructing a title
library according to
another exemplary embodiment, with reference to Fig. 4, as a preferred mode of
execution in the embodiments of the present invention, the process of
constructing the
title library includes:
[0089] S301 - subjecting plural collected first sample titles to a term-
segmenting process, and
obtaining a second term-segmenting result;
[0090] S302 - employing a preset first keyword extracting method to extract a
first keyword from
the sample titles;
[0091] S303 - intersecting a first keyword set with the second term-segmenting
result, and
obtaining an input data set;
[0092] S304 - taking data of the input data set as input, taking the target
titles as output, and
training out the third generating model based on a preset algorithm; and
[0093] S305 - inputting the second term-segmenting result and the first
keyword in the third
generating model, and obtaining plural new titles, wherein the title library
consists of the
new titles.
[0094] Specifically, the third generating model here is likewise a natural
language processing
13
Date Regue/Date Received 2022-06-29
CA 03166556 2022-06-29
model pretrained by means of a preset algorithm (such as the Seq2Seq
algorithm). When
training data is prepared for the third generating model, it is possible to
calculate the
intersection of the first keyword set with the second term-segmenting result
obtained in
the foregoing steps, obtain an input data set, then take data of the input
data set as input,
and take the target titles as output to train out the third generating model
based on a preset
algorithm (such as the Seq2Seq algorithm). In addition, the specific
implementation
processes of steps S301, S302 and S305 can be inferred with reference to the
specific
implementation processes of the foregoing steps S201 to S203, while no
repetition is
redundantly made on a one-by-one basis.
[0095] Moreover, the model under different training states (namely different
steps or epochs) can
be further employed to repeat the foregoing steps, so as to further expand the
titles.
Merely with the aid of existing titles (the first sample titles), the method
employs a
particular extracting mode to construction input and output to train out the
third
generating model, so as to make it possible to obtain great quantities of
titles with flexible
sentence patterns in a short time, save manpower cost, and enhance production
efficiency.
[0096] Fig. 5 is a flowchart illustrating the process of constructing a first
generating model
according to an exemplary embodiment, with reference to Fig. 5, as a preferred
mode of
execution in the embodiments of the present invention, the method further
comprises a
process of constructing the first generating model, which process includes the
following.
[0097] S401 - subjecting plural collected second sample titles and
introductions corresponding
to the second sample titles to a term-segmenting process.
[0098] Specifically, in the embodiments of the present invention, a keyword
matching method is
employed to mine the inherent relation between titles and introductions, and
to match and
correspond one title to plural introductions, so that training data of the
first generating
model can be greatly expanded, such problems as overfitting and inferior
generation
effect caused by insufficient training data are avoided, and the generation
effect of the
first generating model is effectively enhanced. During specific
implementation, certain
amounts of title-introduction pairs are firstly collected in advance, that is,
plural second
14
Date Regue/Date Received 2022-06-29
CA 03166556 2022-06-29
sample titles and introductions corresponding to the second sample titles are
collected,
the second sample titles and introductions corresponding to the second sample
titles are
thereafter subjected to a term-segmenting process, and their term-segmenting
results are
respectively obtained.
[0099] S402 - employing a preset second keyword extracting method to extract a
second
keyword from the second sample titles.
[0100] Specifically, a preset second keyword extracting method is then
employed to extract a
second keyword from the second sample titles, wherein it is possible for the
user to set
the extraction proportion of the second keyword(s) (namely the proportion of
the second
keyword(s) in the sample titles) according to practical requirement. As should
be noted
here, in the embodiments of the present invention, the second keyword
extracting method
is also not specifically defined, and the user can set the method according to
practical
requirement, for instance, by employing a TS-IDF algorithm, etc.
[0101] S403 - intersecting a second keyword set with each term-segmented
second sample title,
and obtaining a target keyword.
[0102] Specifically, a target keyword is extracted from each second sample
title, during specific
implementation, it is possible to calculate the intersection of the second
keyword set with
each term-segmented second sample title, and to take the result obtained by
such
intersection calculation as a target keyword.
[0103] S404 - traversing each said second sample title, matching the target
keyword with the
totally term-segmented introductions corresponding to the second sample
titles, and
obtaining a successfully matched introduction to serve as a new introduction
of the
current second sample title.
[0104] Specifically, an optimal matching criterion is preset according to
practical requirement,
for instance, sorting is effected according to numbers of matched keywords,
and top-ten
introductions with the maximum numbers of matched keywords are selected to
serve as
the introductions to which the title corresponds. Each second sample title is
traversed,
Date Regue/Date Received 2022-06-29
CA 03166556 2022-06-29
target keywords of each second sample title are used to match in the
introductions after
total term segmentation, and plural successfully matched introductions are
obtained
according to the preset optimal matching criterion to serve as new
introductions of the
current second sample title, whereby the volume of data can be greatly
expanded.
[0105] S405 - taking the second sample title as input, taking the introduction
corresponding to
the second sample title and the new introduction as output, and training out
the first
generating model based on a preset algorithm.
[0106] Specifically, the first generating model is also a natural language
processing model
pretrained by means of a preset algorithm (such as the 5eq25eq algorithm).
Finally, the
second sample title is taken as input, the introduction corresponding to the
second sample
title and the new introduction expanded in the foregoing step are taken as
output, and the
first generating model is trained out based on the preset algorithm.
[0107] Fig. 6 is a view schematically illustrating the structure of a device
for generating a target
advertorial based on deep learning according to an exemplary embodiment, with
reference to Fig. 6, the device comprises:
[0108] a title matching module, for receiving relevant information of a target
object, and
matching plural adaptable target titles out of a title library according to
the relevant
information, wherein titles in the title library are derived by expansion of
collected titles
by a third generating model;
[0109] an introduction generating module, for inputting the target titles in a
first generating
model, and generating at least one target introduction;
[0110] a text generating module, for generating at least one piece of input
information that
conforms to a preset structure according to the relevant information and a
preset rule,
inputting the input information in a second generating model, and generating
at least one
target text; and
[0111] an information assembling module, for assembling the target titles, the
target introduction,
and the target text, and obtaining plural target advertorials.
16
Date Regue/Date Received 2022-06-29
CA 03166556 2022-06-29
[0112] As a preferred mode of execution in the embodiments of the present
invention, the text
generating module includes:
[0113] a first term-segmenting unit, for subjecting the relevant information
to a term-segmenting
process, and extracting target segmented terms that satisfy a preset condition
from a first
term-segmenting result as obtained;
[0114] a segmented terms recombining unit, for recombining the target
segmented terms, and
obtaining at least one piece of input information that conforms to a preset
structure; and
[0115] a text generating unit, for inputting the input information in the
second generating model,
and generating at least one target text.
[0116] As a preferred mode of execution in the embodiments of the present
invention, the device
further comprises a first constructing module that includes:
[0117] a second term-segmenting unit, for subjecting plural collected first
sample titles to a term-
segmenting process, and obtaining a second term-segmenting result;
[0118] a first extracting unit, for employing a preset first keyword
extracting method to extract
a first keyword from the first sample titles; and
[0119] a title generating unit, for inputting the second term-segmenting
result and the first
keyword in the third generating model, and obtaining plural new titles,
wherein the title
library consists of the new titles.
[0120] As a preferred mode of execution in the embodiments of the present
invention, the first
constructing module further includes:
[0121] a first intersecting unit, for intersecting a first keyword set with
the second term-
segmenting result, and obtaining an input data set; and
[0122] a first training unit, for taking data of the input data set as input,
taking the target titles as
output, and training out the third generating model based on a preset
algorithm.
[0123] As a preferred mode of execution in the embodiments of the present
invention, the device
17
Date Regue/Date Received 2022-06-29
CA 03166556 2022-06-29
further comprises a second constructing module that includes:
[0124] a third term-segmenting unit, for subjecting plural collected second
sample titles and
introductions corresponding to the second sample titles to a term-segmenting
process;
[0125] a second extracting unit, for employing a preset second keyword
extracting method to
extract a second keyword from the second sample titles;
[0126] a second intersecting unit, for intersecting a second keyword set with
each term-
segmented second sample title, and obtaining a target keyword;
[0127] an introduction expanding unit, for traversing each said second sample
title, matching the
target keyword with the totally term-segmented introductions corresponding to
the
second sample titles, and obtaining a successfully matched introduction to
serve as a new
introduction of the current second sample title; and
[0128] a second training unit, for taking the second sample title as input,
taking the introduction
corresponding to the second sample title and the new introduction as output,
and training
out the first generating model based on a preset algorithm.
[0129] In summary, the technical solutions provided by the embodiments of the
present invention
bring about the following advantageous effects.
[0130] 1. In the method of and device for generating a target advertorial
based on deep learning
as provided by the embodiments of the present invention, by receiving relevant
information of a target object, matching plural adaptable target titles out of
a title library
according to the relevant information, wherein titles in the title library are
derived by
expansion of collected titles by a third generating model, inputting the
target titles in a
first generating model, generating at least one target introduction,
generating at least one
piece of input information that conforms to a preset structure according to
the relevant
information and a preset rule, inputting the input information in a second
generating
model, generating at least one target text, assembling the target titles, the
target
introduction, and the target text, and obtaining plural target advertorials,
by means of the
deep learning and natural language processing technologies, it is made
possible to realize
18
Date Regue/Date Received 2022-06-29
CA 03166556 2022-06-29
automatic, intellectualized, and diversified generation of marketing
advertorials, save
input of operating personnel, enhance production efficiency of marketing
advertorials,
effectively avoid the problem concerning low efficiency by manual writing, and
avoid
the problem concerning the stereotyped style of template generation at the
same time.
[0131] 2. In the method of and device for generating a target advertorial
based on deep learning
as provided by the embodiments of the present invention, by subjecting plural
collected
first sample titles to a term-segmenting process, obtaining a second term-
segmenting
result, employing a preset first keyword extracting method to extract a first
keyword from
the first sample titles, inputting the second term-segmenting result and the
first keyword
in the third generating model, and obtaining plural new titles, currently
available and
limited titles are used to expand the number of titles in the title library.
[0132] 3. In the method of and device for generating a target advertorial
based on deep learning
as provided by the embodiments of the present invention, by subjecting plural
collected
second sample titles and introductions corresponding to the second sample
titles to a term-
segmenting process, employing a preset second keyword extracting method to
extract a
second keyword from the second sample titles, intersecting a second keyword
set with
each term-segmented second sample title, obtaining a target keyword,
traversing each
said second sample title, matching the target keyword with the totally term-
segmented
introductions corresponding to the second sample titles, obtaining a
successfully matched
introduction to serve as a new introduction of the current second sample
title, taking the
second sample title as input, taking the introduction corresponding to the
second sample
title and the new introduction as output, and training out the first
generating model based
on a preset algorithm, training data of an introduction generating model is
expanded, and
such problems as overfitting and inferior generation effects easily causable
by insufficient
training data are avoided.
[0133] As should be noted, when the device for generating a target advertorial
based on deep
learning as provided by the foregoing embodiment triggers a target advertorial
generating
business, it is merely exemplarily described by being divided into the
aforementioned
19
Date Regue/Date Received 2022-06-29
CA 03166556 2022-06-29
various functional modules, whereas in practical application, it is possible
to assign these
functions to different functional modules to be completed there according to
requirements,
that is to say, the internal structure of the device can be divided into
different functional
modules to complete the entire or partial functions mentioned above. In
addition, the
device for generating a target advertorial based on deep learning as provided
by the
foregoing embodiment pertains to the same conception as the method of
generating a
target advertorial based on deep learning as provided by the method
embodiment, that is
to say, the device is based on the method of generating a target advertorial
based on deep
learning ¨ see the method embodiments for details of its specific
implementation process,
while no repetition is made in this context.
[0134] As can be understood by persons ordinarily skilled in the art, the
entire or partial steps
that realize the aforementioned embodiments can be completed via hardware, and
can
also be completed via a program that instructs relevant hardware, the program
can be
stored in a computer-readable storage medium, and the storage medium can be a
read-
only memory, a magnetic disk or an optical disk, etc.
[0135] What the above describes is merely directed to preferred embodiments of
the present
invention, and is not meant to restrict the present invention. Any
modification, equivalent
substitution, and improvement makeable within the spirit and principle of the
present
invention shall all be covered by the protection scope of the present
invention.
Date Regue/Date Received 2022-06-29
