A Multi-Knowledge Graph-Based Adaptive Topic-Optimizing Text Summarization Method

Abstract

This invention provides a method for adaptive topic-selective text summarization based on multiple knowledge graphs, characterized by the following steps: Step S1, preprocessing the document to be identified to obtain preprocessed data; Step S2, obtaining document feature information based on the preprocessed data; Step S3, obtaining a topic context representation vector based on the preprocessed data and document feature information; Step S4, constructing a FastText model as an adaptive text classifier; Step S5, obtaining class feature vectors based on the preprocessed data; Step S6, calculating document category features by combining the class feature vectors and topic context representation vectors; Step S7, obtaining feature matrices corresponding to each external knowledge base using a knowledge vectorization method; Step S8, obtaining connection results based on the knowledge selector and topic context representation vectors; Step S9, inputting the connection results into a decoder to obtain a summary. In summary, this method can generate more accurate corresponding summaries based on documents.

Description

Technical Field

[0001] This invention relates to the field of natural language processing, and more specifically to a method for adaptive topic optimization text summarization based on multiple knowledge graphs. Background Technology

[0002] With the rapid development of internet technology and the continuous improvement of hardware and data volume, machine learning technology has flourished and been widely applied. Deep learning technology is widely used in various fields such as computer vision, natural language processing, and recommendation systems, greatly changing human life. However, the proverb "too much information kills information" reveals the drawbacks of the current information society. In the past two years up to 2021, 90% of the world's data has been created. Most companies only view 12% of this data. According to relevant organizations, by 2025, the amount of data generated will exceed 180 zettabytes. Downloading all the data from the internet now would take one person approximately 181 million years. Therefore, how to utilize internet technology to extract important content from the vast amount of internet information, save users' time, and improve information browsing efficiency is an urgent problem to be solved.

[0003] Currently, deep learning technology, primarily based on neural networks, has achieved significant results in various natural language processing tasks. By combining deep learning and machine learning techniques with text summarization tasks, specific semantic features of text are learned from large corpora. Similar to text generation tasks, enabling machines to understand text, much like humans think when reading, allows computers to produce fluent short texts that accurately reflect the main idea of ​​the original. For over half a century, scholars worldwide have explored and researched text summarization technology from different perspectives, seeking higher-quality research results. Text summarization systems are classified differently based on their input, output, purpose, length, algorithm, domain, and language. Many other factors can be considered when discussing the classification of summaries. In text summarization tasks, understanding semantic information will be a major breakthrough for automatic text summarization, allowing summarization models to incorporate this method of understanding text into the automatic summarization process, just like human thinking—a approach that aligns with both practical logic and human expectations for this task. Many researchers have conducted various studies on the impact of semantic information in the task of summarizing. Although their thinking and methods differ, they all agree that there is still considerable room for research on semantic information, which can further improve the accuracy and readability of the summary content.

[0004] The real world contains countless conceptual entities with intricate connections. Constructing a knowledge base encompassing all these relationships is not only extremely costly but also presents unforeseen challenges. Among existing external knowledge bases, WordNet primarily focuses on semantic, part-of-speech, and tense relationships between lexical entities, as well as the properties of things; ConceptNet focuses on semantic concepts related to lexical expressions and the inherent information related to natural language concepts. Because different knowledge bases have different construction goals, the appropriate document types for each remain different, meaning that existing technical solutions cannot fully utilize external knowledge bases to improve the quality of generated summary content. Summary of the Invention

[0005] This invention is made to solve the above-mentioned problems, and aims to provide a text summarization method for adaptive topic selection based on multiple knowledge graphs.

[0006] This invention provides a multi-knowledge graph-based adaptive topic-optimized text summarization method for processing documents to be identified using multiple existing external knowledge bases to obtain corresponding summaries. The method includes the following steps: Step S1, preprocessing the document to be identified to obtain preprocessed data; Step S2, vectorizing the preprocessed data into paragraphs to obtain document feature information; Step S3, inputting the preprocessed data and document feature information into a bidirectional long short-term memory network, and then processing the output of the bidirectional long short-term memory network according to an attention mechanism to obtain a topic context representation vector; Step S4, constructing a FastText model and training the FastText model using a training set built from multiple existing documents. Step S5: Train the FastText model to obtain the adapted text classifier; Step S6: Input the preprocessed data into the adapted text classifier to obtain class feature vectors; Step S7: Calculate the class feature vectors and topic context representation vectors according to the attention mechanism function and selection function to obtain document category features; Step S8: Calculate the feature matrices of each external knowledge base according to the knowledge vectorization method to obtain the feature matrices corresponding to each external knowledge base; Step S9: Input all feature matrices, document category features, and topic context representation vectors into the knowledge selector, and then concatenate the output of the knowledge selector with the topic context representation vectors to obtain the concatenation result; Step S10: Input the concatenation result into the decoder to obtain the summary.

[0007] The adaptive topic optimization text summarization method based on multiple knowledge graphs provided by the present invention may also have the following feature: wherein, in step S1, the preprocessing is to process the document to be identified using the Stanford CoreNLP tool.

[0008] The adaptive topic optimization text summarization method based on multiple knowledge graphs provided by this invention may also have the following features: In step S2, the specific process of paragraph vectorization is as follows: constructing a Doc2Vec model, inputting document data into the Doc2Vec model, and using the output of the Doc2Vec model as document feature information.

[0009] The adaptive topic-optimized text summarization method based on multiple knowledge graphs provided in this invention may also have the following features: In step S4, the FastText model includes an input layer, a hidden layer, and an output layer. The hidden layer averages different word vectors, and the word vectors are the output of the input layer. The output of the FastText model is the class label of the input document. The FastText model uses a loss function L during training, and the formula for calculating the loss function L is: In the formula, N is the total number of documents in the training set, and x n Let y be the vector representation of the nth document. n W represents the actual tag of the nth document. A and W B All are weight matrices.

[0010] The adaptive topic optimization text summarization method based on multiple knowledge graphs provided by this invention may also have the following feature: wherein, in step S6, the calculation formula for document category features is: In the formula, σ is the selection function, Attn is the attention mechanism function, and W α W β Both b and are learnable parameters. For class feature vectors, This is a vector representing the topic context.

[0011] The adaptive topic-optimized text summarization method based on multiple knowledge graphs provided in this invention may also have the following feature: wherein the multiple external knowledge bases are WordNet knowledge base and ConceptNet knowledge base, and in step S7, the expression of the feature matrix corresponding to the WordNet knowledge base and ConceptNet knowledge base is: In the formula G c The feature matrix corresponding to the ConceptNet knowledge base is shown below. The triples in the ConceptNet knowledge base are represented as (h... c ,r c ,t c ), h c and t c For the two distinct entities in the triple, r c For the relationship between the two entities in this triple, G wLet h be the feature matrix corresponding to the WordNet knowledge base. The triples in the WordNet knowledge base are represented as (h w ,r w ,t w ), h w and t w For the two distinct entities in the triple, r w This represents the relationship between the two entities in the triple.

[0012] The adaptive topic optimization text summarization method based on multiple knowledge graphs provided by this invention may also have the following feature: wherein, in step S8, when the multiple external knowledge bases are WordNet knowledge base and ConceptNet knowledge base, the expression of the connection result is: In the formula The result is a join operation, where conc represents the join operation and σ represents the knowledge selector. As a context representation vector, For document category features.

[0013] The adaptive topic optimization text summarization method based on multiple knowledge graphs provided by this invention may also have the following feature: In step S9, the decoder generates the output of each time step according to the connection result. The output at time step t is the word with the highest probability in the probability distribution of all words at time step t. The outputs of each time step constitute a summary in the order of the time steps. The expression for the probability distribution of all words at time step t is: In the formula P cw,t Let f be the probability distribution of all words at time step t, f be the softmax function, and W be the probability distribution of all words at time step t. ζ and b ζ All are learnable parameters. For the connection result, h cw,t-1 The hidden state generated by the decoder at time step t-1 based on the input of the current time step, the hidden state generated by the encoder at the first time step based on the connection result, and the encoder uses the hidden state generated at the current time step as the input of the encoder at the next time step, until the final output is the termination token.

[0014] The role and effect of invention

[0015] The adaptive topic-selection text summarization method based on multi-knowledge graphs of this invention improves the quality of the generated summaries by using knowledge vectorization to generate a feature matrix of the most suitable dimension for an external knowledge base and applying it to summarization. Then, a knowledge selector selects the feature matrix that best matches the document to be identified from the feature matrices corresponding to multiple existing external knowledge bases and embeds the knowledge, thus avoiding the problem of semantic similarity rather than the same words. Therefore, the adaptive topic-selection text summarization method based on multi-knowledge graphs of this invention can generate more accurate corresponding summaries based on the document. Attached Figure Description

[0016] Figure 1 This is a schematic diagram illustrating the working principle of the adapted topic preferred text summary in an embodiment of the present invention;

[0017] Figure 2 This is a flowchart illustrating the adaptive topic optimization text summarization method based on multiple knowledge graphs in an embodiment of the present invention;

[0018] Figure 3 This is a schematic diagram of Rouge scores corresponding to feature matrices of different dimensions in an embodiment of the present invention. Detailed Implementation

[0019] To make the technical means, creative features, objectives and effects of this invention easier to understand, the following embodiments, in conjunction with the accompanying drawings, specifically illustrate the method for adapting topic optimization text summarization based on multiple knowledge graphs.

[0020] Figure 1 This is a schematic diagram illustrating the working principle of the adapted topic preferred text summary in an embodiment of the present invention.

[0021] like Figure 1 As shown, the document to be identified is vectorized into paragraphs, and then the results of paragraph vectorization are input into a bidirectional long short-term memory network along with the document to be identified to obtain a topic context representation vector. The document to be identified is then input into an adaptive text classifier to obtain a class feature vector. The class feature vector and the topic context representation vector are then processed by an attention mechanism function and a selection function to obtain document category features. At the same time, feature matrices are calculated for multiple existing external knowledge bases, such as external knowledge base A and external knowledge base B, to obtain the feature matrices corresponding to each external knowledge base. These feature matrices are then input into a knowledge selector along with the document category features and the topic context representation vector. The knowledge selector selects the external knowledge that best matches the topic context representation vector for embedding based on self-attention selection. The output of the knowledge selector is then concatenated with the topic context representation vector to obtain a connection result. Finally, the connection result is input into a decoder to obtain the summary corresponding to the document to be identified.

[0022] Figure 2This is a flowchart illustrating the adaptive topic optimization text summarization method based on multiple knowledge graphs in an embodiment of the present invention.

[0023] like Figure 2 As shown, the adaptive topic optimization text summarization method based on multiple knowledge graphs in this embodiment includes the following steps:

[0024] Step S1: Preprocess the document to be recognized to obtain preprocessed data.

[0025] The preprocessing involved using the Stanford CoreNLP tool to process the documents to be identified.

[0026] Step S2: Vectorize the preprocessed data into paragraphs to obtain document feature information.

[0027] The specific process of paragraph vectorization is as follows:

[0028] Construct a Doc2Vec model, input the document data into the Doc2Vec model, and use the output of the Doc2Vec model as the document feature information.

[0029] Step S3: Input the preprocessed data and document feature information into the bidirectional long short-term memory network, and then process the output of the bidirectional long short-term memory network according to the attention mechanism to obtain the topic context representation vector.

[0030] Step S4: Construct the FastText model and train the FastText model using a training set built from existing documents to obtain the trained FastText model as an adaptive text classifier.

[0031] The FastText model consists of an input layer, a hidden layer, and an output layer. The hidden layer averages different word vectors, and the word vectors are the output of the input layer. The output of the FastText model is the class label of the input document.

[0032] The FastText model uses a loss function L during training. The formula for calculating the loss function L is:

[0033]

[0034] In the formula, N is the total number of documents in the training set, and x n Let y be the vector representation of the nth document. n W represents the actual tag of the nth document. A and W B All are weight matrices.

[0035] Step S5: Input the preprocessed data into the adaptive text classifier to obtain class feature vectors.

[0036] Step S6: Calculate the class feature vector and topic context representation vector based on the attention mechanism function and selection function to obtain the document category features.

[0037] The formula for calculating document category features is as follows:

[0038]

[0039] In the formula, σ is the selection function, Attn is the attention mechanism function, and W α W β Both b and are learnable parameters. For class feature vectors, This is a vector representing the topic context.

[0040] Step S7: Calculate the feature matrix of each external knowledge base according to the knowledge vectorization method to obtain the feature matrix corresponding to each external knowledge base. In this embodiment, the knowledge vectorization method is RotatE.

[0041] Among them, several external knowledge bases are the WordNet knowledge base and the ConceptNet knowledge base.

[0042] The expressions for the feature matrices corresponding to the WordNet and ConceptNet knowledge bases are:

[0043]

[0044]

[0045] In the formula G c The feature matrix corresponding to the ConceptNet knowledge base is shown below. The triples in the ConceptNet knowledge base are represented as (h... c ,r c ,t c ), h c and t c For the two distinct entities in the triple, r c For the relationship between the two entities in this triple, G w Let h be the feature matrix corresponding to the WordNet knowledge base. The triples in the WordNet knowledge base are represented as (h w ,r w ,t w ), h w and t w For the two distinct entities in the triple, r w This represents the relationship between the two entities in the triple.

[0046] Step S8: Input all feature matrices, document category features, and topic context representation vectors into the knowledge selector, and then concatenate the output of the knowledge selector with the topic context representation vectors to obtain the concatenation result.

[0047] When multiple external knowledge bases are WordNet and ConceptNet knowledge bases, the expression for the connection result is:

[0048]

[0049] In the formula The result is a join operation, where conc represents the join operation and σ represents the knowledge selector. As a context representation vector, For document category features.

[0050] In this embodiment, multiple documents to be identified can also be input as a dataset into the model constructed according to the multi-knowledge graph-based adaptive topic optimization text summarization method of the present invention to obtain multiple corresponding summaries. In this process, a self-attention matrix M is established by using the softmax function. ij This allows us to focus on key components, using the self-attention matrix M. ij The expression is:

[0051]

[0052] In the formula This represents the connection result corresponding to the j-th document to be identified.

[0053] Step S9: Input the connection result into the decoder to obtain the summary.

[0054] The decoder generates the output for each time step based on the connection results. The output at time step t is the word with the highest probability among all words in the probability distribution at time step t. The outputs of each time step are arranged in time step order to form a summary.

[0055] The expression for the probability distribution of all words at time step t is:

[0056]

[0057] In the formula P cw,t Let f be the probability distribution of all words at time step t, f be the softmax function, and W be the probability distribution of all words at time step t. ζ and b ζ All are learnable parameters. For the connection result, h cw,t-1The hidden state generated by the decoder at time step t-1 based on the input of the current time step, the hidden state generated by the encoder at the first time step based on the connection result, and the encoder uses the hidden state generated at the current time step as the input of the encoder at the next time step, until the final output is the termination token.

[0058] In this embodiment, the model constructed using the multi-knowledge graph-based adaptive topic optimization text summarization method of the present invention is used as the model of the present invention. The model of the present invention is then compared with related models and methods in the field of text summarization. The comparison models include PGNet, Bottom-up, and BART models based on the Sequence-to-sequence model method; PEGASUS model based on the Pre-training model method; MatchSum, VHTM, and ECGS models based on the Semantic model method; and GSum, ConSum, SeqCo, and GOLD models based on the General model method. The above 12 models are evaluated using Rouge metrics on the existing CNN / Daily Mail dataset. Rouge metrics include Rouge-1, Rouge-2, and Rouge-L metrics. The evaluation scores of Rouge-1, Rouge-2, and Rouge-L metrics for each model are shown in the table below:

[0059] Model Name Rouge-1 indicator Rouge-2 Indicators Rouge-L Indicator PGNet model 36.44 15.66 33.42 Bottom-up model 41.22 18.68 38.34 BART model 44.16 21.28 40.90 PEGASUS model 44.17 21.47 41.11 MatchSum model 44.41 20.86 40.55 VHTM model 40.57 18.05 37.18 ECGS model 44.52 21.95 41.39 GSum model 45.94 22.32 42.48 ConSum model 44.53 21.54 41.57 SeqCo model 45.02 21.80 41.75 GOLD model 45.40 22.01 42.25 This invention model 45.96 22.39 42.66

[0060] The first column of the table above lists the model names, the second column lists the Rouge-1 score, the third column lists the Rouge-2 score, and the fourth column lists the Rouge-L score. For example, the cell in the second row and second column shows that the PGNet model has a Rouge-1 score of 36.44 on the CNN / Daily Mail dataset. As shown in the table, the model of this invention achieves the best scores across all three metrics. Therefore, on the CNN / Daily Mail dataset, the multi-knowledge graph-based adaptive topic selection text summarization method of this invention has the best text summarization generation performance compared to the existing methods and models described above.

[0061] In this embodiment, the model of the present invention was compared with related models and methods in the field of text summarization on the existing XSum dataset. The comparison models included PGNet, BART, and BertSum models based on the sequence-to-sequence model method; PEGASUS model based on the pre-training model method; MatchSum model based on the semantic model method; and GSum, ConSum, SeqCo, GOLD, and SimCLS models based on the general model method. The Rouge evaluation index was also calculated for the above 10 models. The evaluation scores of Rouge-1, Rouge-2, and Rouge-L indices for each model are shown in the table below:

[0062] Model Name Rouge-1 indicator Rouge-2 Indicators Rouge-L Indicator PGNet model 29.70 9.21 23.24 BART model 45.14 22.27 31.27 BertSum model 38.81 16.50 31.27 PEGASUS model 47.21 24.56 39.25 MatchSum model 47.21 24.56 39.25 GSum model 45.40 21.89 36.67 ConSum model 47.34 24.67 39.40 SeqCo model 45.65 22.41 37.04 GOLD model 45.75 22.26 37.30 SimCLS model 47.61 24.57 39.44 This invention model 48.02 25.56 40.41

[0063] The first column of the table above lists the model names, the second column lists the Rouge-1 score, the third column lists the Rouge-2 score, and the fourth column lists the Rouge-L score. For example, the cell in the second row and second column shows that the PGNet model has a Rouge-1 score of 29.70 on the XSum dataset. As shown in the table, the model of this invention achieves the best scores across all three metrics. Therefore, on the XSum dataset, the multi-knowledge graph-based adaptive topic optimization text summarization method of this invention has the best text summarization generation performance compared to the existing methods and models described above.

[0064] Therefore, the topic-optimized text summarization method based on multiple knowledge graphs of the present invention has better summary content quality than existing text summarization generation methods.

[0065] In this embodiment, the feature matrix is ​​applied to the summary generation, i.e., the most matching external knowledge base is introduced, thereby avoiding the problem of semantic similarity rather than the same word and improving the quality of the generated summary. In this process, the different dimensions of the feature matrix obtained by the knowledge vectorization method have different effects on the quality of the summary. Therefore, in this embodiment, feature matrices of different dimensions are trained by the knowledge vectorization method RotatE, and different summaries are generated. The summaries are then analyzed based on the Rouge scores, and the data is specifically processed during visualization: the scores of feature matrices of different dimensions on the same Rouge index are removed by removing the same integer part. For example, if the Rouge-1 before processing is 37.16, 37.38, 37.59 and 37.25, then after processing it is 0.16, 0.38, 0.59 and 0.25.

[0066] Figure 3 This is a schematic diagram of Rouge scores corresponding to feature matrices of different dimensions in an embodiment of the present invention.

[0067] like Figure 3 As shown, the horizontal axis represents the Rouge index, and the vertical axis represents the processed index score. For the Rouge-1 index (ROUGE-1'), the score corresponding to the 100-dimensional feature matrix (100-Dimemsion) is 0.16, the score corresponding to the 200-dimensional feature matrix (200-Dimemsion) is 0.38, the score corresponding to the 300-dimensional feature matrix (300-Dimemsion) is 0.59, and the score corresponding to the 400-dimensional feature matrix (400-Dimemsion) is 0.25. For the Rouge-2 index... On the ROUGE-2' scale, the score for a 100-dimensional feature matrix is ​​0.13, for a 200-dimensional feature matrix it is 0.35, for a 300-dimensional feature matrix it is 0.57, and for a 400-dimensional feature matrix it is 0.38. On the ROUGE-L' scale, the score for a 100-dimensional feature matrix is ​​0.05, for a 200-dimensional feature matrix it is 0.26, for a 300-dimensional feature matrix it is 0.17, and for a 400-dimensional feature matrix it is 0.48. Therefore, among the four dimensions mentioned above, the 300-dimensional feature matrix generates the best summary quality.

[0068] The role and effect of the embodiments

[0069] According to the multi-knowledge graph-based adaptive topic-optimized text summarization method involved in this embodiment, the most suitable dimension feature matrix of the external knowledge base is generated using knowledge vectorization and applied to the summary generation. Then, a knowledge selector selects the feature matrix that best matches the document to be identified from the feature matrices corresponding to multiple existing external knowledge bases and performs knowledge embedding, thereby avoiding the problem of semantic similarity rather than the same words, thus improving the quality of the generated summary. In summary, this method can generate more accurate corresponding summaries based on documents.

[0070] The above embodiments are preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention.

Claims

1. A text summarization method based on multiple knowledge graphs for adapting topics and obtaining corresponding summaries from documents to be identified by processing multiple existing external knowledge bases, characterized in that... Includes the following steps: Step S1: Preprocess the document to be identified to obtain preprocessed data; Step S2: The preprocessed data is vectorized into paragraphs to obtain document feature information; Step S3: Input the preprocessed data and the document feature information into the bidirectional long short-term memory network, and then process the output of the bidirectional long short-term memory network according to the attention mechanism to obtain the topic context representation vector; Step S4: Construct the FastText model and train the FastText model using a training set built from multiple existing documents to obtain the trained FastText model as an adaptive text classifier. Step S5: Input the preprocessed data into the adaptive text classifier to obtain class feature vectors; Step S6: Calculate the class feature vector and the topic context representation vector based on the attention mechanism function and the selection function to obtain the document category features; Step S7: Calculate the feature matrix of each of the external knowledge bases according to the knowledge vectorization method to obtain the feature matrix corresponding to each of the external knowledge bases. Step S8: Input all the feature matrices, the document category features, and the topic context representation vector into the knowledge selector, and then concatenate the output of the knowledge selector with the topic context representation vector to obtain the connection result; Step S9: Input the connection result into the decoder to obtain the summary. In step S9, the decoder generates the output for each time step based on the connection result. The output at time step t is the word with the highest probability among all words in the probability distribution at time step t. The outputs at each time step are arranged in time step order to form the summary. The expression for the probability distribution of all words at time step t is: In the formula P cw,t Let f be the probability distribution of all words at time step t, f be the softmax function, and W be the probability distribution of all words at time step t. ζ and b ζ All are learnable parameters. For the connection result, h cw,t-1 The hidden state generated by the decoder at time step t-1 based on the input of the current time step, the hidden state generated by the encoder at the first time step based on the connection result, and the encoder uses the hidden state generated at the current time step as the input of the encoder at the next time step, until the final output of the termination token.

2. The text summarization method for adaptive topic selection based on multiple knowledge graphs according to claim 1, characterized in that: in, In step S1, the preprocessing involves processing the document to be identified using the Stanford CoreNLP tool.

3. The text summarization method for adaptive topic selection based on multiple knowledge graphs according to claim 1, characterized in that: in, In step S2, the specific process of paragraph vectorization is as follows: A Doc2Vec model is constructed, the document data is then input into the Doc2Vec model, and the output of the Doc2Vec model is used as the document feature information.

4. The text summarization method for adaptive topic selection based on multiple knowledge graphs according to claim 1, characterized in that: in, In step S4, the FastText model includes an input layer, a hidden layer, and an output layer. The hidden layer averages different word vectors. The word vectors are the output of the input layer. The output of the FastText model is the class label of the input document. The FastText model uses a loss function L during training, and the formula for calculating the loss function L is as follows: In the formula, N is the total number of documents in the training set, and x n Let y be the vector representation of the nth document. n W represents the actual tag of the nth document. A and W B All are weight matrices.

5. The text summarization method for adaptive topic selection based on multiple knowledge graphs according to claim 1, characterized in that: in, In step S6, the formula for calculating the document category feature is: In the formula, σ is the selection function, Attn is the attention mechanism function, and W α W β Both b and are learnable parameters. For the class feature vector, This is the topic context representation vector.

6. The text summarization method for adaptive topic selection based on multiple knowledge graphs according to claim 1, characterized in that: in, The multiple external knowledge bases mentioned are the WordNet knowledge base and the ConceptNet knowledge base. In step S7, the expressions for the feature matrices corresponding to the WordNet knowledge base and the ConceptNet knowledge base are: In the formula G c The feature matrix corresponding to the ConceptNet knowledge base is represented as (h) in the ConceptNet knowledge base. c ,r c ,t c ), h c and t c For the two distinct entities in the triple, r c For the relationship between the two entities in this triple, G w Let h be the feature matrix corresponding to the WordNet knowledge base. The triples in the WordNet knowledge base are represented as (h w ,r w ,t w ), h w and t w For the two distinct entities in the triple, r w This represents the relationship between the two entities in the triple.

7. The text summarization method for adaptive topic selection based on multiple knowledge graphs according to claim 6, characterized in that: in, In step S8, when the multiple external knowledge bases are WordNet and ConceptNet, the expression for the connection result is: In the formula The result is a join operation, where conc represents the join operation and σ represents the knowledge selector. The context representation vector, These are the document category features.

Images