CN114328899A

CN114328899A - Text summary generation method, device, equipment and storage medium

Info

Publication number: CN114328899A
Application number: CN202111667181.XA
Authority: CN
Inventors: 高建清; 戚婷; 闫莉; 孙境廷
Original assignee: iFlytek Co Ltd
Current assignee: iFlytek Co Ltd
Priority date: 2021-12-30
Filing date: 2021-12-30
Publication date: 2022-04-12
Also published as: WO2023124648A1

Abstract

The application provides a text summary generation method, a device, equipment and a storage medium, wherein the method comprises the following steps: acquiring a target text and a reference text, wherein the reference text is determined based on the target text content concerned by a user; and carrying out summary generation processing on the target text based on the associated content of the reference text positioned in the target text to obtain a target text summary corresponding to the reference text. By adopting the text summary generation method, even for the same target text, when the reference texts are different, the text summary generation processing with different emphasis points can be carried out on the target text by positioning the text contents related to the reference texts from the target text, so as to obtain the target text summary corresponding to the reference text. Therefore, the method can generate the text summary meeting the requirements of different users aiming at the same target text.

Description

Text summary generation method, device, equipment and storage medium

Technical Field

The present application relates to the field of artificial intelligence technologies, and in particular, to a text summary generation method, apparatus, device, and storage medium.

Background

The text summary generation means that content extraction is carried out on a long text, so that information capable of representing the core content of the text is extracted, and the text summary can help people to directly and effectively master the text content.

Conventional text summary generation schemes are typically based on text auto-summarization techniques, extracting points from text and forming generalized text. The text automatic summarization technology may be divided into a decimated summary and a generated summary in a manner of generating a summary. The extraction type abstract is to extract words or sentences from an original text without any change to form an abstract, and the content of the abstract is completely from the original text; the generated abstract allows new words and phrases which are not in the original text to be generated to form the abstract, the abstract is generated by firstly performing semantic understanding on the text content, and generating a section based on the semantic to summarize the given text.

In general, the content of the target text that needs to generate a text summary is multifaceted, and different people may be interested in different aspects of the content, and therefore, the text summary requirements of different people for the same target text are different.

However, the current text summary generation scheme, no matter the abstract summary or the generated summary, cannot generate text summaries of different contents according to different requirements of people, and cannot meet the requirements of different people on the generation of the text summary of the same target text.

Disclosure of Invention

Based on the technical current situation, the application provides a text summary generation method, a text summary generation device, text summary generation equipment and a storage medium.

In order to achieve the above purpose, the present application proposes the following technical solutions:

a text summary generation method, comprising:

acquiring a target text and a reference text, wherein the reference text is determined based on the target text content concerned by a user;

and carrying out summary generation processing on the target text based on the associated content of the reference text positioned in the target text to obtain a target text summary corresponding to the reference text.

A text summary generation apparatus comprising:

the data acquisition unit is used for acquiring a target text and a reference text, wherein the reference text is determined based on the target text content concerned by a user;

and the document generation unit is used for performing document generation processing on the target text based on the associated content of the reference text positioned in the target text to obtain a target text document corresponding to the reference text.

A text summary generation device comprising:

a memory and a processor;

the memory is connected with the processor and used for storing programs;

the processor is used for realizing the text summary generation method by operating the program in the memory.

A storage medium having stored thereon a computer program which, when executed by a processor, implements the text summary generation method described above.

According to the text summary generation method, when a text summary is generated for a target text, a reference text is used as a reference for generating the text summary, and the text summary is generated and processed for the target text by positioning the associated content of the reference text from the target text, so that the target text summary corresponding to the reference text is obtained. When a text summary is generated for a target text, the method jointly applies the target text content and the reference text associated content in the target text to jointly determine the text summary of the target text. By adopting the text summary generation method, even for the same target text, when the reference texts are different, the text summary generation processing with different emphasis points can be carried out on the target text by positioning the text contents related to the reference texts from the target text, so as to obtain the target text summary corresponding to the reference text. Therefore, the method can generate the text summary meeting the requirements of different users aiming at the same target text.

Drawings

In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced below, it is obvious that the drawings in the following description are only embodiments of the present application, and for those skilled in the art, other drawings can be obtained according to the provided drawings without creative efforts.

Fig. 1 is a schematic flowchart of a text summary generation method according to an embodiment of the present disclosure;

FIG. 2 is a schematic structural diagram of a chapter interaction semantic retrieval model provided in an embodiment of the present application;

FIG. 3 is a schematic structural diagram of a semantic retrieval model enhanced based on retrieval candidate information according to an embodiment of the present application;

FIG. 4 is a schematic structural diagram of a text summary generation model based on an attention mechanism according to an embodiment of the present application;

FIG. 5 is a schematic structural diagram of another attention-based text summary generation model provided in an embodiment of the present application;

FIG. 6 is a schematic structural diagram of a word-sentence-chapter hierarchy information coding model provided in an embodiment of the present application;

FIG. 7 is a schematic diagram of word segmentation feature extraction provided in the embodiments of the present application;

FIG. 8 is a schematic structural diagram of an information fusion model provided in an embodiment of the present application;

fig. 9 is a schematic structural diagram of a text summary generation apparatus according to an embodiment of the present application;

fig. 10 is a schematic structural diagram of a text summary generation device according to an embodiment of the present application.

Detailed Description

The technical scheme of the embodiment of the application is suitable for generating the application scene of the text summary, and by the adoption of the technical scheme of the embodiment of the application, the text summary which accords with the user concern can be generated, so that the text summary requirements of different users can be met.

The application scene for generating the text summary specifically refers to a scene needing to generate summary content, and includes, but is not limited to, specific application scenes such as conference summary generation, document summary generation, news gist extraction, and the like.

For example, in a meeting scenario, meeting content is often multifaceted, while content of interest to different participants is often not the same. For example, in a new product planning workshop, the responsible persons such as company design department, product department, and market department participating in the workshop pay attention to different contents. For example, the design department is concerned more about the perfection of the product design scheme, the product department is concerned more about the product definition and development planning, and the market department is concerned more about the market positioning of new products. Thus, the content of the conference summary required by different departments is different.

However, in the conventional text summary generation scheme, no matter the abstract summary or the generated summary, since the text to be summarized is only subjected to technical processing, the main content of the text is determined, and text summaries with different emphasis points cannot be generated for different attention points.

Based on the technical current situation, the embodiment of the application provides a text summary generation scheme, and the scheme can generate a text summary for a target text by referring to the target text content concerned by a user, so that different text summaries can be generated aiming at different concerned points, and the personalized requirements of different users for the text summary content are met.

In the subsequent embodiments of the present application, a conference summary is taken as an example to introduce specific processing contents of the technical solution of the embodiments of the present application, and when the technical solution of the embodiments of the present application is applied to other scenes, a specific execution process of the technical solution of the embodiments of the present application may refer to the introduction of the embodiments of the present application. It should be noted that, the technical scheme of the embodiment of the present application is not only suitable for generating the text summary to obtain the summary of the text form, but also suitable for generating the speech summary to obtain the summary of the text or the speech form, or generating the text summary to obtain the summary of the speech form. When generating the summary of the data in the non-text form or generating the summary of the non-text form, the generation may be implemented by converting the data in the non-text form into text data or converting the generated summary of the text form into the non-text form, where the summary generates the main processing, and reference may still be made to the description of the embodiment of the present application.

The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application, and it is obvious that the described embodiments are only a part of the embodiments of the present application, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present application.

The embodiment of the application provides a text summary generation method, and as shown in fig. 1, the method includes:

s101, acquiring a target text and a reference text.

The target text refers to a text that needs to be generated as an abstract, and the target text may be a text of any content and any language acquired through any way. Specifically, the target text may be a directly acquired text such as an academic document, a news article, a book, or the like, or a text obtained by voice recognition such as a recognition result text obtained by voice recognition of a meeting recording, a recognition result text obtained by recognition of a speech of a speaker, or the like. Theoretically, data contents in any form can be converted into a text form, so that the text can be used as the target text, and the summary generation of the target text is realized through subsequent processing.

The reference text is determined based on the target text content focused by the user. Specifically, the reference text can represent the interest or interest points of the user for the target text content, and simultaneously represent the requirement of the user for the generated text summary content, and is used for providing reference for the text summary generating the target text, so that the target text summary which meets the interest points of the user or contains the interest content of the user can be generated.

The reference text may be input by a user or may be preset before text summary generation is performed. The specific form of the reference text can be a fixed sentence pattern, a keyword or phrase, or a short text sentence or text segment, and even a logical combination of a plurality of texts. For example, the reference text may be a combination of search conditions, such as a related and B related (denoted as a & B), a related or B related (denoted as a | | B), or a related but B unrelated (denoted as a-B), etc., or even a more complex combination of conditions, such as a related and B related but C unrelated (denoted as { a & B } -C), where the specific form of a, B, C is not required, and may be a phrase, keyword, sentence, or other form.

For example, in a conference summary generation scene, a conference recording is subjected to speech recognition processing, so as to obtain a text corresponding to the conference recording, and the text is used as the target text. Meanwhile, reference texts which are input by the user and represent the conference contents which are interested or concerned by the user are obtained. The reference text can be phrases, short sentences summarized by the user based on the conference content, or simple conference records recorded by the user at the conference place, or keywords, phrases, retrieval conditions and the like determined by the user based on the desired conference summary content. When a target text corresponding to the conference and a reference text input by the user are obtained, subsequent text summary generation processing can be executed, and the target text is subjected to summary generation processing to obtain a target text summary which accords with the user concern or contains the content interested by the user.

When the processing device processes the text, the processing device actually processes the text features. Therefore, the "obtaining the target text and the reference text" may be obtaining an original text of the target text and the reference text, and then performing feature extraction on the obtained target text and the original text of the reference text to obtain features of the target text and features of the reference text for subsequent text summary generation processing; or, the features of the target text and the reference text may be directly obtained for the subsequent text summary generation processing.

S102, based on the associated content of the reference text positioned in the target text, carrying out summary generation processing on the target text to obtain a target text summary corresponding to the reference text.

The content associated with the reference text refers to a text content associated with the reference text, for example, a text content with a similarity greater than a set similarity threshold value with the reference text, or a text content similar to or associated with the semantics of the reference text, which can be used as the content associated with the reference text.

For example, in the embodiment of the application, the reference text is sequentially compared with each text segment of the target text, and the text similarity or semantic similarity between the reference text and each text segment of the target text is determined, so that the correlation between the reference text and the target text segment is determined, and the related content related to the reference text is positioned and identified from the target text.

It can be understood that the reference texts in the target texts are associated with contents, and the effect value for generating the text summary corresponding to the reference texts is greater. Because the reference text associated content contains the text information related to the reference text, if the associated content of the reference texts can be considered in important when the text summary generation is carried out on the target text, the finally generated text summary can contain more reference text associated information, and therefore the generated target text summary is matched with the reference text.

Based on the above ideas, in the embodiment of the application, when the text summary of the target text is generated, the text summary of the target text is generated by mainly using the reference text associated content in the target text and by using other content in the target text as an auxiliary, so that the proportion of the content associated with the reference text in the finally generated target text summary is higher, and thus the relevance between the finally generated target text summary and the reference text is higher, that is, the target text summary corresponding to the reference text is obtained.

As an exemplary implementation manner, in the embodiment of the present application, a text fragment related to a reference text is located from a target text, and full-text content of the target text is subjected to summary generation processing, so as to obtain a target text summary corresponding to the reference text.

The text segments can be text sentences, text paragraphs or text phrases, etc. According to the embodiment of the application, text segments related to the reference text are identified from the target text through text comparison, semantic comparison and other methods. For example, as long as the text similarity or semantic similarity between the text segment in the target text and the reference text is not 0, it may be considered to be valuable for generating the target text corresponding to the reference text, and thus it is determined as the text segment related to the reference text.

And then, generating text summary for the full-text content of the target text by combining the positioning result of the text segment related to the reference text. In the generation summary process, the contribution degree of the text segment related to the reference text to the generation of the target text summary is set to be higher than the contribution degree of the text segment unrelated to the reference text to the generation of the target text summary, so that the generated target text summary contains more information of the text segment related to the reference text. Furthermore, different contribution degrees can be set for each text segment related to the reference text according to the degree of correlation between the text segment and the reference text, so that the finally generated target text summary has higher degree of correlation with the reference text.

For example, suppose that the conference original text of a certain conference contains 200 segments of session texts, as shown in the following conference original text (for the sake of simplicity, ellipses are used to represent omitted parts of the conference original text content because the conference original text content is long):

meeting original text:

paragraph 1: today we mainly discuss the marketing plan for new products on the wand.

Paragraph 2: the plan is that 3 months in 2019 are started to be the tour exhibition of about four months, and the scientific and technological innovation date is developed.

…………

Paragraph 16: the following is a comprehensive analysis of product trends to help the product stand out from the intense market. We will select 14 cities to actively find self-media collaboration, the collaboration content including … …

…………

Paragraph 68: youngsters prefer a very cool design and we are to take care of this part of the market to find a point of entry. Promotional effect … …

…………

Paragraph 79: now there is a demo to see. Before the twenty-one release, we performed a more systematic test on the product. The accuracy on the test set can be as high as 98 percent, which exceeds the relative 30 percent of the competitive products and is enough to form the generation difference.

…………

Paragraph 88: the test set data test, no effect was measured. Because the comparison with the competitive products is complete, the comparison is really made on the same scene. If the same set, different people experience different.

…………

Paragraph 88: product orientation should be more young, adding this design to … …

…………

Paragraph 162: the interface design also needs to be interactive and reasonable, the key attribute, jump and other links … … are considered, the aforementioned test problem, the user subjective experience item is also considered, and some subjective experience contrast needs to be set. And the using effect of the product is sensed under special conditions, such as smoothness, multi-row crossing selection and the like.

…………

Paragraph 200: this is the case in conferences today, where later departments are concerned about coordination.

Taking the original conference text as a target text, assuming that the retrieval condition input by the user a is "subjective experience of product effect", taking the retrieval condition input by the user a as a reference text, and aiming at the determined target text and the reference text, by executing the processing of the text summary generation method provided by the embodiment of the application, it can be determined that, in the original conference text, the text content related to the retrieval condition "subjective experience of product effect" input by the user a is "correct, the above-mentioned test problem, and we should also consider the item of subjective experience of the user, and need to set some contrast of subjective experience. And' feeling the using effect of the product for special conditions, such as smoothness, multi-row crossing selection and the like. "through locating out the text content relevant with retrieval condition" product effect subjective experience "from meeting original text, carry out summary generation to this meeting original text and handle, finally obtain meeting summary" the effect of the new product of wand not only sees the data result, still need consider some schemes of subjective experience. Therefore, the finally obtained conference summary is matched with the retrieval condition 'product effect subjective experience' input by the user A, namely the conference summary is the conference summary of the relevant information representing the product effect subjective experience. Therefore, by adopting the technical scheme of the embodiment of the application, the conference summary corresponding to the retrieval condition input by the user A can be generated by positioning the text content related to the retrieval condition input by the user A from the conference original text, so that the conference summary requirement of the user A is met.

For another example, for the above-mentioned original text of the conference, assuming that the search condition input by the user B is "subjective experience of product effect & product location", the original text of the conference is taken as the target text, and the search condition input by the user B is taken as the reference text, and by performing the processing of the text summary generation method proposed in the embodiment of the present application, it can be determined that, in the above-mentioned original text of the conference, the text content related to the search condition "subjective experience of product effect & product location" input by the user B is "product location should be considered more young people, and this part of the design is added. "in the above test problem, we should also consider the subjective experience of the user, and need to set some comparison of subjective experiences. And' feeling the using effect of the product for special conditions, such as smoothness, multi-row crossing selection and the like. "through locating out the text content relevant with retrieval condition" product effect subjective experience & product location "from the meeting original text, carry out summary generation processing to this meeting original text, finally obtain the meeting summary" the new product location of wand should consider that the young person prefers the design of cool appearance, the result of use not only sees the data result, also needs to consider some schemes of subjective experience. Therefore, the finally obtained conference summary is matched with the retrieval condition 'product effect subjective experience & product positioning' input by the user B, namely the conference summary is the conference summary of the relevant information representing the product effect subjective experience and the product positioning. Therefore, by adopting the technical scheme of the embodiment of the application, the conference summary corresponding to the retrieval condition input by the user B can be generated by positioning the text content related to the retrieval condition input by the user B from the conference original text, so that the conference summary requirement of the user B is met.

Therefore, by adopting the text summary generation method provided by the embodiment of the application, even for the same original text of the conference, when the retrieval conditions input by the user are different, the text summary matched with the user retrieval conditions can be generated respectively for different user retrieval conditions, so that the requirements of different users on the conference content are met.

In summary, in the text summary generation method provided in the embodiment of the present application, when generating a text summary for a target text, a reference text is used as a reference for generating the text summary, and the target text is summarized and generated by locating associated content of the reference text from the target text, so as to obtain a target text summary corresponding to the reference text. When a text summary is generated for a target text, the method jointly applies the target text content and the reference text associated content in the target text to jointly determine the text summary of the target text. By adopting the text summary generation method, even for the same target text, when the reference texts are different, the text summary generation processing with different emphasis points can be carried out on the target text by positioning the text contents related to the reference texts from the target text, so as to obtain the target text summary corresponding to the reference text. Therefore, the method can generate the text summary meeting the requirements of different users aiming at the same target text.

As an exemplary implementation manner, in the embodiment of the present application, a text summary generation model based on an attention mechanism is trained in advance, and is used for performing text summary generation processing on a target text according to the target text and a reference text to obtain a target text summary corresponding to the reference text.

The text summary generation model based on the attention mechanism is obtained by training based on pre-collected target text-reference text-target text summary parallel data. For example, a large amount of parallel data of the original text of the conference, the user retrieval condition, and the summary text of the conference are collected in advance, and are used for training the model after data preprocessing.

The conference original text data can collect conference audio data, and corresponding text data can be obtained after voice transcription is carried out on the conference audio data; of course, the meeting text data can also be directly collected, such as the original full manuscript finished by the shorthand for the meeting. The search condition, form, is not limited to fixed sentence pattern of fixed template, but also supports user to input the key word or phrase concerned by the user, or short meeting record, even logic combination of multiple sub-conditions, such as a related and B related (denoted as a & B), a related or B related (denoted as a | | | B) and a related but B unrelated (denoted as a-B), even more complex condition combination, such as a related and B related but C unrelated (denoted as { a & B } -C), where the specific form of a, B, C is not required, and may be phrase, key word, sentence or other form. In the embodiments of the present application, a specific source or specific content of the search condition is not limited. The conference summary aims to highly condense original long conference texts highly related to the retrieval conditions and cover and summarize the important contents of the conference related to the retrieval conditions.

The data preprocessing is to perform sentence division processing on the original text of the conference. The clauses can be divided into clauses or whole sentences according to punctuation marks, or can be divided into clauses according to fixed word number windows and sliding windows, the specific sentence dividing method is not specifically required, and the punctuation marks are adopted in the scheme to divide the original text of the conference according to the whole sentence dividing mode; secondly, the text data of the original conference after the sentence division is processed into an input sequence form of the participles, the text participles can be processed by the prior art, and are not detailed here, for example, the following key work of the text sentence of the original conference is a follow-up of spring meeting. "post-word result" next/major/work/yes/spring meet/one/follow-up. ". For the user search condition, if the search condition is only a single condition, the search condition is directly regarded as a plain text, the preprocessing only needs to process the text data of the search condition into an input sequence form of participles, and if the search condition is a composite of a plurality of sub-conditions, each sub-condition is processed into a sequence form after the participles. For a corresponding conference summary, the preprocessing is processing the text data into a sequential form of segmented words.

The text summary generation model based on the attention mechanism is a time sequence output model, namely, one word segmentation is output in each decoding process, and finally obtained word segmentations can be combined to obtain a summary text. During training, the characteristics of a participle sequence of a conference original text and the characteristics of a participle sequence of a retrieval condition are respectively obtained, the characteristics of the participle sequence of the conference original text and the characteristics of the participle sequence of the retrieval condition are input, a text summary generation model based on the attention mechanism is input, model loss is determined by comparing the conference summary participle sequence output by the model with a conference summary participle sequence obtained by pre-processing, model parameters are corrected based on the model loss, so that the text summary generation model based on the attention mechanism can generate a target text summary corresponding to a reference text by taking the characteristics of the target text and the reference text as input.

Next, taking generation of a meeting summary meeting different user retrieval conditions as an example, a specific processing procedure of the text summary generation method provided in the embodiment of the present application is introduced. In the following embodiments, a conference original text is used to represent the target text, a search condition or a search sub-condition input by a user is used to represent the reference text, a text summary is generated for the conference original text by introducing the process of obtaining a text summary corresponding to the search condition input by the user, and a specific processing procedure of a technical scheme of the embodiment of the application for generating the target text summary corresponding to the reference text based on positioning the associated content of the reference text from the target text is shown.

First, as for the "acquisition target text and the reference text", as shown above, the original text of the target text and the reference text may be acquired, and then feature extraction is performed thereon for subsequent text summary generation processing. Or, the features of the target text and the features of the reference text may be directly obtained and used for the subsequent text summary generation processing.

For a specific process of extracting the target text feature and the reference text feature, reference may be made to the detailed description of the following embodiments.

Then, as an exemplary implementation manner, a processing step of "performing summary generation processing on the target text based on the associated content of the reference text located from the target text to obtain a target text summary corresponding to the reference text" in the text summary generation method provided in the embodiment of the present application may be implemented by the following steps a1-a 2:

a1, determining the relevance of each text segment in the target text and the reference text, and locating the text segment relevant to the reference text from the target text.

Specifically, the text segment may be a text content with any granularity, such as a text sentence, a text segment, or a text phrase. The embodiment of the application divides the target text into text sentences, and each divided text sentence is used as the text segment. The text sentence division can be carried out on the target text according to punctuations of the target text, or the sentence division can be carried out according to a fixed word number window and a sliding window. In the embodiment of the application, sentence division is performed on the target text according to punctuations in the target text.

And then, performing text comparison or semantic comparison on the reference text and each divided target text sentence, and determining the correlation degree of the reference text and each target text sentence. In the embodiment of the application, the relevancy between each text segment in the target text and the reference text is determined through semantic measurement, that is, the relevancy between the reference text and each text segment in the target text is determined by comparing semantic similarities between the reference text and each text segment in the target text.

Illustratively, the features of the target text and the features of the reference text are obtained separately. And then, respectively determining the correlation degree of each text segment in the target text and the reference text according to the characteristics of each text segment in the target text and the characteristics of the reference text. According to the relevance between the reference text and each text segment in the target text, the text segment relevant to the reference text can be positioned from the target text. For example, in the target text, the text segment having a non-zero degree of correlation with the reference text is the text segment related to the reference text.

The features of the target text may be extracted from each text segment in the target text, and then the features of each text segment are combined to obtain the features of the target text, so that the overall features of the target text and the features of each text segment of the target text may be determined based on the features of each text segment in the target text. Or, the feature of the target text can be directly extracted from the whole target text to obtain the features of the target text, and then the features of the corresponding positions are intercepted from the features of the target text according to the positions of the text segments of the target text in the target text to obtain the features of the text segments.

In the embodiment of the present application, the features of the target text are determined by extracting the features of each text segment of the target text, and the features of the reference text are determined by extracting the features of each text segment of the reference text, and a specific feature extraction processing procedure will be described in the following embodiments.

As an alternative implementation, for example, in an application scenario generated by a conference summary, assuming that a user inputs a plurality of search sub-conditions, semantic similarity between each search sub-condition and each text in a conference original text is determined for each search sub-condition. At this time, for each retrieval sub-condition, the semantic similarity between the retrieval sub-condition and each sentence of text in the conference original text is determined, which can be realized by the above processing procedure for determining the semantic similarity between the text segment of the target text and the reference text.

For example, the embodiment of the present application constructs a semantic fuzzy retrieval model, and extracts a semantic similarity score between each sub-condition of the user retrieval condition and each text in the conference original text, that is, determines the correlation between the reference text and each text segment in the target text.

The semantic fuzzy retrieval model takes a retrieval sub-condition text and a conference original text as input, and outputs each sentence of text in a certain retrieval sub-condition text and the conference original text and the semantic similarity score of the text. Based on the semantic fuzzy retrieval model, after the conference original text and the text sequence of the user retrieval sub-condition are input into the semantic fuzzy retrieval model, the semantic similarity scores of 1 to n (n is the total number of the sentences of the conference original text) sentences in the conference original text and the search sub-condition A for the search sub-condition A and the n sentences of the conference original text can be obtained

Representing, for the search sub-condition B, the semantic similarity score with n sentences of the conference original text

And (4) showing.

The embodiment of the application provides two semantic fuzzy retrieval model frameworks: the system comprises a chapter interaction semantic retrieval model and a semantic retrieval model enhanced based on retrieval candidate information, wherein the semantic retrieval model is used for measuring semantic similarity scores of retrieval sub-conditions (reference texts) and a conference original text (target text) sentence.

As shown in fig. 2, the chapter interaction semantic retrieval model includes a word encoder, a sentence encoder, and a retrieval sub-condition text and conference text interaction module. The word encoder adopts a BERT pre-training model, firstly inputs a conference transcription text sentence S, and searches a sub-condition Q into the word encoder (wherein Q ═ { w ═ w%₁,w₂,…,w_nDenotes n participles, S, contained in the search sub-condition Q_j＝{w_j,1,w_j,2,…w_j,mRepresenting m participles contained in the jth conference text sentence), obtaining context coding vectors of each word in the sentence, and extracting [ CLS }]Word vector characterization of

And

as a transcription text sentence code and a retrieval sub-condition sentence code, for representing information of the whole sentence.

Then, the transliteration text sentence coding introduces context information into the current sentence through modeling of a two-layer Transformer sentence coder, and supplements the omitted information of the current sentence bearing context, thereby obtaining more accurate sentence representation

Secondly, through the interactive module of the search sub-condition text and the conference text, each of the conference original texts is searchedAnd carrying out interactive operation based on an attention mechanism to obtain the sentence codes of the retrieval sub-conditions after the information is complete. The interactive module for the search sub-condition text and the conference text is composed of an attention structure and codes the search sub-condition coding sentences

The method comprises the following steps that an inquiry Q used as an attention mechanism takes conference transcription sentence codes as K and V, so that conference content information is merged into a search sub-condition sentence code Q, and conference text content can well supplement information omitted in the search sub-condition under the condition that the search sub-condition is short or information such as description is not complete; meanwhile, the code q has a global view of the whole conference content, and is more beneficial to selecting a better retrieval result in the transcribed text.

Finally, the search sub-condition sentence code q is coded with each conference text sentence

Splicing is carried out to generate a final interaction vector S_jAnd inputting the semantic similarity score of the sentence and the retrieval sub-condition into an output layer for predicting.

In the above formula, the first and second carbon atoms are,

representing the dot product operation of the retrieval sub-condition and the text sentence j, representing the similarity degree of the two;

the information difference indicating the search sub-condition and the text sentence j can obtain more comprehensive similarity determination information by comparing a plurality of view angles.

Semantic retrieval model based on retrieval candidate information enhancement compared with discourse interactive semantic retrieval modelThe conditions of the rigging are modified interactively with the text sentence. In particular, sentence encoding upon obtaining search sub-conditions

And obtaining sentence codes of each text sentence of the original text of the conference

And finally, carrying out interactive operation based on an attention mechanism on the sentence codes of the N selected text sentences and the sentence codes of the retrieval sub-conditions to obtain the sentence codes of the retrieval sub-conditions with complete information.

E.g. by text sentence coding of a meeting

Calculating cosine distance to obtain TopN initial matching retrieval results most similar to the retrieval sub-conditions, namely, selecting N text sentences with highest similarity to the retrieval sub-conditions from the conference original text according to the similarity of each text sentence in the conference original text to the retrieval sub-conditions. As shown in fig. 3, r ═ 1 denotes selected TopN preliminary matching sentences, and r ═ 0 denotes other less relevant sentences.

Then, the sub-conditional sentence codes are retrieved

Match results with TopN high quality

And performing interaction through attention, and updating the code vector of the search sub-condition to be q. When the N setting is small, such as 2, the confidence of the N retrieval candidates related to the retrieval sub-condition is high, andto supplement the information of the search sub-condition more accurately; meanwhile, the N searches with higher quality are used as the prompt information, so that the model can be guided to select the results similar to the initial high-quality search results, the phenomenon that the semantic correlation difference between the finally selected search results is too large is avoided, and the user experience is reduced.

In consideration of the advantage of the BM25 in the exact matching scenario, the embodiment of the present application fuses the BM25 scheme with the output result of the semantic retrieval model. Specifically, according to the encoding of each text sentence in the conference original text and the encoding of the retrieval sub-condition, the semantic similarity between each text sentence in the conference original text and the retrieval sub-condition is calculated and determined through the BM25 algorithm. And then, performing fusion processing, for example, weighting fusion, on the semantic similarity score between each text sentence in the conference original text output by the semantic fuzzy retrieval model and the retrieval sub-condition, and the semantic similarity between each text sentence in the conference original text and the retrieval sub-condition calculated and determined by the BM25 algorithm to obtain the semantic similarity score between each text sentence in the fused conference original text and the retrieval sub-condition, that is, obtaining the correlation between each text sentence in the fused conference original text and the retrieval sub-condition.

In addition, in consideration of the fact that the sentences related to the retrieval sub-conditions are not in accordance with the fact that the description contents of the actual retrieval sub-conditions are concentrated when the spans of the sentences related to the retrieval sub-conditions are too large in the conference text, the embodiment of the present application restricts the spans of the position distribution of the original text sentences of the conference related to the retrieval sub-conditions, that is, corrects the correlation degree between each text sentence in the conference text and the retrieval sub-conditions according to the position distribution of each text sentence in the conference text.

In the embodiment of the application, a second number of text segments with highest correlation degree with the retrieval sub-conditions are selected from all text sentences in the conference text;

then, determining the punishment degree of the correlation degree of the other text sentences in the conference text and the retrieval sub-condition according to a rule that the larger the distance between the other text sentences in the conference text and the selected second number of text sentences is, the higher the punishment degree of the correlation degree of the other text sentences and the retrieval sub-condition is; and punishment is carried out on the correlation degrees of other text sentences in the conference text and the retrieval sub-conditions according to the punishment degree of the correlation degrees of other text sentences in the conference text and the retrieval sub-conditions.

Specifically, the embodiment of the present application sets that the Top2 of the semantic similarity score between each text sentence in the merged original conference text and the search sub-condition is relatively accurate, and the semantic similarity scores between other text sentences in the original conference text and the search sub-condition should be penalized correspondingly according to the distance Top-2. Because Top2 has two sentences, the penalty suffered by the semantic similarity score of other text sentences and retrieval sub-conditions should only be selected to carry out penalty with smaller distance from Top 2; based on the processing, the semantic similarity score of other text sentences in the conference original text and the retrieval sub-condition is the original score minus the distance penalty.

Furthermore, the embodiment of the present application further performs an abnormal text sentence filtering process on the text sentence identified from the original text of the conference and related to the retrieval sub-condition, that is, further accurately determines the similarity score between the conference text sentence and the retrieval sub-condition. Firstly, selecting a third number of text sentences with highest correlation degree with the retrieval sub-conditions from the conference text according to the correlation degree of each text sentence in the conference text with the retrieval sub-conditions; then, selecting a text sentence, of which the correlation degree with the retrieval sub-condition is greater than a first correlation degree threshold value or the correlation degree with the retrieval sub-condition is greater than a second correlation degree threshold value and the normalized correlation degree with the retrieval sub-condition is greater than a third correlation degree threshold value, from the third number of text sentences as text sentences related to the retrieval sub-condition according to the correlation degree of each text sentence with the retrieval sub-condition in the selected third number of text sentences;

wherein the first correlation threshold is greater than the second correlation threshold, and the second correlation threshold is greater than the third correlation threshold.

Specifically, after extracting a text sentence of the degree of correlation TopK with the search sub-condition from the conference text, the text sentence of which the number of non-stop words is 1 or less is deleted to filter a low information amount sentence. If the similarity score of the remaining text sentence and the retrieval sub-condition is less than t1 (for example, t1 is 0.6), the text sentence is considered to be low in reliability, and the text sentence is considered to be deleted; if the similarity score of the remaining text sentences and the search sub-condition is low, for example, both are less than t1, it is difficult to specify the content of the search sub-condition, such as being too short or having strong generalization, so the threshold t2 is adjusted downward (e.g., t2 ═ 0.3), and the normalized score _ norm of the remaining text sentences and the search sub-condition is further checked, and the threshold t3 is set (e.g., t3 ═ 0.2). Finally, only the text sentences whose similarity score with the retrieval sub-condition is greater than t1, or whose similarity score with the retrieval sub-condition is greater than t2 and whose normalized similarity score with the retrieval sub-condition is greater than t3 are retained as the text sentences related to the retrieval sub-condition.

Specifically, the normalized score _ norm is calculated as follows:

the final strategy for selecting relevant text sentences is that for the text sentence i in the original text of the conference, the similarity score with the retrieval sub-condition needs to be satisfied (score)_i＞t₁)||((score_i＞t₂)&&(score_norm_i＞t₃))。

Through the above-described processing, text sentences related to the search condition or the search sub-condition can be located from the conference original text. The text half sentence related to the search condition or the search sub-condition can also be output to the user, so that the user can apply or know which text contents related to the search condition or the search sub-condition exist in the conference original text.

Based on the processing, the semantic similarity of the retrieval sub-condition and each text sentence in the conference text can be respectively determined, and the text sentence related to the retrieval sub-condition can be identified from the text of the conference original text.

When there are a plurality of retrieval sub-conditions, it is necessary to integrate each retrieval sub-condition with the semantic similarity score of each text sentence in the conference text, and determine the semantic similarity score of each text sentence in the conference text as a whole with the retrieval conditions.

Specifically, when the search condition includes a plurality of search sub-conditions (i.e., the number of texts equivalent to the reference text is greater than 1, where the number of texts may be determined according to the number of texts in the granularity of text sentences, text paragraphs, and the like), the embodiment of the present application determines, for each text sentence in the conference text, its relevance to the search condition, that is, its similarity score to the search condition, by:

firstly, according to the characteristics of the text sentence and the characteristics of each retrieval sub-condition, the relevance of the text sentence and each retrieval sub-condition is determined.

Specifically, referring to the description of the above-mentioned embodiment, for the text sentence, the similarity score for determining each search sub-condition of the text sentence and the search condition may be calculated based on the feature of the text sentence and the feature of each search sub-condition.

And then, according to the relation between each retrieval sub-condition, carrying out fusion processing on the correlation degree of the text sentence and each retrieval sub-condition, and determining the correlation degree of the text sentence and the retrieval condition.

Specifically, when a user sets a plurality of search sub-conditions to constitute a complete search condition, the search condition is generally obtained by logically combining the plurality of search sub-conditions. Therefore, each search sub-condition in the search condition has an explicit logical relationship.

Based on the above logical relationship, in the embodiments of the present application, after the similarity scores between the text sentence and each of the retrieval sub-conditions are determined, the similarity scores between the text sentence and each of the retrieval sub-conditions are logically combined according to the logical relationship between the retrieval sub-conditions, so as to determine the similarity score between the text sentence and the retrieval condition as a whole.

The above-mentioned logical combination of the similarity scores between the text sentence and each search sub-condition can be seen from table 1:

TABLE 1

According to table 1, for the first text sentence in the original text of the conference, the similarity score with the search sub-condition a is assumed to be

Its similarity score with the search sub-condition B is

Then, the text sentence is associated with search condition A&B a similarity score of

The similarity score between the text sentence and the retrieval condition A | | B is

The similarity score between the text sentence and the search condition A-B is

Furthermore, the embodiment of the present application further performs normalization processing on the similarity score between each text sentence in the original meeting text determined in the above manner and the search condition, so that the similarity score between each text sentence in the original meeting text and the search condition is processed between 0 and 1, which is convenient for more intuitively representing the correlation between the text sentence in the original meeting text and the search condition, and makes the correlation between different text sentences and the search condition comparable. After the above-mentioned treatment, the above-mentioned material is passed through the treatment,for 1 st to n text sentences in the conference original text, the relevance of the text sentences to the retrieval conditions is p₁,p₂…,p_nAnd (4) showing.

A2, performing summary generation processing on the full-text content of the target text at least based on the relevance between each text segment in the target text related to the reference text and the reference text, and obtaining a target text summary corresponding to the reference text.

Specifically, when each text segment related to the reference text is located from the target text, the full-text content of the target text can be generated and processed according to each text segment related to the reference text, so as to obtain the target text summary corresponding to the reference text.

For example, a text summary is generated by taking the content of each text segment related to the reference text in the target text as the main content and taking other text contents in the target text as auxiliary contents, and the obtained main content in the target text summary is the content related to the reference text.

Or setting the contribution degree of each relevant text segment to the generation of the target text summary according to the relevance degree of each text segment relevant to the reference text in the target text and the reference text, so that the higher the relevance degree of the text segment relevant to the reference text is, the higher the contribution degree of the text segment relevant to the generation of the target text summary is, and further the proportion of the text segment content relevant to the reference text in the finally generated target text summary is in direct proportion to the relevance degree of the reference text.

As a preferred implementation manner, in the embodiment of the present application, relevance between each text segment in the target text and the reference text is comprehensively considered, and the full-text content of the target text is subjected to summary generation processing. That is, a summary of the full-text content of the target text is generated by performing the following steps A21-A22:

a21, determining the contribution degree of each text segment in the target text to the generation of the text summary corresponding to the reference text according to the relevance degree of each text segment in the target text and the reference text.

Specifically, according to the rule that the higher the correlation degree between the text segment in the target text and the reference text is, the greater the contribution degree of the text segment in the target text to the generation of the text summary corresponding to the reference text is, the contribution degree of each text segment in the target text to the generation of the text summary corresponding to the reference text is respectively determined.

As an exemplary implementation, the embodiment of the application adopts a text summary generation model based on an attention mechanism to generate a text summary of a target text. The text summary generation model based on the attention mechanism can determine the contribution degree of each text segment in the target text to the generation of the text summary corresponding to the reference text according to the correlation degree of each text segment in the target text and the reference text, and further generate the text summary of the target text based on the contribution degree.

Because the text summary generation model is a text decoding model based on an attention mechanism, the model can obtain a text summary decoding result meeting the requirement by adjusting attention coefficients of all text segments of the input target text. When the attention of the decoding process is different for different text segments, the content of the text summary finally decoded can be changed. Since the text summary generation model based on the attention mechanism is a time sequence output model, the text summary decoding at the current time point has the attention coefficient for each text segment in the target text and possibly also relates to the target text summary content which is output before the current time point.

Through training, the model can determine the attention coefficient distribution of the text summary decoding to each text segment of the target file at the current moment according to the decoding result of the preamble, namely, the correct attention coefficient is distributed to each text segment of the input target text, so that the attention coefficient of the model to each text segment in the target text can be determined when the model generates the text summary of the target text.

And then determining the contribution degree of each text segment in the target text to the generation of the text summary corresponding to the reference text according to the attention coefficient of the text summary of the generated target text to each text segment in the target text and the correlation degree of each text segment in the target text and the reference text.

Because the final purpose of the embodiment of the application is to generate the text summary corresponding to the reference text, it is not enough to make the finally generated target text summary correspond to the reference text only by determining the attention coefficient of the text summary generated for each text segment of the target text. In order to enable the finally generated target text summary to correspond to the reference text, the embodiment of the application further combines the attention coefficient of the text summary of the generated target text for each text segment in the target text with the correlation degree of each text segment in the target text and the reference text, and jointly determines the contribution degree of each text segment in the target text to the generation of the text summary corresponding to the reference text.

Illustratively, the attention coefficient of the text summary of the generated target text to each text segment in the target text is multiplied by the correlation between each text segment in the target text and the reference text, then the product result corresponding to each text segment is normalized, and the finally obtained normalized value corresponding to each text segment is used as the contribution degree of each text segment in the target text to the generation of the text summary corresponding to the reference text.

The above solution for determining the contribution degree can also refer to the following examples of the specific process for generating the target text summary.

A22, according to at least the contribution degree of each text segment in the target text to the generation of the text summary corresponding to the reference text, carrying out summary generation processing on the full text content of the target text to obtain the target text summary corresponding to the reference text.

Specifically, the text summary generation model based on the attention mechanism generates text summary decoding features according to the features of the target text and the contribution degree of each text segment in the target text to the generation of the text summary corresponding to the reference text; then, according to the text summary decoding feature, text summary decoding processing is carried out to generate a text summary of the target text.

According to the method and the device for determining the characteristics of the target text, the characteristics of each text segment of the target text are determined by acquiring the characteristics of each text segment of the target text, so that the characteristics of each text segment of the target text are definite in advance. And the contribution degree of each text segment of the target text to the generation of the text summary is different, so that the text summary decoding feature can be generated directly according to the feature of each text segment of the target text and the contribution degree of each text segment in the target text to the generation of the text summary corresponding to the reference text.

And further decoding the text summary decoding characteristics in a preset dictionary range to obtain a decoding result. And decoding the full target text by the model according to the method to obtain the target text summary corresponding to the reference text.

Still taking the example of generating a conference summary meeting the user retrieval condition for the conference original text, fig. 4 shows the structure of a text summary generation model based on the attention mechanism and the processing procedure of decoding the text summary generation model to generate a target text summary.

Suppose that for a certain search condition of a user, the sequence of conference summary text words solved by the model history is represented as y₁,y₂,…,y_t-1The sentence hiding layer of each sentence (assuming n sentences) in the conference original text is characterized in that

t represents the current decoding time, and the relevance of the 1 st to the n th text sentences in the conference original text to the retrieval condition is p₁,p₂…,p_nAnd (4) showing. The sentence hiding layer characteristic of each sentence in the conference original text

And the relevance p of each sentence text to the search condition₁,p₂…,p_nInputting the model, in particular inputting the solution of the modelAnd the code end and the original text interaction attention module.

Historically solved text word sequence y of meetings₁,y₂,…,y_t-1After passing through a decoding hidden layer feature expression module of the model, obtaining a hidden layer state feature d at the current decoding moment_t. The decoding hidden layer feature expression module can input a certain retrieval condition for a user, and output hidden layer state features of the current decoding moment by utilizing the historical conference summary text word sequence solved by the model. The network structure of the decoding hidden layer feature expression module can utilize a decoder partial coding model under a Transformer scheme or a unidirectional LSTM structure and the like.

In the module for interactive attention between the decoding end and the original text, firstly, based on the attention mechanism, the hidden layer state feature d at the current decoding time is determined_tAttention coefficient to sentence hiding layer characteristics of jth text sentence in conference original text

Then, according to the hidden layer state characteristic d_tAttention coefficient to sentence hiding layer characteristics of jth text sentence in conference original text

And the correlation p between the jth text sentence in the conference original text and the retrieval condition_jCalculating and determining the contribution degree of the jth text sentence in the conference original text to the generation of the conference summary

Finally, generating text summary decoding characteristics according to sentence hidden layer characteristics of each text sentence in the conference original text and contribution degrees of each text sentence in the conference original text to generation of a conference summary

The specific calculation procedure of the above processing is as follows:

where j is 1, 2, … n, which represents n text sentences in the original text of the conference. The Attention () represents the Attention mechanism calculation function, and can adopt self-Attention, additive Attention and the like. In the embodiment of the application, the decoding end fully considers the influence of information related to the content of the retrieval condition on the generation of the conference summary in all text sentences of the conference original text, and optimizes and improves the original attention coefficient to be

Text summary decoding feature

And selecting related contents for a decoding end with semantic fuzzy retrieval characteristics, and paying attention to the context vector representation with different degrees of the hidden layer characteristics of the original sentence of the conference.

From

And

the calculation process shows that if the correlation degree of a certain sentence in the original text of the conference and the retrieval condition is higher, namely the retrieval matching characteristic value is larger, the attention coefficient value of the sentence after corresponding optimization is larger, and the final text summary decoding characteristic is carried out

The larger the contribution degree of the original text content is, the capability of selecting and retrieving related original text content of the conference summary generation model based on the attention mechanism provided by the embodiment of the application is ensured.

The out-word prediction module of FIG. 4 is an input text summary decode feature

And calculating the word output probability distributed in the size of the dictionary, and outputting the word corresponding to the current decoding moment. The network structure of the word-out prediction module can utilize a linear layer to carry out nonlinear activation function layer, and the decoding algorithm can utilize a beamsearch algorithm. According to the algorithm, the model respectively determines the text summary decoding characteristics based on the conference text at each moment and performs decoding output, so that the conference summary which corresponds to the conference text and meets the retrieval conditions can be obtained.

As a preferred implementation manner, since the target text summary finally generated in the embodiment of the present application needs to correspond to the reference text, in addition to defining the contribution degree of each text segment in the target text to the generation of the text summary corresponding to the reference text, so that the finally generated target text summary includes target text content related to the reference text, the embodiment of the present application directly uses the feature of the reference text for generating the text summary of the target text, thereby further improving the relevance between the generated target text summary and the reference text.

Based on the above thought, in the embodiment of the application, when the text summary decoding feature is generated, the text summary decoding feature is generated according to the feature of the target text, the feature of the reference text, and the contribution degree of each text segment in the target text to the generation of the text summary corresponding to the reference text.

Illustratively, in the attention-based text summary generation model shown in fig. 4, except for the question-hidden layer feature of each sentence of text in the conference original text

And the relevance p of each sentence text to the search condition₁,p₂…,p_nInputting the characteristics of the search condition into the model in addition to the model, for example, the word hidden layer characteristics of each word of the search condition

The model is input, so that when the model is subjected to decoding processing, the text content of the original text of the conference can be decoded by referring to the search condition, and the correlation between the decoding result and the search condition is improved.

For example, the sentence hiding layer characteristic of each sentence text in the meeting original text

Relevance p of each sentence text and retrieval condition₁,p₂…,p_nAnd word hidden layer characteristics of each word of the search condition

And inputting a decoding end of the model and an original text interaction attention module, so that the model can use the retrieval condition as a reference when determining the attention coefficient of each text sentence of the conference original text, determining the contribution degree of each text sentence of the conference original text to the generation of the conference summary corresponding to the retrieval condition and generating the text summary decoding characteristic, thereby enabling the correlation degree of the finally decoded conference summary and the retrieval condition to be higher and avoiding the conference summary from deviating from the retrieval condition.

As an optional implementation manner, an embodiment of the present application proposes an attention-based text summary generation model structure as shown in fig. 5, where, with respect to the model structure shown in fig. 4, a decoding end and a retrieval interaction attention module are added before a decoding end and an original text interaction attention module, and the decoding end and the retrieval interaction attention module mainly implement interaction between a model hidden layer state feature and a retrieval condition feature to generate a reference decoding feature, and then implement interaction with a conference original text feature through the decoding end and the original text interaction attention module.

Based on the model structure shown in fig. 5, when generating text summary decoding features, generating reference decoding features according to the features of a reference text; and then generating text summary decoding characteristics according to the reference decoding characteristics, the characteristics of the target text and the contribution degree of each text segment in the target text to the generation of the text summary corresponding to the reference text.

Specifically, referring to fig. 5, the features of the search condition and the hidden state features are input into the decoding end of the model and the search interaction attention module, so that the features of the search condition and the hidden state features of the model are fused to obtain the reference decoding features.

Furthermore, in the embodiment of the present application, the features of the reference text are determined by referring to the features of each text segment of the text, and the influence of each text segment of the reference text on generating the target text summary corresponding to the reference text is also different. For example, key entity words in the reference text can largely express the semantics of the reference text, so that the reference value of the key entity words for generating the target text summary corresponding to the reference text is greater, and the reference value of non-entity words in the reference text, such as the mood words, the modifiers and the like, for generating the target text summary corresponding to the reference text is relatively smaller. Therefore, in the embodiment of the present application, with reference to the scheme described in the above embodiment for determining the contribution degree of each text segment in the target text to the generation of the text summary corresponding to the reference text, the contribution degree of each text segment in the reference text to the generation of the text summary corresponding to the reference text is determined. The text segments in the reference text can be text contents with any granularity, such as words, phrases, sentences, text paragraphs and the like.

And after the contribution degree of each text segment in the reference text to the generation of the text summary corresponding to the reference text is clarified, generating a reference decoding feature according to the feature of the reference text and the contribution degree of each text segment in the reference text to the generation of the text summary corresponding to the reference text.

Specifically, the embodiment of the present application determines the features of the reference text by obtaining the features of each text segment of the reference text, and the contribution degrees of each text segment of the reference text to the generation of the text summary corresponding to the reference text are different, so the embodiment of the present application generates the reference decoding features according to the features of each text segment of the reference text and the contribution degrees of each text segment in the reference text to the generation of the text summary corresponding to the reference text.

Continuing with the above-mentioned example of generating a meeting summary meeting the user search condition for the meeting original text, by means of the text summary generation model based on the attention mechanism as shown in fig. 5, it is assumed that the word hidden layer feature of each word of the search condition is

Hiding the current decoding time with the state feature d_tAnd word hidden layer characteristics of each word of search condition

The module firstly determines the contribution degree of each word in the retrieval condition to the generation of the text summary corresponding to the retrieval condition by using an attention mechanism, and then generates reference decoding characteristics according to the characteristics of each word in the retrieval condition and the contribution degree of each word in the retrieval condition to the generation of the text summary corresponding to the retrieval condition. The specific calculation process is as follows:

where i is 1, 2, … m1+ m2, which indicates m1+ m2 words in the search condition, and m1+ m2 indicates the sum of the numbers of words of two search sub-conditions included in the search condition. The Attention () represents the Attention mechanism calculation function, and can adopt self-Attention, additive Attention and the like.

Representing the current decoding time, decoding hidden layer state feature attention coefficient to the word hidden layer feature of the ith word in the search condition, and the coefficient also represents the search barThe contribution of the ith word in the piece to the generation of the text summary corresponding to the search condition.

In the embodiment of the present application, a decoding end is named as a reference decoding feature for representing context vectors with different degrees of hidden features of a search condition word of interest obtained after interacting with a search. It can be understood that the technical scheme of the embodiment of the application fully considers the contribution degree of the information in the retrieval condition to the generation of the conference summary at the decoding end.

Referring to FIG. 5, in obtaining the reference decoding feature

And then inputting the reference decoding characteristics, the characteristics of each text sentence of the original conference text and the correlation degree of each text sentence in the original conference text and the retrieval condition into a decoding end and original text interaction attention module, and determining the contribution degree of each text sentence in the original conference text to the generation of the text summary corresponding to the retrieval condition and the text summary decoding characteristics by the model through attention interaction operation. The specific calculation process is as follows:

wherein,

the context vector after the decoding end interacts with the search at the time t, that is, the reference decoding feature.

The syntax hidden layer characteristic in the jth syntax in the conference original text is i ═ 1, 2, … n. The Attention () represents the Attention mechanism calculation function, and can adopt self-Attention, additive Attention and the like.

And the attention coefficient of the decoding end and the context vector after retrieval interaction to the feature attention coefficient of the jth text sentence in the original text of the conference at the current decoding moment is shown, namely the attention coefficient of the decoding end to the jth text sentence in the original text of the conference when generating the text summary corresponding to the retrieval condition. p is a radical of_jAnd the similarity between the retrieval conditions and the text of the jth sentence in the conference original text is shown.

By combining the introduction of the above embodiments, it can be seen that, in the embodiment of the present application, when a text summary is decoded, not only the contribution degree of each text segment of a target text to the generation of the target text summary corresponding to a reference text is considered, but also the contribution degree of each text segment of the reference text to the generation of the target text summary corresponding to the reference text is considered. Therefore, the whole target text summary generation process has the capability of selecting and retrieving related target text content and reference text content, the relevance of the finally generated target text summary and the reference text and the related content of the reference text in the target text is improved, and the finally generated target text summary corresponds to the reference text.

It should be noted that, in the above embodiment, the obtaining process of the reference decoding feature and the text summary decoding feature and the processing process of finally decoding to obtain the target text summary are described with the aid of the features of each text segment of the target text and the features of each text segment of the reference text. In practical implementation of the embodiment of the present application, the global feature of the target text and the global feature of the reference text may be directly input into the text summary generation model based on the attention mechanism, and the reference decoding feature and the text summary decoding feature are obtained, and at this time, both the training process of the model and the specific processing process of the model may be performed with reference to the description of the above embodiment.

In addition, the names of the processing modules of the text summary generation model based on the attention mechanism shown in fig. 4 and 5 are named in combination with specific processing objects, and when the target text and the reference text to be actually processed are texts of other types, not the conference original text and the search condition, the names of the processing modules can be adaptively changed according to the actual processing objects. The embodiment of the present application does not limit the names of the processing modules of the text summary generation model based on the attention mechanism, but mainly introduces the functions and processing contents of the processing modules, thereby specifically introducing the processing procedures and implemented functions of the text summary generation model based on the attention mechanism.

In the above embodiments, specific embodiments of measuring the relevance between each text segment of the target text and the reference text, and performing text summary generation processing on the target text to obtain a target text summary corresponding to the reference text are respectively described. As for the computer device, it processes the text, which is essentially to process the features of the text, that is, the content of the text processing included in the text summary generation method provided in the embodiment of the present application is essentially to process the features of the text. Therefore, the accuracy of the text feature will directly affect the accuracy of the text summary generation process. Next, in the embodiments of the present application, an example of a manner of acquiring features of a target text and features of a reference text will be described.

In general, an encoder is used to encode a text to obtain text features. For example, a word-level coding encoder structure is adopted, so that the coding characteristics of the target text and the reference text can be acquired. However, the target text of the summary, which is usually a long text, needs to be generated, for example, the main feature of the conference text is that the text length is long, and a conference with one hour of duration may contain 1-2 ten thousand words. However, if the features of the conference text are obtained by using a conventional word-level encoder, a large amount of content is consumed, long-distance dependent information cannot be well captured, and the extracted text features are inaccurate or incomplete, which also results in a conventional generation method of the document, which can only perform abstract generation for scenes such as news, mails and few turns of interpersonal conversations, which cannot be competent for generating tasks of the document of the long target text.

In order to improve the effect of text feature extraction, as an exemplary implementation manner, in the embodiments of the present application, a feature of a target text is determined by acquiring features of each text segment in the target text, and a feature of a reference text is determined by acquiring features of each text segment in the reference text.

When the features of each text segment in the target text are obtained, the following steps B1-B4 are executed to realize feature extraction of each text segment:

and B1, performing text segment division processing on the target text, and determining each text segment contained in the target text.

And B2, performing word segmentation processing on each text segment in the target text respectively, and determining each word segmentation contained in each text segment.

And B3, extracting the participle characteristics of the fusion context information of the participles contained in each text segment.

And B4, determining the characteristics of each text segment according to the segmentation characteristics of the fusion context information of each segmentation contained in each text segment.

Specifically, in the embodiment of the present application, a word-sentence-chapter hierarchical information coding model is constructed to extract sentence-level hidden layer features and chapter-level hidden layer features of the fusion context information of the target text, that is, to extract features of each text segment of the target text and overall features of the target text.

The method and the device for segmenting the target text comprise the steps of firstly segmenting the target text, segmenting words of the segmented text and determining each segmented word contained in each text segment. The text segment division may be performed on the target text, for example, the text sentence division is performed according to punctuation marks, or the text segment is extracted by sliding on the target text through a sliding window with a fixed number of words. The word segmentation of the text segment can be realized by adopting the existing word segmentation algorithm, and the embodiment of the application is not described in detail. Based on the text segment division and the word segmentation processing, the text segment characteristics and the chapter characteristics of the target text can be extracted by means of the word-sentence-chapter hierarchical information coding model aiming at the word segmentation contained in each text segment of the target text.

Referring to fig. 6, the word-sentence-chapter hierarchical information coding model includes a word hidden layer feature expression module, a sentence expression extraction module, a sentence hidden layer feature expression module, and a chapter feature extraction module. Next, a processing procedure of extracting features of each text segment of the target text and overall features of the target text is introduced by taking the conference original text as a target text, each text sentence in the conference original text as each text segment in the target text, and extracting features of each text sentence in the conference original text and features of chapters of the conference original text as examples.

The term hidden layer feature expression module is used for inputting the term representation of each term for each text sentence in the original text of the conference and outputting the term hidden layer feature fused with the context information of the current sentence. The network structure of the word hidden layer feature expression module can utilize an encoder partial model under a Transformer scheme or a bidirectional LSTM structure and the like. Suppose that n sentences are provided in total after the original text of the conference is divided into sentences, and the word sequence contained in each sentence is W_n1,W_n2,…,W_nmnWherein n represents the nth sentence of the original text of the conference, m_nIndicating the total number of words contained in the nth sentence.

The term hidden layer characteristics of each term in the 1 st sentence in the text of the conference after the context information of the current sentence is fused are shown, and m1 shows that m1 terms are in total in the 1 st sentence. In the same way

The word hidden layer characteristics of m2 words in the 2 nd sentence in the text of the conference original text after the context information of the current sentence is fused are shown,

representing m in the nth sentence in the text of the original text of the conference_nAnd the word hidden layer characteristics of the words after the context information of the current sentence is fused.

The sentence expression extraction module compresses word expressions of a plurality of words in the input sequence to obtain a sentence expression vector and all word hidden layer characteristics of the 1 st sentence of the original text of the conference

After the sentence expression and extraction module, the sentence expression vector of the 1 st sentence is obtained as s₁. By analogy, sentence representation vectors of the 1 st sentence to the nth sentence in the conference text can be represented as a sequence s₁,s₂…,s_n. The network structure of the sentence expression extraction module is not limited in the embodiment of the application, and techniques such as attention mechanism or pooling can be adopted.

The sentence hiding layer feature expression module is used for inputting all sentence expression vectors of the original text of the conference and outputting the sentence hiding layer features fused with the context information of the current sentence. Similar to the above-mentioned word hidden layer feature expression module, the network structure of the sentence hidden layer feature expression module can use an encoder partial model or a bidirectional LSTM and other structures under a Transformer scheme.

And the sentence hiding layer characteristics of n sentences in the conference text after the context information is fused are represented.

The above-mentioned chapter feature extraction module, similar to the above-mentioned sentence expression extraction module, compresses the sentence-hidden layer feature expression of multiple sentences in the input sequence to obtain a chapter expression vector. Sentence representation vectors of sentences 1 to n of the original text of the conference can be represented as a sequence

And obtaining the discourse character u of the original text of the conference after the discourse character extraction module. The embodiment of the present application does not limit the network structure of the chapter feature extraction module, and may adopt an attention mechanism orPooling and the like.

Through the word-sentence-chapter hierarchical information coding model, the characteristics of each text sentence of the conference original text and the chapter characteristics of the conference original text, namely the overall characteristics of the conference original text, can be respectively obtained. Moreover, the characteristics of the text sentence of the conference original text, the characteristics of the participles contained in the text sentence, and the chapter characteristics of the conference original text are the characteristics of the integrated context information, so that the long-distance dependency information in the long-distance conference text can be captured better by the conference original text characteristic extraction scheme of the embodiment of the application, and more accurate conference text characteristics can be obtained.

As an optional implementation, the word-sentence-chapter hierarchical information coding model may also omit the sentence hidden layer feature expression module therein, and directly use the sentence expression vector s of each text sentence output by the sentence expression extraction module₁,s₂…,s_nThe characteristics of the individual text sentences as conference text, and s₁,s₂…,s_nAnd determining the chapter characteristics u of the conference text.

When obtaining the features of each text segment in the reference text, the feature extraction of each text segment is realized by executing the following steps C1-C2:

and C1, performing word segmentation processing on the reference text, and determining each word segmentation contained in the reference text.

And C2, extracting the participle characteristics of the fused context information of each participle contained in the reference text respectively.

Specifically, the embodiment of the application determines the overall characteristics of the reference text by extracting the characteristics of each text segment of the reference text. The text segment of the reference text may be text content with any granularity, such as a word, a phrase, a text sentence, a text segment, and the like in the reference text. In the embodiment of the application, the features of the reference text are determined by extracting the features of the word segments of the reference text.

Therefore, the reference text is first subjected to word segmentation processing, for example, by a word segmentation model or a word segmentation algorithm, so as to determine each word segment contained in the reference text. Then, the word segmentation characteristics of the fusion context information of each word segmentation contained in the reference text are respectively extracted, and the overall characteristics of the reference text are obtained through combination according to the word segmentation characteristics of each word segmentation contained in the reference text.

Further, if the number of texts included in the reference text is greater than 1, that is, the reference text includes a plurality of text sentences, in the embodiment of the present application, the text sentences included in the reference text are integrated into one reference text after being merged or screened according to the relationship between the reference texts, and then the reference text is subjected to word segmentation, word segmentation feature extraction, and reference text feature extraction.

Illustratively, in the embodiment of the present application, the word segmentation characteristics of each word segmentation of the reference text are extracted by a word hidden layer characteristic expression module as shown in fig. 7.

Take the feature of extracting the search condition that the user inputs in the conference summary of obtaining the original text of the conference as an example. If the user retrieval condition is a single condition, after segmenting the user retrieval condition, inputting the retrieval word sequence into the word hidden layer feature expression module, and outputting the word hidden layer features of the context information of the fusion retrieval condition, so as to obtain the segmentation features of the fusion context information of each segmentation contained in the retrieval condition.

If the user search condition is a combination of a plurality of search sub-conditions, the word sequence of the single or a plurality of search sub-conditions is input into the word hidden layer feature expression module according to the method shown in the following table 2, and the word hidden layer feature expression module correspondingly outputs the word hidden layer features of the fusion context information of the single or a plurality of search conditions.

TABLE 2

Composite conditions	Input deviceDescription of the invention	Number of input sequence
			A&B	Splicing A and B sub-conditional sequences into a sequence input	1
A\|\|B	Inputting A and B sub-condition sequences respectively	2
			A-B	Inputting only A sub-conditional sequences	1

According to the concept of merging or screening search conditions shown in table 2, for a more complex search sub-condition complex case, such as { a & B } -C, a text sequence can be processed by the above method, i.e. a text sequence input after splicing the sub-conditions a and B.

The network structure of the word hidden layer feature expression module can utilize an encoder partial model under a Transformer scheme or a bidirectional LSTM structure and the like. As shown in FIG. 7, assume that an input sequence in a search condition has m1 words

After the word sequence is input into the word hidden layer feature expression module, the hidden layer feature of each word of the search condition is obtained and expressed as

In particular, if the search condition has multiple sequences after being decomposed according to the method, the extraction process is similar, for example, another input sequence in the figure has m2 words, which are respectively

After the word sequence is input into the word hidden layer feature expression module, the hidden layer feature of each word is obtained and expressed as

According to the method, the word segmentation characteristics of each retrieval sub-condition can be respectively obtained, and finally, the word segmentation characteristics of each word segmentation included in the retrieval sub-conditions are spliced according to the word segmentation sequence, so that the overall characteristics of the retrieval conditions can be obtained.

After the features of each text segment of the target text and the features of each text segment of the reference text are respectively determined through the processing, the feature fusion processing is further performed on the features of each text segment of the target text and the features of each text segment of the reference text to obtain the target text features fused with the reference text features and/or the reference text features fused with the target text features.

That is, in the embodiment of the present application, the features of the reference text are merged into the features of the target text, and/or the features of the target text are merged into the features of the reference text, so that the features of the reference text and/or the target text not only include the features of the reference text and/or the features of the target text, but also include the features of the other party.

In the embodiment of the application, the features of the reference text are merged into the features of the target text, and meanwhile, the features of the target text are merged into the features of the reference text. When the technical solution of the embodiment of the present application is actually implemented, the features of one of the two may be selectively incorporated into the other according to the introduction of the embodiment of the present application.

When the target text features are merged into the reference text features, chapter features of the target text and/or features of each text segment of the target text, features of each text segment of the reference text, or overall features of the reference text may be merged. The chapter characteristics of the target text are determined according to the characteristics of each text segment of the target text.

In the embodiment of the application, chapter characteristics of the target text are determined according to characteristics of each text segment of the target text. Then, integrating the chapter characteristics of the target text and the characteristics of each text segment of the target text into the characteristics of each text segment of the reference text respectively; in addition, the characteristics of each text segment of the reference text are blended into the characteristics of each text segment of the target text. Finally, the obtained characteristics of each text segment of the reference text are fused with the chapter characteristics and each text segment characteristics of the target text, and the obtained characteristics of each text segment of the target text are fused with the characteristics of each text segment of the reference text.

In an exemplary embodiment, the above-mentioned original meeting text is used to represent the target text, and the user search condition is used to represent the reference text, in the embodiment of the present application, information fusion between the user search condition and the original meeting text is fully considered, that is, when each hidden layer feature of each sentence in the final original meeting text is extracted, relevant search condition information is fused, and meanwhile, when each hidden layer feature of each word in the final user search condition is extracted, the original meeting text information is also fused.

The embodiment of the application builds an information fusion model for realizing information fusion of the user retrieval condition and the conference original text. Referring to fig. 8, the information fusion model includes a word hidden layer feature expression module, a word feature extraction module, and an information mutual fusion module.

The functions and processing procedures of the hidden layer feature expression module can be described with reference to fig. 7.

The word feature extraction module inputs the original text chapter-level hidden layer features u of the conference and the hidden layer feature representation of each word of the retrieval conditions

Hidden layer characteristics of each word of retrieval conditions after outputting and fusing original text chapter information

In the embodiment of the application, the word feature extraction module adopts a recursive network structure, and the discourse text-level hidden layer features u of the conference are an initial state tableThe calculation process of the hidden layer characteristics of each word in the search condition after the information of the original text chapter of the fused conference is obtained by recursion is shown as follows:

…

the recursive network structure may be LSTM or GRU.

In particular, if the search condition is a composite of a plurality of search sub-conditions, there are a plurality of word hidden layer features after the above processing. As shown in FIG. 8, for another word hidden layer feature representation

The calculation process of the hidden layer characteristics of each word in the retrieval conditions after extracting the information of the original text chapters of the merged conference is consistent with the above process, and finally the hidden layer characteristics of each word are obtained

The information mutual fusion module inputs the hidden layer characteristics of each word in the retrieval conditions after the information of the original text chapters of the conference is fused and the sentence hidden layer characteristics of each text sentence of the original text of the conference

Outputting hidden layer characteristics of each word of search conditions further fused with original sentence information

And sentence hiding layer characteristics of each sentence in conference original text fused with related search condition word information

The information mutual fusion module can utilize self-attention mechanism or bidirectional LSTM and other structures.

As can be seen from the above description, when the feature extraction is performed on the target text and the reference text, the feature extraction method and the device can not only extract the features of each text segment of the target text and the reference text, which are fused with context information, but also realize the fusion of the features of the target text and the reference text, so that the feature information of the target text and the reference text is richer, and the text summary generation processing is more favorably performed on the target text to obtain the target text summary corresponding to the reference text.

For example, based on the features of the target text extracted in the above manner, text summary generation processing is performed, or the features of the target text extracted in the above manner and the features of the reference text are combined and applied to perform text summary generation processing.

In the embodiment of the present application, the features of each text segment of the target text extracted in the above manner and the features of each text segment of the reference text are input into the text summary generation model based on the attention mechanism as shown in fig. 5, so as to generate the target text summary corresponding to the reference text, and refer to the description of the above embodiment for a specific summary generation process.

Corresponding to the text summary generation method, an embodiment of the present application further provides a text summary generation apparatus, as shown in fig. 9, the apparatus includes:

a data acquisition unit 100 configured to acquire a target text and a reference text, wherein the reference text is determined based on target text content focused by a user;

an epoch generating unit 110, configured to perform epoch generating processing on the target text based on the associated content of the reference text located in the target text, so as to obtain a target text epoch corresponding to the reference text.

As an optional implementation manner, performing summary generation processing on the target text based on the associated content of the reference text located from the target text to obtain a target text summary corresponding to the reference text, includes:

and based on the text segment related to the reference text positioned in the target text, carrying out summary generation processing on the full-text content of the target text to obtain a target text summary corresponding to the reference text.

As an optional implementation manner, based on locating a text segment related to the reference text from the target text, performing summary generation processing on full-text content of the target text to obtain a target text summary corresponding to the reference text, including:

determining the correlation degree of each text segment in a target text and a reference text, and locating the text segment related to the reference text from the target text;

and performing summary generation processing on the full-text content of the target text at least based on the correlation degree between each text segment in the target text, which is related to the reference text, and the reference text to obtain a target text summary corresponding to the reference text.

As an optional implementation manner, performing summary generation processing on full-text content of the target text based on at least the correlation degree between each text segment in the target text, which is related to the reference text, and the reference text to obtain a target text summary corresponding to the reference text, includes:

determining the contribution degree of each text segment in the target text to the generation of a text summary corresponding to the reference text according to the correlation degree of each text segment in the target text and the reference text;

and carrying out summary generation processing on the full-text content of the target text at least according to the contribution degree of each text segment in the target text to the generation of the text summary corresponding to the reference text, so as to obtain the target text summary corresponding to the reference text.

As an optional implementation manner, determining the relevance of each text segment in the target text to the reference text includes:

respectively acquiring the characteristics of a target text and the characteristics of a reference text;

and respectively determining the correlation degree of each text segment in the target text and the reference text according to the characteristics of each text segment in the target text and the characteristics of the reference text.

As an optional implementation, the summary generation unit 110 is further configured to:

and performing interactive operation based on an attention mechanism on the characteristics of each text segment in the target text and the characteristics of the reference text to obtain the characteristics of the reference text with the information being complete.

As an optional implementation manner, performing an interactive operation based on an attention mechanism on the features of each text segment in the target text and the features of the reference text to obtain reference text features with complete information, including:

calculating and determining the similarity between each text segment in the target text and the reference text according to the characteristics of each text segment in the target text and the characteristics of the reference text;

selecting a first number of text segments with highest similarity with the reference text from the target text according to the similarity between each text segment in the target text and the reference text;

and performing interactive operation based on an attention mechanism on the features of the first number of text segments selected from the target text and the features of the reference text to obtain the reference text features with the information being complete.

As an optional implementation manner, when the number of texts in the reference text is greater than 1, determining, according to the features of each text segment in the target text and the features of the reference text, the relevance between each text segment in the target text and the reference text respectively includes:

for each text segment in the target text, the relevance of the text segment to the reference text is determined by the following processes:

determining the relevance of the text segment and each reference text according to the characteristics of the text segment and the characteristics of each reference text;

and according to the relation between the reference texts, carrying out fusion processing on the correlation degrees of the text segment and the reference texts, and determining the correlation degrees of the text segment and the reference texts.

calculating and determining semantic similarity between each text segment in the target text and the reference text through a BM25 algorithm according to the characteristics of each text segment in the target text and the characteristics of the reference text;

and carrying out fusion processing on the correlation degree of each text segment in the target text and the reference text and the semantic similarity of each text segment in the target text and the reference text to obtain the correlation degree of each text segment in the fused target text and the reference text.

and correcting the correlation degree of each text segment in the target text and the reference text according to the position distribution of each text segment in the target text.

As an optional implementation manner, modifying the correlation between each text segment in the target text and the reference text according to the position distribution of each text segment in the target text, includes:

selecting a second number of text segments with highest correlation degree with the reference text from all the text segments in the target text;

determining the punishment degree of the correlation degree of the other text segments in the target text and the reference text according to a rule that the larger the distance between the other text segments in the target text and the selected second number of text segments is, the higher the punishment degree of the correlation degree of the other text segments and the reference text is;

and punishing the relevance between the other text segments in the target text and the reference text according to the punishment of the relevance between the other text segments in the target text and the reference text.

selecting a third number of text segments with highest correlation degree with the reference text from the target text according to the correlation degree of each text segment in the target text and the reference text;

selecting a text segment with the correlation degree larger than a first correlation degree threshold value with the reference text or with the correlation degree larger than a second correlation degree threshold value with the reference text and with the normalized correlation degree larger than a third correlation degree threshold value with the reference text from the third number of text segments as a text segment related to the reference text according to the correlation degree between each text segment in the third number of selected text segments and the reference text;

As an optional implementation manner, determining, according to a relevance between each text segment in a target text and a reference text, a contribution degree of each text segment in the target text to generating a text summary corresponding to the reference text includes:

determining attention coefficients of text excerpts generating the target text for each text segment in the target text;

determining the contribution degree of each text segment in the target text to the generation of the text summary corresponding to the reference text according to the attention coefficient of the text summary of the target text to each text segment in the target text and the correlation degree of each text segment in the target text and the reference text.

As an optional implementation manner, performing summary generation processing on full-text content of the target text according to at least a contribution degree of each text segment in the target text to generating a text summary corresponding to the reference text, to obtain a target text summary corresponding to the reference text, includes:

generating text summary decoding characteristics at least according to the characteristics of the target text and the contribution degree of each text segment in the target text to the generation of a text summary corresponding to the reference text;

and generating a text summary of the target text according to the text summary decoding characteristics.

As an optional implementation manner, generating a text summary decoding feature according to at least the feature of the target text and the contribution degree of each text segment in the target text to generating a text summary corresponding to the reference text includes:

and generating text summary decoding characteristics at least according to the characteristics of the target text, the characteristics of the reference text and the contribution degree of each text segment in the target text to the generation of a text summary corresponding to the reference text.

As an optional implementation manner, generating a text summary decoding feature according to at least the feature of the target text, the feature of the reference text, and the degree of contribution of each text segment in the target text to the generation of a text summary corresponding to the reference text, includes:

generating a reference decoding feature according to at least the feature of the reference text;

and generating text summary decoding characteristics according to the reference decoding characteristics, the characteristics of the target text and the contribution degree of each text segment in the target text to the generation of a text summary corresponding to the reference text.

determining the contribution degree of each text segment in the reference text to generating a text summary corresponding to the reference text;

generating a reference decoding feature at least according to the feature of the reference text, including:

and generating reference decoding characteristics according to the characteristics of the reference text and the contribution degree of each text segment in the reference text to the generation of the text summary corresponding to the reference text.

As an optional implementation manner, generating a reference decoding feature according to the feature of the reference text and the contribution degree of each text segment in the reference text to generating a text summary corresponding to the reference text includes:

generating reference decoding characteristics according to the characteristics of each text segment of the reference text and the contribution degree of each text segment in the reference text to the generation of a text summary corresponding to the reference text;

generating text summary decoding characteristics according to the reference decoding characteristics, the characteristics of the target text and the contribution degree of each text segment in the target text to generating a text summary corresponding to the reference text, including:

and generating text summary decoding characteristics according to the reference decoding characteristics, the characteristics of each text segment of the target text and the contribution degree of each text segment in the target text to the generation of a text summary corresponding to the reference text.

As an optional implementation manner, generating a text summary decoding feature according to the feature of the target text and the contribution degree of each text segment in the target text to generating a text summary corresponding to the reference text includes:

and generating text summary decoding characteristics according to the characteristics of each text segment in the target text and the contribution degree of each text segment in the target text to the generation of the text summary corresponding to the reference text.

As an optional implementation manner, obtaining a target text and a reference text, performing summary generation processing on the target text based on the associated content of the reference text located in the target text, and obtaining a target text summary corresponding to the reference text, includes:

acquiring the characteristics of a target text and the characteristics of a reference text;

inputting the characteristics of a target text and the characteristics of a reference text into a pre-trained attention-based text summary generation model, enabling the attention-based text summary generation model to perform summary generation processing on the target text based on the associated content of the reference text positioned in the target text, and obtaining a target text summary corresponding to the reference text.

As an optional implementation manner, the characteristics of the target text are obtained by acquiring the characteristics of each text segment in the target text; and the characteristics of the reference text are obtained by acquiring the characteristics of each text segment of the reference text.

As an optional implementation manner, obtaining characteristics of each text segment in the target text includes:

performing text segment division processing on the target text, and determining each text segment contained in the target text;

performing word segmentation processing on each text segment in the target text respectively, and determining each word segmentation contained in each text segment;

respectively extracting the word segmentation characteristics of the fusion context information of each word segmentation contained in each text segment;

and determining the characteristics of each text segment according to the segmentation characteristics of the fusion context information of each segmentation contained in each text segment.

As an optional implementation manner, the obtaining characteristics of each text segment in the target text further includes:

and performing fusion coding processing on the characteristics of each text segment to obtain the text segment characteristics of the fusion context information of each text segment.

As an optional implementation manner, obtaining characteristics of each text segment of the reference text includes:

performing word segmentation processing on the reference text, and determining each word segmentation contained in the reference text;

and respectively extracting the word segmentation characteristics of the fusion context information of each word segmentation contained in the reference text.

As an optional implementation manner, when the number of texts in the reference text is greater than 1, before performing word segmentation processing on the reference text and determining each word segmentation contained in the reference text, the method further includes:

and merging or screening the reference texts according to the relationship among the reference texts.

As an optional implementation, the method further includes:

and carrying out feature fusion processing on the features of each text segment of the target text and the features of each text segment of the reference text to obtain the target text features fused with the reference text features and/or the reference text features fused with the target text features.

As an optional implementation manner, performing feature fusion processing on features of each text segment of the target text and features of each text segment of the reference text to obtain reference text features fusing target text features, includes:

performing feature fusion processing on chapter features of the target text and/or features of each text segment of the target text and features of each text segment of the reference text to obtain reference text features fusing the features of the target text;

the chapter characteristics of the target text are determined according to the characteristics of each text segment of the target text.

As an optional implementation manner, performing feature fusion processing on features of each text segment of the target text and features of each text segment of the reference text to obtain a target text feature fused with the reference text feature and a reference text feature fused with the target text feature, includes:

determining chapter characteristics of the target text according to the characteristics of each text fragment of the target text;

performing feature fusion processing on the chapter features of the target text and the features of each text segment of the reference text to obtain text segment features of each text segment of the reference text, which are fused with the target text chapter features;

and performing feature fusion processing on the text segment features of the fusion target text chapter features of each text segment of the reference text and the features of each text segment of the target text to obtain the target text features of the fusion reference text features and the reference text features of the fusion target text features.

Specifically, the specific working contents of each part in each embodiment of the text summary generation device are referred to the specific contents of the corresponding processing steps in each embodiment of the text summary generation method, and the description thereof is not repeated here.

Another embodiment of the present application further provides a text summary generating device, as shown in fig. 10, the device including:

a memory 200 and a processor 210;

wherein, the memory 200 is connected to the processor 210 for storing programs;

the processor 210 is configured to implement the text summary generation method disclosed in any of the above embodiments by running the program stored in the memory 200.

Specifically, the text summary generating device may further include: a bus, a communication interface 220, an input device 230, and an output device 240.

The processor 210, the memory 200, the communication interface 220, the input device 230, and the output device 240 are connected to each other through a bus. Wherein:

a bus may include a path that transfers information between components of a computer system.

The processor 210 may be a general-purpose processor, such as a general-purpose Central Processing Unit (CPU), microprocessor, etc., an application-specific integrated circuit (ASIC), or one or more integrated circuits for controlling the execution of programs in accordance with the present invention. But may also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), an off-the-shelf programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components.

The processor 210 may include a main processor and may also include a baseband chip, modem, and the like.

The memory 200 stores programs for executing the technical solution of the present invention, and may also store an operating system and other key services. In particular, the program may include program code including computer operating instructions. More specifically, memory 200 may include a read-only memory (ROM), other types of static storage devices that may store static information and instructions, a Random Access Memory (RAM), other types of dynamic storage devices that may store information and instructions, a disk storage, a flash, and so forth.

The input device 230 may include a means for receiving data and information input by a user, such as a keyboard, mouse, camera, scanner, light pen, voice input device, touch screen, pedometer, or gravity sensor, among others.

Output device 240 may include equipment that allows output of information to a user, such as a display screen, a printer, speakers, and the like.

Communication interface 220 may include any device that uses any transceiver or the like to communicate with other devices or communication networks, such as an ethernet network, a Radio Access Network (RAN), a Wireless Local Area Network (WLAN), etc.

The processor 210 executes the program stored in the memory 200 and invokes other devices, which can be used to implement the steps of any one of the text summary generation methods provided in the above-described embodiments of the present application.

Another embodiment of the present application further provides a storage medium, where a computer program is stored on the storage medium, and when the computer program is executed by a processor, the computer program implements the steps of any one of the text summary generation methods provided in the foregoing embodiments of the present application.

Specifically, the specific working content of each part of the text summary generation device and the specific processing content of the computer program on the storage medium when being executed by the processor may refer to the content of each embodiment of the text summary generation method, which is not described herein again.

While, for purposes of simplicity of explanation, the foregoing method embodiments have been described as a series of acts or combination of acts, it will be appreciated by those skilled in the art that the present application is not limited by the order of acts or acts described, as some steps may occur in other orders or concurrently with other steps in accordance with the application. Further, those skilled in the art should also appreciate that the embodiments described in the specification are preferred embodiments and that the acts and modules referred to are not necessarily required in this application.

It should be noted that, in the present specification, the embodiments are all described in a progressive manner, each embodiment focuses on differences from other embodiments, and the same and similar parts among the embodiments may be referred to each other. For the device-like embodiment, since it is basically similar to the method embodiment, the description is simple, and for the relevant points, reference may be made to the partial description of the method embodiment.

The steps in the method of each embodiment of the present application may be sequentially adjusted, combined, and deleted according to actual needs, and technical features described in each embodiment may be replaced or combined.

The modules and sub-modules in the device and the terminal in the embodiments of the application can be combined, divided and deleted according to actual needs.

In the several embodiments provided in the present application, it should be understood that the disclosed terminal, apparatus and method may be implemented in other manners. For example, the above-described terminal embodiments are merely illustrative, and for example, the division of a module or a sub-module is only one logical division, and there may be other divisions when the terminal is actually implemented, for example, a plurality of sub-modules or modules may be combined or integrated into another module, or some features may be omitted or not executed. In addition, the shown or discussed mutual coupling or direct coupling or communication connection may be an indirect coupling or communication connection through some interfaces, devices or modules, and may be in an electrical, mechanical or other form.

The modules or sub-modules described as separate parts may or may not be physically separate, and parts that are modules or sub-modules may or may not be physical modules or sub-modules, may be located in one place, or may be distributed over a plurality of network modules or sub-modules. Some or all of the modules or sub-modules can be selected according to actual needs to achieve the purpose of the solution of the present embodiment.

In addition, each functional module or sub-module in the embodiments of the present application may be integrated into one processing module, or each module or sub-module may exist alone physically, or two or more modules or sub-modules may be integrated into one module. The integrated modules or sub-modules may be implemented in the form of hardware, or may be implemented in the form of software functional modules or sub-modules.

Those of skill would further appreciate that the various illustrative elements and algorithm steps described in connection with the embodiments disclosed herein may be implemented as electronic hardware, computer software, or combinations of both, and that the various illustrative components and steps have been described above generally in terms of their functionality in order to clearly illustrate this interchangeability of hardware and software. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the implementation. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present application.

The steps of a method or algorithm described in connection with the embodiments disclosed herein may be embodied directly in hardware, in a software unit executed by a processor, or in a combination of the two. The software cells may reside in Random Access Memory (RAM), memory, Read Only Memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

Finally, it should also be noted that, herein, relational terms such as first and second, and the like may be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Also, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising an … …" does not exclude the presence of other identical elements in a process, method, article, or apparatus that comprises the element.

The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the application. Thus, the present application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A text summary generation method, comprising:

2. The method of claim 1, wherein performing summary generation processing on the target text based on the associated content of the reference text located from the target text to obtain a target text summary corresponding to the reference text comprises:

3. The method of claim 2, wherein generating an epoch of the full-text content of the target text based on locating a text segment associated with the reference text from the target text, resulting in a target text epoch corresponding to the reference text, comprises:

4. The method according to claim 3, wherein performing summary generation processing on the full-text content of the target text based on at least the correlation degree of each text segment in the target text related to the reference text and the reference text to obtain a target text summary corresponding to the reference text comprises:

5. The method of claim 3, wherein determining the relevance of each text segment in the target text to the reference text comprises:

6. The method of claim 5, wherein after obtaining the features of the target text and the reference text respectively, the method further comprises:

7. The method as claimed in claim 6, wherein performing an attention-based interactive operation on the features of each text segment in the target text and the features of the reference text to obtain information-refined reference text features comprises:

8. The method according to claim 5, wherein when the number of texts in the reference text is greater than 1, determining the relevance of each text segment in the target text to the reference text according to the feature of each text segment in the target text and the feature of the reference text respectively comprises:

9. The method of claim 5, further comprising:

10. The method of claim 9, further comprising:

11. The method of claim 10, wherein the correcting the correlation between each text segment in the target text and the reference text according to the position distribution of each text segment in the target text comprises:

12. The method of claim 11, further comprising:

13. The method of claim 4, wherein determining the contribution degree of each text segment in the target text to generating the text summary corresponding to the reference text according to the relevance degree of each text segment in the target text to the reference text comprises:

14. The method according to claim 4, wherein performing summary generation processing on full-text content of the target text according to at least the contribution degree of each text segment in the target text to generating a text summary corresponding to the reference text to obtain the target text summary corresponding to the reference text comprises:

15. The method of claim 14, wherein generating text summary decoding features based on at least the features of the target text and the degree of contribution of each text segment in the target text to generating a text summary corresponding to the reference text comprises:

16. The method of claim 15, wherein generating text summary decoding features based on at least the features of the target text, the features of the reference text, and the degree of contribution of each text segment in the target text to generating a text summary corresponding to the reference text comprises:

17. The method of claim 16, further comprising:

18. The method of claim 17, wherein generating reference decoding features according to the features of the reference text and the contribution of each text segment in the reference text to generating a text summary corresponding to the reference text comprises:

19. The method of claim 14, wherein generating text summary decoding features according to the features of the target text and the degree of contribution of each text segment in the target text to generating a text summary corresponding to the reference text comprises:

20. The method of claim 1, wherein obtaining a target text and a reference text, and performing summary generation processing on the target text based on the associated content of the reference text located from the target text to obtain a target text summary corresponding to the reference text comprises:

21. The method according to any one of claims 1 to 20, wherein the characteristics of the target text are obtained by obtaining the characteristics of each text segment in the target text; and the characteristics of the reference text are obtained by acquiring the characteristics of each text segment of the reference text.

22. The method of claim 21, wherein obtaining the characteristics of each text segment in the target text comprises:

23. The method of claim 22, further comprising:

24. The method of claim 21, wherein obtaining the characteristics of each text segment of the reference text comprises:

25. The method of claim 24, wherein when the number of texts in the reference text is greater than 1, before performing word segmentation processing on the reference text to determine each word segment contained in the reference text, the method further comprises:

26. The method of claim 21, further comprising:

27. The method according to claim 26, wherein performing feature fusion processing on features of each text segment of the target text and features of each text segment of the reference text to obtain reference text features fused with the target text features comprises:

28. The method according to claim 26, wherein performing feature fusion processing on features of each text segment of the target text and features of each text segment of the reference text to obtain target text features fused with the reference text features and reference text features fused with the target text features comprises:

29. A text summary generation apparatus, comprising:

30. A text summary generation device, characterized by comprising:

a memory and a processor;

the memory is connected with the processor and used for storing programs;

the processor is configured to implement the text summary generation method according to any one of claims 1 to 28 by executing a program in the memory.

31. A storage medium having stored thereon a computer program which, when executed by a processor, implements a text summary generation method according to any one of claims 1 to 28.