Multi-text abstract generation method and device, equipment, medium and program product

Abstract

The invention provides a multi-text abstract generation method which can be applied to the field of artificial intelligence. The multi-text abstract generation method comprises the steps of obtaining at least two to-be-recognized texts, wherein the to-be-recognized texts are different answer texts for the same question; extracting keywords in each to-be-recognized text, filtering statements which do not contain the keywords in the to-be-recognized text, and serially connecting the reserved statements with the corresponding keywords to obtain a plurality of serially connected statements; constructing a knowledge graph based on each serial statement, and converting the knowledge graph into a plurality of embedded representations; inputting each embedded representation into a pre-trained abstract generation model to obtain probability distribution of the predicted words; and obtaining an answer abstract based on the probability distribution of the predicted words. The invention further provides a multi-text abstract generation device and equipment, a storage medium and a program product.

Description

Technical Field

[0001] This application relates to the financial field, specifically to the field of artificial intelligence, and more specifically to a method, apparatus, device, medium, and program product for generating multi-text summaries. Background Technology

[0002] Traditional multi-document summarization techniques simply concatenate text and treat it as a single-text summarization task, which not only increases computational cost but also struggles to generate concise and informative summaries. Furthermore, while two-stage summarization methods that prioritize retrieval before generation achieve good performance, this approach can lead to high redundancy and a lack of coherence in the generated summaries. Transforming multiple documents into a graph model alleviates hardware memory limitations to some extent, but ensuring lossless information output to the generative model and accurately selecting core information from the massive graph structure remains a significant challenge. Summary of the Invention

[0003] In view of the above problems, this application provides a method, apparatus, device, medium and program product for generating multi-text summaries that improves comprehensiveness and conciseness.

[0004] According to a first aspect of this application, a multi-text summarization method is provided, the method comprising: acquiring at least two texts to be identified, each text being a different answer text for the same question; extracting keywords from each text to be identified, filtering statements in the text that do not contain the keywords, and concatenating the retained statements with the corresponding keywords to obtain multiple concatenated statements; constructing a knowledge graph based on each concatenated statement, and converting the knowledge graph into multiple embedded representations; inputting each embedded representation into a pre-trained summarization generation model to obtain a probability distribution of predicted words; and obtaining an answer summary based on the probability distribution of the predicted words.

[0005] According to an embodiment of this application, constructing a knowledge graph based on each of the concatenated statements includes: mapping each of the concatenated statements and the statement containing the question to multiple vertices of the knowledge graph, and establishing corresponding edges between each of the vertices; and calculating the semantic similarity between each of the vertices, determining the connection weight of each of the edges based on the semantic similarity, and obtaining the knowledge graph.

[0006] According to an embodiment of this application, converting the knowledge graph into multiple embedded representations includes: generating a word-level embedded representation, a graph structure embedded representation, and a positional embedded representation for each vertex in the knowledge graph, wherein the word-level embedded representation is used to characterize the semantic information of the vertex corresponding to the answer text, the graph structure embedded representation is used to characterize the topological relationship and neighbor information of the vertex in the knowledge graph, and the positional embedded representation is used to characterize the positional information of the vertex in the answer text; and summing the word-level embedded representation, the graph structure embedded representation, and the positional embedded representation to obtain the embedded representation of each vertex.

[0007] According to an embodiment of this application, the step of inputting each of the embedded representations into a pre-trained summary generation model to obtain the probability distribution of predicted words includes: performing hierarchical encoding on each of the embedded representations based on a multi-head self-attention target encoder to obtain hidden vector representations; fusing the connection weights of the edges in the knowledge graph with the hidden vector representations of each of the vertices to generate a structure-enhanced graph encoding representation; performing sequence generation based on a multi-head self-attention target decoder according to the graph encoding representation, and fusing the semantic features of the vertices corresponding to the question and the structural features of the knowledge graph to obtain predicted words; and calculating the mixed probability distribution of the generation and duplication of the predicted words through a pointer generation mechanism.

[0008] According to an embodiment of this application, training the summary generation model includes: calculating the prediction loss of the predicted word based on the minimum negative log-likelihood function, and calculating the coverage loss of the attention distribution based on the coverage loss function; weighting the prediction loss and the coverage loss through hyperparameters to obtain a total loss function, wherein the hyperparameters are used to balance the contribution of different loss terms; and adjusting the parameters of the summary generation model based on the total loss function until the total loss function satisfies a preset convergence condition.

[0009] According to an embodiment of this application, the step of extracting keywords from each of the texts to be identified and filtering out statements in the texts to be identified that do not contain the keywords includes: segmenting the texts to be identified into words and extracting keywords from the segmented texts to be identified, wherein the keywords include named entities and keywords; determining whether the keywords exist in each statement in the texts to be identified; filtering the statement when the keywords do not exist in the statement; or retaining the statement when the keywords exist in the statement.

[0010] According to an embodiment of this application, the method further includes: performing data cleaning on the text to be identified, removing web page tags and target characters from the text to be identified to obtain plain text data; and organizing the plain text data into a combination of question and answer sequences, wherein the order of the answer sequences is sorted according to the relevance of each answer text to the corresponding question.

[0011] A second aspect of this application provides a multi-text summarization generation apparatus, comprising: an acquisition module for acquiring at least two texts to be identified, each text being a different answer text for the same question; a concatenation module for extracting keywords from each text to be identified, filtering statements in the texts that do not contain the keywords, and concatenating the retained statements with the corresponding keywords to obtain multiple concatenated statements; a construction module for constructing a knowledge graph based on each concatenated statement, and converting the knowledge graph into an embedded representation; a generation module for inputting the embedded representation into a pre-trained summarization generation model to obtain a probability distribution of predicted words; and a summarization module for obtaining an answer summary based on the probability distribution of the predicted words.

[0012] A third aspect of this application provides an electronic device comprising: one or more processors; and a memory for storing one or more computer programs, wherein the one or more processors execute the one or more computer programs to implement the steps of the method described above.

[0013] A fourth aspect of this application also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0014] The fifth aspect of this application also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described method. Attached Figure Description

[0015] The above-mentioned contents, other objects, features and advantages of this application will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0016] Figure 1 The illustrations depict application scenarios of the multi-text digest generation method, apparatus, device, medium, and program product according to embodiments of this application.

[0017] Figure 2 A schematic diagram illustrating a summary generation model according to an embodiment of this application is shown.

[0018] Figure 3 A flowchart illustrating a multi-text digest generation method according to an embodiment of this application is shown schematically.

[0019] Figure 4 This illustration schematically shows another flowchart of a multi-text digest generation method according to an embodiment of the present application;

[0020] Figure 5 This schematically illustrates a structural block diagram of a multi-text summarization generation apparatus according to an embodiment of this application; and

[0021] Figure 6 A block diagram schematically illustrates an electronic device suitable for implementing a multi-text summarization method according to an embodiment of this application. Detailed Implementation

[0022] The embodiments of this application will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of this application. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of this application for ease of explanation. However, it will be apparent that one or more embodiments may be implemented without these specific details. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application.

[0023] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0024] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0025] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0026] In the technical solution of this application, the user information (including but not limited to user personal information, user image information, user device information, such as location information) and data (including but not limited to data used for analysis, stored data, and displayed data) involved are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of related data all comply with relevant laws, regulations, and standards, take necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation entry points for users to choose to authorize or refuse.

[0027] In scenarios involving automated decision-making using personal information, the methods, devices, and systems provided in this application all offer users corresponding entry points for choosing to agree to or reject the automated decision-making results. If the user chooses to reject, the process proceeds to the expert decision-making stage. Here, "automated decision-making" refers to the activity of automatically analyzing and evaluating an individual's behavioral habits, interests, or economic, health, and credit status through computer programs, and then making a decision. Here, "expert decision-making" refers to the activity of making decisions by personnel who specialize in a particular field, possess specialized experience, knowledge, and skills, and have reached a certain level of professional expertise.

[0028] Embodiments of this application provide a method, apparatus, device, medium, and program product for generating multitext summaries.

[0029] Figure 1 The illustrations depict application scenarios of the multitext digest generation method, apparatus, device, medium, and program product according to embodiments of this application.

[0030] like Figure 1 As shown, application scenario 100 according to this embodiment may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 serves as a medium for providing a communication link between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. The network 104 may include various connection types, such as wired or wireless communication links, or fiber optic cables, etc.

[0031] Users can use the first terminal device 101, the second terminal device 102, and the third terminal device 103 to interact with the server 105 via the network 104 to receive or send messages, etc. Various communication client applications can be installed on the first terminal device 101, the second terminal device 102, and the third terminal device 103, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social media platform software, etc. (for example only).

[0032] The first terminal device 101, the second terminal device 102, and the third terminal device 103 can be various electronic devices with displays and support web browsing, including but not limited to smartphones, tablets, laptops, and desktop computers.

[0033] Server 105 can be a server that provides various services, such as a backend management server that supports websites browsed by users using the first terminal device 101, the second terminal device 102, and the third terminal device 103 (this is just an example). The backend management server can analyze and process data such as received user requests, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices.

[0034] It should be noted that the multi-text summarization method provided in this application embodiment can generally be executed by server 105. Correspondingly, the multi-text summarization device provided in this application embodiment can generally be located in server 105. The multi-text summarization method provided in this application embodiment can also be executed by a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105. Correspondingly, the multi-text summarization device provided in this application embodiment can also be located in a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105.

[0035] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0036] The following will be based on Figure 1 The described scene, through Figures 2-6 A multi-text digest generation method according to embodiments of this application will be described in detail.

[0037] First, the technical terms used in this article are explained and clarified as follows.

[0038] Graph knowledge transformation refers to converting textual knowledge into graph form without weakening the knowledge itself.

[0039] Knowledge learning refers to the ability of a model to understand the meaning of knowledge, know the carrier of knowledge, and to memorize, retain, recognize, and reproduce knowledge through training.

[0040] Multi-document summary generation refers to the process of integrating and processing multiple paragraphs from multiple articles on the same or similar topics to generate a comprehensive, key, and coherent summary summary.

[0041] The inventors discovered that multi-text answer summarization models in related technologies are mainly divided into extractive and generative types. Extractive models generate summaries by extracting and merging sentences from the input document, while generative models generate summaries using arbitrary words and expressions based on document understanding. Early researchers attempted to simply transfer single-text summarization models to multi-document summarization tasks, primarily by training the model on large-scale single-text summarization datasets and fine-tuning it to adapt to multi-document summarization tasks. To overcome the limitations of multi-document summarization model training and the hardware limitations caused by excessively long input text in multi-document summarization, researchers have attempted to introduce graph structures into text representation. The basic idea is to mine the relationships between input documents or sentences, incorporate them into a relationship graph, calculate the importance of sentences or words in the document, and combine pre-trained language models and graph models to facilitate multi-document summarization tasks.

[0042] In embodiments of this application, training the summary generation model includes: calculating the prediction loss of the predicted word based on the minimum negative log-likelihood function, and calculating the coverage loss of the attention distribution based on the coverage loss function; weighting the prediction loss and the coverage loss through hyperparameters to obtain a total loss function, wherein the hyperparameters are used to balance the contribution of different loss terms; and adjusting the parameters of the summary generation model based on the total loss function until the total loss function satisfies a preset convergence condition.

[0043] To ensure data diversity and usability, this embodiment utilizes programming languages ​​to collect data from real-world internet-based community-based question-and-answer platforms, obtaining a large number of user-generated question-and-answer pairs as training data. Then, the collected raw data can be preprocessed, such as removing webpage tags, special characters, segmenting words, and removing stop words. Each data pair can then be divided according to a certain ratio (e.g., 8:1:1) to form a training set, a validation set, and a test set. The multi-text answer summarization task can be defined as: given a question Q and L corresponding correct answers... The goal is to generate a concise summary S from these correct but lengthy answer paragraphs that contains the core of all the answer paragraphs, and to write a reference answer based on the highest-ranking answer in terms of relevance between the question and the different answers.

[0044] In the task of summarizing, the model needs to predict the next most likely word. To measure the accuracy of the model's prediction, the Negative Log Likelihood Loss (NLL Loss) function can be used to calculate the prediction loss for the predicted word. Meanwhile, during the summarizing process, to avoid duplicate generation or omission of important information, it is necessary to ensure that the attention mechanism can evenly and comprehensively focus on different parts of the input text. Therefore, a Coverage Loss function can be introduced to measure the reasonableness of the attention distribution.

[0045] The minimum negative log-likelihood function and the coverage loss function can together constitute the total loss function for model training. The minimum negative log-likelihood function L... loss The mathematical expression is as follows:

[0046]

[0047] Where T represents the length of the abstract, y t P(y) represents the actual word at time step t. t ) indicates that the model predicts y t The probability of the next word.

[0048] The mathematical expression for the coverage loss function is as follows:

[0049]

[0050] Where T represents the length of the abstract, This represents the attention weight of the model for the i-th word in the input text at time step t. This represents the cumulative attention weight of the model for the i-th word in the input text up to the t-th time step (it can be initialized to 0 and updated according to the attention weight at each time step).

[0051] To balance the impact of the minimum negative log-likelihood function and the coverage loss function on model training, a hyperparameter λ can be introduced. l and λ a We then perform a weighted summation of these two loss terms to obtain the total loss function.

[0052]

[0053] in, , These are all hyperparameters, used to balance different losses; they can be adopted... .

[0054] During model training, optimization algorithms can be used to calculate the gradient of the total loss function with respect to the model parameters, and the parameters can be updated in the opposite direction of the gradient to gradually reduce the value of the total loss function. Simultaneously, preset convergence conditions can be set, such as the total loss function value being less than a preset threshold, or the model's evaluation metric on the validation set reaching a preset level. When the model has been trained continuously and meets the preset convergence conditions, training can be stopped, and the model parameters can be saved.

[0055] Figure 2 A schematic diagram of a summary generation model according to an embodiment of this application is shown.

[0056] After the model training is completed, such as Figure 2 As shown, the model can be tested on a validation set, and the model that performs best on the validation set can be saved. Then, ROUGE (Recall-Oriented Understudy for Gisting Evaluation) can be used for automatic summary evaluation. This involves using the standard ROUGE metric to automatically evaluate the model and calculate the F1 scores of ROUGE-1, ROUGE-2, and ROUGE-L in the embodiments of this application, measuring the similarity between the reference summary and the generated summary based on word overlap, bigram overlap, and longest common sequence, respectively. The specific calculation method is as follows.

[0057]

[0058]

[0059]

[0060]

[0061] Where n=1, 2, g is an n-gram, and S and R represent the summary and reference, respectively. Two multisets A∩B are represented as the intersection of A and B, defined as the multiset containing all multiples of their shared elements. LCS(S,R) is the length of the longest common subsequence of S and R. m and n represent the lengths of the reference summary and the automatic summary (generally the number of words contained), respectively, and R... lcs and P lcs It displays recall and precision, then reports the F1 score. B can be set to a large number. When there are multiple reference abstracts, the ROUGE points for each reference can be calculated and the average returned.

[0062] Because generative models have many parameters and a long training cycle, large graphics processors can be used to train the models and save the training models for each training epoch. The optimal model is selected using the validation set, and finally the trained detection model is tested using the test set to ensure that the model can generate answer summaries with high fluency and readability.

[0063] Understandably, by fusing the minimum negative log-likelihood and the coverage loss function, we can not only optimize the syntactic accuracy and semantic coherence of the generated summaries, but also reduce redundancy and ensure the coverage of key information. Furthermore, by introducing hyperparameters, the weights of prediction loss and coverage loss can be flexibly controlled, thus adapting to summarization tasks in different domains.

[0064] Figure 3 A flowchart illustrating a multi-text digest generation method according to an embodiment of this application is shown schematically.

[0065] like Figure 3 As shown, the multi-text summary generation method of this embodiment includes operations S310 to S350.

[0066] In operation S310, at least two texts to be identified are obtained, each of which is a different answer text for the same question.

[0067] In embodiments of this application, user consent or authorization may be obtained before acquiring at least two texts to be identified. For example, before operation S310, a request may be sent to the user to acquire at least two texts to be identified. Operation S310 is performed only if the user consents or authorizes the acquisition of at least two texts to be identified.

[0068] In operation S320, keywords are extracted from each of the texts to be identified, and statements in the texts to be identified that do not contain the keywords are filtered out. The retained statements are then concatenated with the corresponding keywords to obtain multiple concatenated statements.

[0069] In operation S330, a knowledge graph is constructed based on each of the concatenated statements, and the knowledge graph is converted into multiple embedded representations.

[0070] In operation S340, the embedded representations are input into a pre-trained summary generation model to obtain the probability distribution of the predicted words.

[0071] In operation S350, an answer summary is obtained based on the probability distribution of the predicted words.

[0072] In the embodiments of this application, a corresponding operation entry can be provided to the user, allowing the user to choose to agree to or reject the automated decision result. That is, before obtaining the answer summary processing / decision based on the probability distribution of predicted words, the user can input their agreement or rejection of the processing / decision through the corresponding operation entry. If the user agrees to the processing / decision, the answer summary processing / decision is obtained based on the probability distribution of predicted words, i.e., step S350 is executed. If the user rejects the processing / decision, the expert decision-making process is initiated.

[0073] First, at least two differentiated answer texts for the same question can be obtained, meaning each text to be identified has a similarity greater than or equal to a threshold. Then, for each text to be identified, keywords can be extracted to summarize the core content, providing a basis for subsequent sentence filtering and concatenation. Based on the extracted keywords, sentences in each text to be identified can be filtered out, removing those that do not contain keywords, i.e., redundant information or descriptions not closely related to the core question. Furthermore, the retained sentences can be concatenated with corresponding keywords to obtain multiple concatenated sentences for different keywords. Based on these concatenated sentences, a knowledge graph can be constructed, where key points serve as nodes relating sentences; that is, keywords in the concatenated sentences are used as nodes in the knowledge graph, and the content expressed by the sentences reflects the relationships between nodes. The constructed knowledge graph is then converted into multiple embedded representations, mapping the nodes and relationships in the knowledge graph to a low-dimensional vector space, preserving the semantic and structural information in the knowledge graph. This is then input into a summarization generation model to predict the probability distribution of the next word. Finally, based on the predicted word probability distribution, an appropriate decoding strategy can be used to generate the final answer summary.

[0074] For example, three different answers to the question "How to assess the investment value of a stock in an emerging technology company?" can be obtained first. These answers can come from professional financial forums, authoritative financial websites, and investment research reports. Then, keywords can be extracted from each text to be identified. Keywords may include "revenue growth rate," "market share," "industry competitive landscape," "technological innovation capability," and "macroeconomic environment." These keywords are then concatenated with corresponding statements. For instance, a statement containing the keyword "revenue growth rate" can be combined with that keyword to form a concatenated statement centered on that keyword. Based on these concatenated statements, a knowledge graph about the emerging technology company's stock can be constructed. The "revenue growth rate" node in the knowledge graph may be connected to the "company profitability" node, indicating that revenue growth rate is an important indicator of a company's profitability. Furthermore, the knowledge graph can be converted into an embedded representation and input into a summary generation model. The model can generate a series of word probability distributions based on the embedded representation of the knowledge graph. During the generation process, the model can select words with higher probabilities to combine into meaningful sentences, gradually constructing a complete summary. For example, the final summary could be: "Based on a comprehensive assessment, the stock of this emerging technology company has certain investment value. Its revenue growth rate is good, and its market share is gradually expanding, demonstrating strong market competitiveness. However, the industry is highly competitive, and uncertainties in the macroeconomic environment may affect the company's performance; investors should make prudent decisions."

[0075] It is understood that this application provides a multi-text answer summary generation method based on graph knowledge transformation and learning, which aims to summarize multiple correct answers and extract a comprehensive and concise answer summary. From the perspective of text representation, a graph knowledge transformation and learning method is proposed to avoid the problem of slow model training convergence caused by excessively long input text; from the perspective of text generation, graph learning and text learning are combined, and decoding is performed under the guidance of the question, thereby ensuring the fluency and readability of the generated summary.

[0076] In embodiments of this application, the method further includes: performing data cleaning on the text to be identified, removing web page tags and target characters from the text to be identified to obtain plain text data; and organizing the plain text data into a combination of question and answer sequences, wherein the order of the answer sequences is sorted according to the relevance of each answer text to the corresponding question.

[0077] Figure 4 Another flowchart illustrating a multi-text summary generation method according to an embodiment of this application is shown schematically.

[0078] like Figure 4As shown, firstly, data can be collected to obtain multiple texts to be recognized. Then, the data can be preprocessed, i.e., data cleaning of the dirty data in the texts to be recognized according to certain rules, retaining only the question and its corresponding answer text to form the initial dataset required by the model. For example, firstly, web page tags are removed, and rule-based methods are used to roughly filter the data, removing target characters such as emoticons, special characters, and unrecognizable characters, ensuring that each question-answer pair after filtering is pure text data. Further, the pure text data can be processed, and the processed data format is a combination of text "question + answer 1 + answer 2 + ... + answer L", where the order of the L answers can be sorted according to their relevance to the corresponding question. It should be noted that all the answers here are correct answers to the corresponding questions.

[0079] To determine the relevance of an answer to a question, various methods can be used, such as keyword matching and semantic similarity calculation. For example, for the question "Is it worthwhile to buy the stock of a certain emerging technology company?", the relevance of the answer to the question can be determined by analyzing whether the answer contains keywords related to the question (such as "revenue growth", "technological innovation capability", "industry competition", "macroeconomic environment", etc.) and the semantic similarity, and then sorting the answers from highest to lowest relevance.

[0080] Understandably, cleaning and organizing the text to be identified can improve text quality, facilitate efficient subsequent processing and accurate retrieval, and enhance information utilization efficiency.

[0081] Based on the above embodiments, in this embodiment, the step of extracting keywords from each of the texts to be identified and filtering out statements in the texts to be identified that do not contain the keywords includes: segmenting the texts to be identified into words and extracting keywords from the segmented texts to be identified, wherein the keywords include named entities and keywords; determining whether the keywords exist in each statement in the texts to be identified; filtering the statement when the keywords do not exist in the statement; or retaining the statement when the keywords exist in the statement.

[0082] Continue to refer to Figure 4 Because correct answers may contain a lot of noise—that is, while an answer may clarify the user's question, it may contain many sentences irrelevant to the question, increasing the user's reading burden and reducing the user experience—after data preprocessing, for each answer paragraph A... iFirst, natural language processing tools can be used for word segmentation and named entity recognition, such as names of people, places, organizations, times, dates, currencies, and percentages. Since named entities themselves are insufficient to cover the main information of a document, keyword extraction algorithms can be further used to obtain additional keywords. Then, each sentence in the text to be recognized can be evaluated; if a sentence does not contain keywords, it can be discarded, thus reducing the training burden on the model to some extent. If a sentence contains keywords, the sentence and keywords are concatenated as the model's original input format.

[0083] Only statements containing keywords are retained, potentially leading to the accidental deletion of semantically related statements that do not explicitly mention keywords (such as the semantic connection between "using loop structures" and the keyword "iteration"). Therefore, the keyword set of each text can be extracted and dynamically expanded to include synonyms, near-synonyms, hyponyms, and high-frequency phrases, forming a more comprehensive semantic coverage. Then, the semantic relevance between statements and keywords can be calculated using the expanded keyword set, retaining statements with relevance above a threshold. This effectively retains semantically related statements that do not explicitly contain the initial keywords while filtering redundant information, improving the completeness and accuracy of information retention.

[0084] Understandably, filtering out statements that do not contain keywords can significantly reduce the amount of text data that needs further analysis, speed up processing, and ensure that the analysis process focuses on the core content of the text.

[0085] Based on the above embodiments, in this embodiment, constructing a knowledge graph based on each of the concatenated statements includes: mapping each of the concatenated statements and the statement containing the question to multiple vertices of the knowledge graph, and establishing corresponding edges between each of the vertices; and calculating the semantic similarity between each of the vertices, determining the connection weight of each of the edges based on the semantic similarity, and obtaining the knowledge graph.

[0086] In the knowledge transformation layer, multiple answer paragraphs can be represented as topic-based interaction graphs. Text sentences are decomposed into several topic-centric text clusters, each representing a key aspect of the article. Each cluster and topic forms a vertex in the graph, and the edges between vertices can be obtained by calculating the semantic relationships between them.

[0087] After obtaining named entities and keywords (collectively referred to as keywords), the keywords in answer set A can be obtained. Then, all the sentences in the answers can be associated with the keywords they contain. That is, if the keyword k appears in a sentence... n Then you can choose answer A. i The j-th sentence s ijKeywords are assigned. Since a sentence may contain multiple keywords, a knowledge graph allows a sentence to be associated with multiple keywords, implying relationships between two topics. Sentences without any keywords are considered to lack key information useful for answer generation and can be filtered out. For questions, since understanding sentences in candidate answers is crucial, a special vertex called `que` containing the question text can be added during graph construction. Ultimately, the words in each vertex represent a keyword aspect of the answer, sentences, and the keyword `k` to which these sentences belong. n A vertex v is formed in the interaction graph. i Then, based on the relevance and logic of the text content, edges are established between the vertices to represent the semantic or logical relationships between these vertices. For edge calculation in the interaction graph, a content-based method can be used to calculate the vertex v. i With v j The text content similarity (i.e., semantic similarity) between the edges is used as the connection weight of the edges. The weight is used to represent the strength of the relationship between vertices. For example, the weight value can be a number between 0 and 1, with a larger value indicating a stronger relationship.

[0088] Understandably, the semantic similarity-driven edge weight allocation mechanism can automatically filter redundant or weakly related content, thereby forming a hierarchical knowledge network centered on questions in the knowledge graph. This not only preserves the core relevance of multi-source answers, but also enhances the ability of subsequent summary generation models to focus on key information through weight differences, ultimately outputting more compact and semantically coherent answer summaries.

[0089] Based on the above embodiments, in this embodiment, converting the knowledge graph into multiple embedded representations includes: generating a word-level embedding representation, a graph structure embedding representation, and a positional embedding representation for each vertex in the knowledge graph, wherein the word-level embedding representation is used to characterize the semantic information of the vertex corresponding to the answer text, the graph structure embedding representation is used to characterize the topological relationship and neighbor information of the vertex in the knowledge graph, and the positional embedding representation is used to characterize the positional information of the vertex in the answer text; and summing the word-level embedding representation, the graph structure embedding representation, and the positional embedding representation to obtain the embedded representation of each vertex.

[0090] For the embedded representation e of each vertex in the knowledge transformation graph i Besides the embedded representation of words In addition, graph structure embedding representations can be added. And the embedded representation of the position of each sentence in each answer. Therefore, the final embedded representation of each vertex is:

[0091]

[0092] Word-level embeddings can aggregate pre-trained word vectors (such as average pooling or attention weighting) of the text corresponding to each vertex (keywords / sentences / questions) to capture the semantic information of the vertex text, ensuring the model understands the answer content. Graph-structured embeddings can encode the topological relationships of vertices in the graph using deep learning models specifically designed for graph-structured data, capturing their association features with neighboring vertices and revealing the logical connections between answer elements. Positional embeddings encode the positional information of vertices in the original answer text (such as sentence number or paragraph position), and can be implemented using a learnable positional encoding matrix.

[0093] Understandably, by fusing three modalities—word-level embedding, graph structure embedding, and positional embedding—a vertex representation with semantic depth, structural relevance, and positional awareness is generated, providing richer contextual information for subsequent summary generation.

[0094] Based on the above embodiments, in this embodiment, the step of inputting each of the embedded representations into a pre-trained summary generation model to obtain the probability distribution of predicted words includes: performing hierarchical encoding on each of the embedded representations based on a multi-head self-attention target encoder to obtain hidden vector representations; fusing the connection weights of the edges in the knowledge graph with the hidden vector representations of each vertex to generate a structure-enhanced graph encoding representation; performing sequence generation based on a multi-head self-attention target decoder according to the graph encoding representation, and fusing the semantic features of the vertex corresponding to the question and the structural features of the knowledge graph to obtain predicted words; and calculating the mixed probability distribution of the generation and duplication of the predicted words through a pointer generation mechanism.

[0095] At the encoding layer, a target encoder, such as a vertex encoder based on a multi-head self-attention mechanism, can be used to encode the embedded representation of each vertex in the interaction graph into a vector representation. The input of the first layer is the embedded representation e. i The input to each subsequent layer is the output of the previous layer. After L... T After layer encoding, the hidden vector of the last layer is used. It is a hidden vector representation representing the entire vertex.

[0096] After obtaining the hidden vector representation of each vertex in the graph, a graph encoder can be used to integrate the edge weight information of the interaction graph into the vertex vectors. This encoder encodes the graph using the graphical structure of the constructed topic interaction graph. In this process, the graph encoder can model the content of the graph vertices and fully utilize the structural information of the graph. Given an adjacency matrix... ,in Add an edge pointing to the node itself to each vertex in the interaction graph:

[0097]

[0098] in, It belongs to the identity matrix. Let D be a diagonal matrix. ,but:

[0099]

[0100] in, Represents a standardized symmetric adjacency matrix. Represents a learnable matrix. The ReLU activation function is used. Furthermore, to prevent the graph encoder from becoming overly smooth, residual connections between layers can be added.

[0101]

[0102] Finally, the final output of the graph encoder can be obtained. A feedforward layer is added to the final output of the graph encoder to obtain the final graph encoded representation of the graph encoder output. :

[0103]

[0104] For the decoding layer, a target decoder can be used to convert the context information generated by the encoder into the output of the target sequence. Each layer of the decoder can include a masked self-attention mechanism, a multi-head attention mechanism, a feedforward neural network, residual connections, and layer normalization. The masked self-attention mechanism ensures that when generating the output at each position, attention is only focused on that position and its preceding positions, preventing information leakage. The multi-head attention mechanism can capture information from different representation subspaces, enhancing expressive power. For the initial state of the decoder, weights can be calculated for the question and answer, and the initial decoding transition H can be selected based on the weight of each candidate answer. A The final output of the encoder, according to the graph, represents... The decoder can perform sequence generation, producing a series of words. For the current step t of decoding, the decoder bases its work on the predicted word generated in the previous step. Generate a hidden state s t To simultaneously consider the impact of semantic information within both the question and the answer on the generated words during the decoding stage, the question text encoding (i.e., the semantic features of the vertices) and the graph encoding output (i.e., the structural features of the knowledge graph) can be taken into account at the same time:

[0105]

[0106]

[0107]

[0108]

[0109] Among them, H Q The semantic encoding output of question Q, a t c represents the learning weights of the problem and the graph in the current decoding step. t This represents the probability distribution of the source word. Finally, a pointer generation network based on a pointer generation mechanism can be used to obtain the final probability distribution between the source word and the copied word:

[0110]

[0111] Among them, P vocab Let be the probability distribution of the vocabulary, representing the probability of generating each word from a fixed vocabulary; s t c represents the hidden state of the decoder at the current step t; t The context vector is represented by W1 and W2 (calculated through the attention mechanism and fused with the encoder output); W1 and W2 represent learnable weight matrices used to map the concatenated hidden state and context vector to the vocabulary space; b1 and b2 are bias terms; tan h The activation function introduces non-linearity; softmax represents normalization to a probability distribution.

[0112]

[0113] Among them, P gen σ represents the generation probability, used to determine whether the current step generates a new word (from the vocabulary) or copies a word from the input; σ is the sigmoid function, outputting the probability in the interval [0,1]; w c ,w y ,w s ,w q Let H represent the learnable weight vectors, corresponding to the contributions of the context vector, the predicted word from the previous step, the decoder hidden state, and the global representation of the question, respectively; y represents the predicted word (word embedding vector) from the previous step; H q b represents the global semantic representation of the problem text (such as the encoded average value); b is the bias term.

[0114]

[0115] Where P represents the final probability distribution, balancing generation and replication. For attention weights, ω i=ω indicates filtering input words that are the same as the word to be generated.

[0116] Understandably, combining knowledge graph edge weights with vertex vectors strengthens entity relationships and enhances structured semantic representation. Simultaneously, dynamically balancing generation and duplication probabilities through a pointer generation mechanism improves the accuracy and diversity of generated text.

[0117] This application provides a multi-text summarization method based on graph knowledge transformation and learning. It collects data from real internet community platforms, preprocesses and labels the data to form the initial dataset for the model. Combining the ideas of knowledge graphs and graph neural networks, a knowledge transformation model for long texts is proposed. Natural language processing tools are used to perform named entity recognition and keyword-based node formation on multiple texts. A content-based method is used to calculate the weights of edges between nodes, forming a relationship graph and removing redundant information from long texts. In model training, the idea of ​​a multi-layer neural network model is incorporated, along with graph-based knowledge learning. Using graphs to guide the summarization process facilitates the generation of coherent and concise summaries. This application fully leverages the characteristics of massive data content on internet community platforms, enabling automatic and effective management of information content, increasing the storage capacity of effective and high-quality content, thereby improving the overall quality of internet content and enhancing user readability and information retrieval speed.

[0118] Based on the above-described multi-text summarization method, this application also provides a multi-text summarization apparatus. The following will combine... Figure 5 The device is described in detail.

[0119] Figure 5 A schematic block diagram of a multi-text summarization generation apparatus according to an embodiment of this application is shown.

[0120] like Figure 5 As shown, the multi-text summarization generation device 500 of this embodiment includes an acquisition module 510, a concatenation module 520, a construction module 530, a generation module 540, and a summarization module 550.

[0121] The acquisition module 510 is used to acquire at least two texts to be identified, each of which is a different answer text for the same question. In one embodiment, the acquisition module 510 can be used to perform the operation S310 described above, which will not be repeated here.

[0122] The concatenation module 520 is used to extract keywords from each of the texts to be identified, filter out sentences in the texts that do not contain the keywords, and concatenate the retained sentences with the corresponding keywords to obtain multiple concatenated sentences. In one embodiment, the concatenation module 520 can be used to perform the operation S320 described above, which will not be repeated here.

[0123] The construction module 530 is used to construct a knowledge graph based on each of the concatenated statements and convert the knowledge graph into an embedded representation. In one embodiment, the construction module 530 can be used to perform the operation S330 described above, which will not be repeated here.

[0124] The generation module 540 is used to input the embedded representation into a pre-trained summary generation model to obtain the probability distribution of the predicted words. In one embodiment, the generation module 540 can be used to perform the operation S340 described above, which will not be repeated here.

[0125] The summarization module 550 is used to obtain an answer summary based on the probability distribution of the predicted words. In one embodiment, the summarization module 550 can be used to perform the operation S350 described above, which will not be repeated here.

[0126] According to an embodiment of this application, the construction module 530 is specifically used to: map each of the concatenated statements and the statement containing the question to multiple vertices of the knowledge graph, and establish corresponding edges of the knowledge graph between each of the vertices; and calculate the semantic similarity between each of the vertices, determine the connection weight of each of the edges based on the semantic similarity, and obtain the knowledge graph.

[0127] According to an embodiment of this application, the construction module 530 is specifically configured to: generate a word-level embedding representation, a graph structure embedding representation, and a positional embedding representation for each vertex in the knowledge graph, wherein the word-level embedding representation is used to characterize the semantic information of the vertex corresponding to the answer text, the graph structure embedding representation is used to characterize the topological relationship and neighbor information of the vertex in the knowledge graph, and the positional embedding representation is used to characterize the positional information of the vertex in the answer text; and sum the word-level embedding representation, the graph structure embedding representation, and the positional embedding representation to obtain the embedding representation of each vertex.

[0128] According to an embodiment of this application, the generation module 540 is specifically configured to: perform hierarchical encoding on each of the embedded representations based on a multi-head self-attention target encoder to obtain hidden vector representations; fuse the connection weights of the edges in the knowledge graph with the hidden vector representations of each of the vertices to generate a structure-enhanced graph encoding representation; perform sequence generation based on a multi-head self-attention target decoder according to the graph encoding representation, and fuse the semantic features of the vertices corresponding to the question and the structural features of the knowledge graph to obtain predicted words; and calculate the mixed probability distribution of the generation and copying of the predicted words through a pointer generation mechanism.

[0129] According to an embodiment of this application, the generation module 540 can also be used to train the summary generation model, including: calculating the prediction loss of the predicted word based on the minimum negative log-likelihood function, and calculating the coverage loss of the attention distribution based on the coverage loss function; weighting the prediction loss and the coverage loss through hyperparameters to obtain a total loss function, wherein the hyperparameters are used to balance the contribution of different loss terms; and adjusting the parameters of the summary generation model based on the total loss function until the total loss function satisfies a preset convergence condition.

[0130] According to an embodiment of this application, the concatenation module 520 is specifically used for: segmenting the text to be identified into words and extracting keywords from the segmented text to be identified, the keywords including named entities and keywords; determining whether the keyword exists in each sentence of the text to be identified; filtering the sentence when the keyword does not exist in the sentence; or retaining the sentence when the keyword exists in the sentence.

[0131] According to an embodiment of this application, the acquisition module 510 can also be used to: perform data cleaning on the text to be identified, remove web page tags and target characters from the text to be identified, and obtain plain text data; organize the plain text data into a combination of question and answer sequences, wherein the order of the answer sequence is sorted according to the relevance of each answer text to the corresponding question.

[0132] According to embodiments of this application, any multiple modules among the acquisition module 510, the serialization module 520, the construction module 530, the generation module 540, and the summarization module 550 can be merged into one module, or any one of these modules can be split into multiple modules. Alternatively, at least some of the functions of one or more of these modules can be combined with at least some of the functions of other modules and implemented in one module. According to embodiments of this application, at least one of the acquisition module 510, the serialization module 520, the construction module 530, the generation module 540, and the summarization module 550 can be at least partially implemented as hardware circuitry, such as a field-programmable gate array (FPGA), a programmable logic array (PLA), a system-on-a-chip, a system-on-a-substrate, a system-on-package, an application-specific integrated circuit (ASIC), or implemented in hardware or firmware by any other reasonable means of integrating or packaging the circuitry, or implemented in software, hardware, or firmware, or in any suitable combination of any of these three implementation methods. Alternatively, at least one of the acquisition module 510, the concatenation module 520, the construction module 530, the generation module 540, and the summarization module 550 may be implemented at least partially as a computer program module, which can perform corresponding functions when the computer program module is run.

[0133] Figure 6 A block diagram schematically illustrates an electronic device suitable for implementing a multi-text summarization method according to an embodiment of this application.

[0134] like Figure 6 As shown, an electronic device 600 according to an embodiment of this application includes a processor 601, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 602 or a program loaded from a storage portion 608 into a random access memory (RAM) 603. The processor 601 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 601 may also include onboard memory for caching purposes. The processor 601 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of this application.

[0135] RAM 603 stores various programs and data required for the operation of electronic device 600. Processor 601, ROM 602, and RAM 603 are interconnected via bus 604. Processor 601 executes various operations of the method flow according to embodiments of this application by executing programs in ROM 602 and / or RAM 603. It should be noted that the programs may also be stored in one or more memories other than ROM 602 and RAM 603. Processor 601 may also execute various operations of the method flow according to embodiments of this application by executing programs stored in said one or more memories.

[0136] According to embodiments of this application, the electronic device 600 may further include an input / output (I / O) interface 605, which is also connected to a bus 604. The electronic device 600 may also include one or more of the following components connected to the input / output (I / O) interface 605: an input section 606 including a keyboard, mouse, etc.; an output section 607 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 608 including a hard disk, etc.; and a communication section 609 including a network interface card such as a LAN card, modem, etc. The communication section 609 performs communication processing via a network such as the Internet. A drive 610 is also connected to the input / output (I / O) interface 605 as needed. A removable medium 611, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 610 as needed so that computer programs read from it can be installed into the storage section 608 as needed.

[0137] This application also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of this application.

[0138] According to embodiments of this application, the computer-readable storage medium can be a non-volatile computer-readable storage medium, such as including but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to embodiments of this application, the computer-readable storage medium may include ROM 602 and / or RAM 603 and / or one or more memories other than ROM 602 and RAM 603 described above.

[0139] Embodiments of this application also include a computer program product comprising a computer program containing program code for performing the methods shown in the flowchart. When the computer program product is run on a computer system, the program code enables the computer system to implement the multitext digest generation method provided in the embodiments of this application.

[0140] When the computer program is executed by the processor 601, it performs the functions defined in the system / apparatus of this application embodiment. According to the embodiments of this application, the systems, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0141] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and downloaded and installed via the communication section 609, and / or installed from the removable medium 611. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0142] In such an embodiment, the computer program can be downloaded and installed from a network via the communication section 609, and / or installed from the removable medium 611. When the computer program is executed by the processor 601, it performs the functions defined in the system of this application embodiment. According to the embodiments of this application, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0143] According to embodiments of this application, program code for executing the computer programs provided in the embodiments of this application can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. The program code can be executed entirely on the user's computing device, partially on the user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0144] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0145] Those skilled in the art will understand that the features described in the various embodiments of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application. All such combinations and / or combinations fall within the scope of this application.

Claims

1. A method for generating multi-text summaries, characterized in that, The method includes: Obtain at least two texts to be identified, each of which is a different answer text to the same question; Extract keywords from each of the texts to be identified, filter out statements in the texts to be identified that do not contain the keywords, and concatenate the retained statements with the corresponding keywords to obtain multiple concatenated statements; Based on the concatenated statements, a knowledge graph is constructed, and the knowledge graph is converted into multiple embedded representations; The embedded representations are input into a pre-trained summary generation model to obtain the probability distribution of predicted words; and Based on the probability distribution of the predicted words, an answer summary is obtained.

2. The method according to claim 1, characterized in that, The construction of the knowledge graph based on the concatenated statements includes: Map each of the concatenated statements and the statement containing the question to multiple vertices of the knowledge graph, and establish corresponding edges between the vertices; and Calculate the semantic similarity between each vertex, determine the connection weight of each edge based on the semantic similarity, and obtain the knowledge graph.

3. The method according to claim 2, characterized in that, The step of converting the knowledge graph into multiple embedded representations includes: For each vertex in the knowledge graph, a word-level embedding representation, a graph structure embedding representation, and a positional embedding representation are generated. The word-level embedding representation characterizes the semantic information of the vertex corresponding to the answer text; the graph structure embedding representation characterizes the topological relationships and neighbor information of the vertex in the knowledge graph; and the positional embedding representation characterizes the positional information of the vertex in the answer text. The word-level embedding representation, the graph structure embedding representation, and the positional embedding representation are summed to obtain the embedding representation of each vertex.

4. The method according to claim 2 or 3, characterized in that, The step of inputting each of the embedded representations into a pre-trained summary generation model to obtain the probability distribution of predicted words includes: Based on a multi-head self-attention target encoder, the embedded representations are hierarchically encoded to obtain hidden vector representations; The connection weights of the edges in the knowledge graph are fused with the hidden vector representations of each vertex to generate a structure-enhanced graph encoding representation. A target decoder based on multi-head self-attention generates sequences according to the graph encoding representation and fuses the semantic features of the vertices corresponding to the question with the structural features of the knowledge graph to obtain predicted words; and The mixed probability distribution of the generation and copying of the predicted word is calculated using a pointer generation mechanism.

5. The method according to claim 1, characterized in that, Training the summary generation model includes: The prediction loss of the predicted word is calculated based on the minimum negative log-likelihood function, and the coverage loss of the attention distribution is calculated based on the coverage loss function. The prediction loss and the coverage loss are weighted and summed using hyperparameters to obtain the total loss function, where the hyperparameters are used to balance the contributions of different loss terms; and Based on the total loss function, the parameters of the summary generation model are adjusted until the total loss function meets the preset convergence condition.

6. The method according to claim 1, 2 or 3, characterized in that, The step of extracting keywords from each of the texts to be identified and filtering out statements in the texts to be identified that do not contain the keywords includes: The text to be identified is segmented into words, and keywords are extracted from the segmented text. The keywords include named entities and keywords. Determine whether the keyword exists in each statement of the text to be identified; If the keyword is not present in the statement, the statement will be filtered; or If the keyword is present in the statement, the statement will be retained.

7. The method according to claim 1 or 5, characterized in that, The method further includes: The text to be identified is cleaned by removing web page tags and target characters to obtain plain text data. The plain text data is organized into a combination of question and answer sequences, with the order of the answer sequences sorted according to the relevance of each answer text to the corresponding question.

8. A multi-text summarization generation device, characterized in that, The device includes: The acquisition module is used to acquire at least two texts to be identified, each of which is a different answer text to the same question; The concatenation module is used to extract keywords from each of the texts to be identified, filter out statements in the texts to be identified that do not contain the keywords, and concatenate the retained statements with the corresponding keywords to obtain multiple concatenated statements; A construction module is used to construct a knowledge graph based on each of the concatenated statements and convert the knowledge graph into an embedded representation; A generation module is used to input the embedded representation into a pre-trained summary generation model to obtain the probability distribution of predicted words; and The summarization module is used to obtain an answer summary based on the probability distribution of the predicted words.

9. An electronic device, comprising: One or more processors; Memory, used to store one or more computer programs. The characteristic feature is that the one or more processors execute the one or more computer programs to implement the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 7.

11. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 7.

Images