Multi-document summarization extraction method and system based on capsule-biGRU network and event automatic classification

Abstract

The application discloses a multi-document summarization extraction method and system based on a Capsule-BiGRU network and event automatic classification, and belongs to the technical field of natural language processing. Firstly, text keywords are extracted, and preliminary clustering is completed according to the extraction result; a local feature matrix of the text is extracted by using a capsule network, a global feature matrix of the text is extracted by using a bidirectional gated recurrent unit network, text similarity fusion analysis is carried out according to the extracted local feature matrix and global feature matrix, and a multi-level similarity vector of the text is obtained; then, text similarity is determined, and accurate clustering is carried out according to the text similarity determination result; finally, the calculation of a minimal dominating set is carried out on each type of document in the text clustering result, the theme and the semantics are fused, and a multi-document summarization extraction result is obtained. The method fuses multi-layer clustering, global features and local features to improve the accuracy of multi-document summarization extraction, and can solve the problem that accurate clustering is difficult in the current multi-document summarization extraction process.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of natural language processing, and particularly relates to a multi-document summarization extraction method and system based on a Capsule-BiGRU network and event automatic classification. BACKGROUND

[0002] In the current era of massive information, there are a large number of redundant and repetitive texts in the description of a certain event or a certain topic. For the public, this is not conducive to timely understanding of the truth, and may cause a certain degree of public opinion fermentation. For analysts, a large amount of repetitive information is not conducive to ensuring the timeliness of the analysis results. Therefore, timely classification and extraction of summaries of multiple documents can enable relevant departments to understand the content of the results found in the shortest time and respond and process in time, so as to minimize the impact of the event that has occurred. At the same time, for some professional domain texts, there may be professional terms that cannot be recognized or filtered, which is not conducive to the integrity of the summary content.

[0003] As a key part of understanding the core content of an article, the summary should have the important role of highlighting the key points, providing abundant information, and feeding back the most important part to the reader. With the rapid development of natural language processing technology, as one of the most important branches of "text summarization", the results have been continuously improved with the improvement of the technical level. At present, text summarization mainly has two kinds of generative summary and extractive summary. The extractive document summary directly extracts complete sentences from the original text to form a summary, so as to ensure the description of the summary and the grammatical correctness. Traditional text summary extraction methods such as Textrank based on graph model have achieved good results in generalization and ease of use. With the continuous development of deep learning, more and more researchers have begun to use deep neural network-based methods for text summary extraction. Earlier extractive text summary algorithms using deep neural networks use recurrent neural network-based models for summary extraction, or use memory networks to help the model better understand the content of the article. With the better performance of Transformer in various tasks, researchers have explored the use of Transformer in extractive text summarization tasks and found that it can improve the quality of extracted summaries. Today, pre-trained models such as Bert have gradually become the mainstream of research, and by adjusting the downstream training and parameters, they can better complete multiple tasks including text summary extraction.

[0004] Multi-document summarization is to extract the summary by extracting and compressing the content of multiple documents under the same topic or different topics. Due to the huge number of media and the large number of users of social platforms at present, there may be a large number of similar or even overlapping contents for a certain topic or event, which will bring a lot of repetitive work to the auditors or text analysis workers, increase their task quantity, and thus a method is needed to realize accurate clustering of multi-document summary, reduce the workload of auditors or text analysis workers, and further improve the efficiency. SUMMARY

[0005] In order to overcome the above-mentioned defects of the prior art, the purpose of the present application is to provide a multi-document summary extraction method and system based on Capsule-BiGRU network and event automatic classification, which solves the problem that multi-document summary extraction is difficult to accurately cluster.

[0006] In order to achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0007] The present application discloses a multi-document summary extraction method based on Capsule-BiGRU network and event automatic classification, comprising the following steps:

[0008] S1, extract the text keywords of each document for all the current texts to be classified, complete the preliminary clustering according to the extraction results of the keywords, and obtain the primary classification results of all the documents;

[0009] S2, based on the primary classification results of all the documents obtained after clustering in step S1, extract the local features and global features of each document in each clustering result by using the capsule network and the bidirectional gated recurrent unit network respectively, and obtain the feature matrix of each document fused with global and local features;

[0010] S3, perform text similarity fusion analysis according to the feature matrix extracted in step S2, and obtain a multi-level similarity vector of each document;

[0011] S4, determine the similarity of each text in each major category based on the multi-level similarity vector of the text obtained in step S3, and perform secondary accurate clustering based on the preliminary clustering according to the text similarity determination result;

[0012] S5, perform minimum dominating set calculation on each category of text in the accurate clustering result obtained in step S4, fuse the theme and the semantics, and obtain the multi-document summary extraction result.

[0013] Preferably, in step S1, the text keywords are extracted by fusing the field adaptation of the transfer learning field and the conditional random field with the bidirectional long short-term memory network field.

[0014] Preferably, the specific steps of extracting the text keywords of each document are: collecting articles and keywords of the relevant field in a specific network as target domain data, and collecting the unsupervised data text set to be extracted as the source domain data; then performing word vector encoding, minimizing the cross entropy of the similar features and keywords classification between the target domain data and the source domain data, and completing the keyword extraction task of the text set.

[0015] Preferably, in step S1, the standard of preliminary clustering is: extracting the keywords of each document, the number of keywords of each document ranges from 3 to 7, and the coincidence degree of keywords between each document is greater than or equal to 0.8, then the documents are automatically clustered.

[0016] Preferably, in step S2, the specific steps of extracting the local feature matrix are: inputting the word vector matrix into the capsule network for convolution operation, performing capsule convolution operation through the main capsule layer, and outputting the main capsule layer after the squeezing function operation, and connecting to the classification capsule layer after the dynamic routing protocol mechanism operation, and the output result of the classification capsule layer is unfolded as the local feature vector of the text.

[0017] Preferably, in step S2, the specific steps of extracting the global feature matrix are: inputting the word vector matrix into the BiGRU network, using the bidirectional GRU network to extract the information of the text from two directions, and obtaining the global feature vector of the text.

[0018] Preferably, in step S4, the specific steps of similarity determination are: respectively performing similarity analysis on the local features and global features of the texts to obtain a similarity matrix, inputting the similarity matrix into a full connection network, and finally determining whether the texts are similar through a classifier.

[0019] Preferably, in step S4, if the determination result of the similar documents does not reach the similarity threshold value greater than or equal to 0.7, the precise clustering is performed; if the similarity threshold value greater than or equal to 0.7 is reached, it is considered that the accuracy of the preliminary clustering result meets the requirements, and the result can be directly used as the precise clustering result and enters step S5.

[0020] Preferably, step S5 is: taking the clustered document set as a node, obtaining a sentence graph model through a Bayesian topic model and a word vector model, fusing the sentence topic probability distribution and the sentence semantic similarity to obtain the final relevance of the sentence, combining the topic information and the semantic information as the edge weight of the sentence graph model, and realizing the multi-document abstract extraction by means of the minimum dominating set method of the sentence graph model.

[0021] The application also discloses a multi-document abstract extraction system based on Capsule-BiGRU network and event automatic classification, which comprises:

[0022] A keyword extraction module is configured to extract text keywords of each document, and complete preliminary clustering according to the extraction result of the keywords;

[0023] A matrix extraction module is configured to extract a local feature matrix of the text by using a capsule network and extract a global feature matrix of the text by using a bidirectional gated recurrent unit network based on the preliminary classification result of all the documents obtained after clustering by the keyword extraction module;

[0024] A fusion analysis module is configured to perform text similarity fusion analysis according to the local feature matrix and the global feature matrix extracted by the matrix extraction module, and obtain a multi-level similarity vector of the text;

[0025] A determination module is configured to determine the similarity of the text according to the multi-level similarity vector of the text obtained by the fusion analysis module, and perform accurate clustering according to the text similarity determination result;

[0026] An abstract extraction module is configured to perform minimum dominating set calculation on each class of documents in the text clustering result obtained by the determination module, fuse the theme and the semantics, and obtain a multi-document abstract extraction result.

[0027] Compared with the prior art, the present application has the following beneficial effects:

[0028] The multi-document abstract extraction method based on the Capsule-BiGRU network and the event automatic classification provided by the present application has the following advantages: 1. The event classification step is divided into two steps, the more accurate text similarity calculation is performed on the result of the preliminary clustering after the keyword extraction, to realize accurate clustering of the documents and improve the accuracy. 2. The global feature and the local feature are combined to calculate the text similarity, the multi-level similarity vector is obtained by fusing the feature matrix, the text similarity deviation caused by only considering the semantic calculation is avoided to the greatest extent, the multi-dimensional semantic calculation is realized, and the accuracy is improved. The method fuses the multi-level clustering, the global feature and the local feature, realizes the maximum degree of document clustering precision and abstract extraction perfection by means of secondary clustering and combination of the theme and the semantics, and the obtained result can accurately describe the core theme of each class of documents in all the current documents. The method solves the problem that the multi-document abstract is difficult to cover the key content or describe the redundancy under the condition of a large number of documents, realizes the efficiency and accuracy of massive text analysis, and improves the accuracy of the multi-document abstract extraction, thereby solving the problem of difficult accurate clustering in the current multi-document abstract extraction process.

[0029] Further, the field adaptation and conditional random field and bidirectional long short-term memory network in the field of transfer learning are used to extract the text keywords. The conditional random field combines the bidirectional long short-term memory network model to convert the keyword recognition task into a sequence labeling task, the bidirectional long short-term memory network efficiently uses the context information, and the conditional random field layer uses the label information, and the transfer learning can realize the cross-domain migration of some proper nouns. Such setting can consider professional knowledge in the process of document keyword extraction, and avoid that some keywords are not extracted due to low frequency.

[0030] Further, the articles and keywords related to the field in the specific network such as the HowNet are collected as target domain data, and the unsupervised data multi-document to be extracted as summary is used as source domain data, then the word vector encoding is performed by using the Bert, the similar features between the target domain data and the source domain data are minimized, and the cross entropy of keyword classification is completed. Such setting realizes the accurate document keyword extraction involving professional fields on the basis of the conditional random field combining the bidirectional long short-term memory network model and the transfer learning.

[0031] Further, the preliminary clustering is completed according to the keyword extraction result, which is the first clustering result of all documents, and the keyword clustering can ensure that the theme described by each type of document has a certain correlation or belongs to a theme, thereby reducing the calculation complexity required by the subsequent accurate classification.

[0032] Further, the GloVe model is used to map the document into a 300-dimensional word vector matrix as the input of the model, and the attention mechanism is used to assign weights, and the result is input into the capsule network (Capsule) to effectively extract the local feature vector of the text.

[0033] Further, the GloVe model is used to map the document into a 300-dimensional word vector matrix input into the BiGRU network, and the bidirectional GRU network is used to extract the information of the text from two directions to obtain the global feature vector of the text.

[0034] Further, the twin neural network structure is used to encode the global feature vector matrix and the local feature vector matrix into the same vector space, the local features and the global features of the text are respectively analyzed for similarity, a similarity matrix is obtained, the similarity matrix is used as the input of the full connection network, and the sigmoid function is used as the classifier in the last layer of the full connection network to judge whether the texts are similar. The combination of the local features and the global features of the text can more comprehensively measure the similarity between the texts, and the misjudgment caused by the similar single index due to the similar description, language, or grammar in Chinese can be avoided as much as possible.

[0035] Further, on the basis of the completion of the initial clustering, the text similarity of different documents in the same clustering result is analyzed, and if the similarity threshold of similar documents is 0.7 or higher, it is considered that the initial clustering result has reached precision, that is, the process of precise clustering has been completed, and the subsequent summary extraction operation can be directly performed on the clustering result; if the threshold is not reached, the similarity of each document in the category with other documents is numerically determined, and the secondary precise clustering is performed on the basis of the initial clustering according to the text similarity result. The precise clustering is specific to the content of each document, so the clustering result is more accurate, and the documents belonging to the same precise clustering achieve the accuracy rate in the process of extracting the multi-document summary.

[0036] Further, each sentence in each precise clustering result is taken as a node to construct a sentence graph model, the sentence topic probability distribution based on the Bayesian topic model and the sentence semantic similarity based on the word vector model are fused to obtain the final relevance of each sentence, the edge weight is calculated in combination with the topic information and the semantic information, the sentence is extracted by using the minimum dominating set principle, and the multi-document summary is obtained.

[0037] It can be understood that the beneficial effects of the above-mentioned second aspect can be referred to the related description in the above-mentioned first aspect, which will not be repeated here.

[0038] In summary, the application realizes precise clustering on document clustering by combining topic information and semantic information, and realizes multi-level text similarity measurement calculation by combining local features and global features, finally realizes high-accuracy document clustering and multi-document summary operation, greatly improves the efficiency of text analysis operation, and reduces the labor cost. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 is a flow chart of a multi-document summary extraction method based on Capsule-BiGRU network and event automatic classification;

[0040] Figure 2 is a field self-adaptive unsupervised text keyword extraction model diagram based on transfer learning;

[0041] Figure 3 is a text similarity analysis diagram based on Capsule-BiGRU network;

[0042] Figure 4 is a multi-document summary extraction diagram based on Capsule-BiGRU network. DETAILED DESCRIPTION

[0043] In the following, the technical solutions in the embodiments of the present application will be described clearly and completely in conjunction with the accompanying drawings of the embodiments of the present application, so that those skilled in the art can better understand the technical solutions of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor should fall within the scope of protection of the present application.

[0044] It should be noted that the terms "first", "second", and the like in the specification and claims of the present application and the above-described drawings are used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0045] The present application will be described in further detail below in conjunction with the accompanying drawings:

[0046] As shown in Figure 1 The multi-document summarization extraction method proposed by the present application is composed of four modules of keyword extraction, text similarity analysis, text clustering and summary extraction. The specific implementation method is as follows:

[0047] S1, keyword extraction

[0048] As shown in Figure 2 The domain text keyword extraction method that fuses the self-adaptation and conditional random field in the field of transfer learning and the bidirectional long short-term memory network (BiLSTM) can well solve the realistic dilemma of strong knowledge of domain text data, more professional vocabulary and unsupervised. The method is based on the shared similar features and classification target loss degree of the labeled target domain and unlabeled source domain to build a calculation model, which can realize the transfer learning of the unsupervised domain text keyword extraction method and complete the keyword extraction task of the domain text.

[0049] S101, target domain data acquisition

[0050] The main task of this part is to collect labeled data in the adjacent field of the text to be extracted (such as articles and keywords in the relevant field of HowNet), construct a target domain data set, perform keyword sequence labeling processing on the data set, apply a word vector encoding model, represent the target domain text in word embedding vectors based on the input text, and then input the BiLSTM model to obtain the feature sequence representation of the target domain.

[0051] S102, source domain data processing

[0052] This part is to regard the domain text to be extracted as a source domain data set, apply the same word vector encoding model as the target domain, represent the source domain text in word embedding vectors based on the input text, and then use the BiLSTM model to obtain the feature sequence representation.

[0053] S103, text keyword extraction

[0054] The transfer learning module is based on the domain adaptation and BiLSTM-CRF model in transfer learning, and constructs a domain text keyword extraction model. The feature sequence representation output by the target domain and source domain modules is used for keyword classification model training. In model training, the sum of shared similar features and keyword classification target loss is used as the overall target for parameter optimization, and the gradient descent is used to solve the parameters to obtain the final model. Finally, the keyword label of the domain text to be extracted is obtained, and the domain text keyword is obtained.

[0055] S2, preliminary clustering

[0056] The number of extracted keywords and the coincidence threshold are determined. The number of keywords extracted from each document ranges from 3 to 7, and by default, 5 keywords are extracted and the keyword coincidence degree is required to reach 0.8 or above. This standard is used as the clustering standard for preliminary clustering.

[0057] S3, text similarity analysis

[0058] As shown in Figure 3 The capsule-BiGRU-based text similarity analysis method extracts the local feature matrix of the text by the capsule network and the global feature matrix of the text by the BiGRU network, respectively, and performs similarity analysis to obtain the similarity matrix of the text. The similarity matrix is fused to obtain a multi-level similarity vector of the two texts, so as to determine the text similarity.

[0059] S201, text word vector matrix input

[0060] The pre-trained GloVe model is used to map the text into a 300-dimensional word vector matrix, the word vector matrix is used as the input of the model, and the attention mechanism is used to assign weights.

[0061] S202, text local feature matrix extraction

[0062] The word vector matrix is input into a capsule network for convolution operation, and the capsule convolution operation is performed on the main capsule layer. After the squeezing function operation, the output of the main capsule layer is obtained. After the operation of the dynamic routing protocol mechanism, it is connected to the classification capsule layer. The output result of the classification capsule layer is unfolded as the local feature vector of the text.

[0063] S203, text global feature matrix extraction

[0064] The word vector matrix is input into the BiGRU network, and the information of the text is extracted from two directions using the bidirectional GRU network to obtain the global feature vector of the text.

[0065] S204, similarity fusion analysis

[0066] In the feature vector extraction stage, a twin neural network structure is used, that is, the same network structure is used to process two word vector matrices, so that the two word vector matrices are encoded into the same vector space. Finally, the local features and global features of the texts are analyzed for similarity respectively, and a similarity matrix is obtained. The similarity matrix is input into the full connection network, and the sigmoid function is used as the classifier in the last layer of the full connection network to determine whether the texts are similar.

[0067] S4, accurate clustering

[0068] On the basis of preliminary clustering, text similarity analysis is performed on different documents in the same clustering result. The similarity threshold of similar documents is 0.7 and above by default. If the standard is not met, more detailed clustering is performed.

[0069] S5, multi-document summary extraction

[0070] As shown in Figure 4 The clustered document set is taken as a node by sentence, and the sentence graph is obtained by using related technologies such as the Bayesian topic model and the word vector model. The minimum dominating set of the sentence graph is calculated by using the sentence graph model based on the TF-IDF algorithm, so as to realize the extraction of the multi-document summary. By fusing the multi-information sentence graph model, the theme information, semantic information and relationship information between sentences are combined, which can effectively improve the comprehensive performance of the extracted summary.

[0071] S501, constructing a sentence graph model

[0072] Each sentence in multiple documents in each clustering result is taken as a node to construct a sentence graph model.

[0073] S502, sentence correlation calculation

[0074] The sentence topic probability distribution and the sentence semantic similarity obtained by the sentence-based Bayesian topic model and the word vector model are fused to obtain the final relevance of the sentence;

[0075] S503, sentence graph model weight calculation

[0076] The theme information and the semantic information are combined as the edge weight of the sentence graph model;

[0077] S504, multi-document summarization extraction

[0078] The multi-document summarization is realized by means of the summarization method of the sentence graph minimum domination set.

[0079] In one embodiment, the application provides a multi-document summarization extraction system based on Capsule-BiGRU network and event automatic classification, which can be used to realize the multi-document summarization extraction method based on Capsule-BiGRU network and event automatic classification, and specifically, the multi-document summarization extraction system based on Capsule-BiGRU network and event automatic classification comprises a keyword extraction module, a matrix extraction module, a fusion analysis module, a judgment module and a summarization extraction module.

[0080] The keyword extraction module is used to extract text keywords and complete preliminary clustering according to the extraction result of the keywords.

[0081] The matrix extraction module is used to extract the local feature matrix of the text by using the capsule network based on the initial classification result of all documents obtained after clustering by the keyword extraction module, and extract the global feature matrix of the text by using the bidirectional gated recurrent unit network.

[0082] The fusion analysis module is used to perform text similarity fusion analysis according to the local feature matrix and the global feature matrix extracted by the matrix extraction module, to obtain a multi-level similarity vector of the text.

[0083] The judgment module is used to judge the similarity of the text according to the multi-level similarity vector of the text obtained by the fusion analysis module, and to perform accurate clustering according to the text similarity judgment result.

[0084] The summarization extraction module is used to calculate the minimum domination set of each class of documents in the text clustering result obtained by the judgment module, fuse the theme and the semantic, and obtain the multi-document summarization extraction result.

[0085] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below. Obviously, the described embodiments are some of the embodiments of the present application but not all the embodiments. The components of the embodiments of the present application described herein and shown in the drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided herein is not intended to limit the scope of the claimed present application, but merely represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the scope of the present application.

[0086] Specific application examples:

[0087] 300 long microblogs are collected manually on Sina microblog as a total data set, the discussion content of which contains hot news events, covering 10 topics, each topic containing 30 documents. The selected corpus is representative and distinctive, including various types such as political news, social news, international news, emergency events, hot issues, etc. Among them, 60 are used as a validation set and 30 are used as a test set. The text abstract of the test set is extracted manually. The extraction method is: three people are selected to extract the abstract and the extraction process is independent of each other.

[0088] Manual abstract extraction has strong subjectivity. The differences in text understanding, knowledge background and subjective judgment of the extraction personnel will produce different extraction results. If the difference is too large, it means that the test set has certain controversial nature, so the credibility of the extracted abstract is low. If the difference is small, it means that the controversial nature of the test set is small, and then it means that the credibility is high. ROUGE is the n-gram co-occurrence times between the generated abstract and the existing abstract, which is the most representative and persuasive evaluation index in the problem of text abstract extraction and generation. The higher the ROUGE value, the better the quality of the generated abstract. In the evaluation index, ROUGE (Recall Oriented Understudy for Gisting Evaluation) is used as a measurement index to calculate the extraction result scores of the three extraction abstract personnel and the system respectively. The results are shown in Table 1:

[0089] Table 1 Data results obtained by using ROUGE as a measurement index

[0090]

[0091] Among them, ROUGE-1, ROUGE-2 and ROUGE-3 represent the ROUGE values of the three extracted abstracts, and ROUGE-S represents the result value of the system abstract.

[0092] At the same time, in view of the consideration of the theme information and the semantic features in the abstract extraction process of the application, the abstract extraction under the condition of only considering the theme information and only considering the semantic features is compared respectively, the ROUGE (Recall Oriented Understudy for Gisting Evaluation) index is used to measure the extraction result, and the result is shown in Table 2.

[0093] Table 2 Data results obtained by using ROUGE as a measurement index

[0094]

[0095] Among them, ROUGE-LDA and ROUGE-sem represent the ROUGE values of the abstracts extracted by only considering the clustering text theme and only considering the clustering text semantic information, and ROUGE-S represents the result value of the system abstract. It can be seen that the experimental results of the method considering the theme information and the semantic features are better than the single consideration.

[0096] In summary, the multi-document abstract extraction method and system based on the Capsule-BiGRU network and the event automatic classification of the application fuse the theme information and the semantic features on the basis of completing the accurate clustering of the text, realize the accurate extraction of the abstract under the condition of the multi-document, and the abstract extraction effect realized by the application is obviously better than the artificial extraction and the automatic extraction mode considering the semantic features or the theme information, so that the accuracy and the degree of simplification of analyzing a large amount of text are maximized.

[0097] The above content only illustrates the technical idea of the application, and cannot limit the protection scope of the application, and any modification made according to the technical idea of the application on the basis of the technical scheme falls within the protection scope of the claims of the application.

Claims

1. A multi-document summarization extraction method based on Capsule-BiGRU network and automatic event classification, characterized in that, Includes the following steps: S1. Extract text keywords from each document of all texts to be classified, and perform preliminary clustering based on the keyword extraction results to obtain the initial classification results of all documents; S2. Based on the initial classification results of all documents obtained after clustering in step S1, local and global features of the text are extracted for each document in each clustering result using capsule network and bidirectional gated recurrent unit network, respectively, to obtain a feature matrix of each document that integrates global and local features. S3. Based on the feature matrix extracted in step S2, perform text similarity fusion analysis to obtain a multi-level similarity vector for each document. S4. Based on the text multi-level similarity vector obtained in step S3, determine the similarity of each text in each major category. Based on the text similarity determination results, perform secondary precise clustering on the basis of the preliminary clustering. S5. Calculate the minimum dominant set for each class of documents in the precise text clustering results obtained in step S4, and integrate the topic and semantics to obtain the multi-document summary extraction results. The calculation of the minimum dominant set for each class of documents in the precise text clustering results obtained in step S4 involves taking the clustered document set as sentences and using a Bayesian topic model and a word vector model to obtain a sentence graph model. The sentence topic probability distribution and sentence semantic similarity are then fused to obtain the final relevance of the sentences. The method of fusing topic and semantic information to obtain multi-document summary extraction results combines topic information and semantic information as edge weights of a sentence graph model, and uses the minimum dominance set method of the sentence graph model to achieve multi-document summary extraction.

2. The multi-document digest extraction method based on Capsule-BiGRU network and automatic event classification according to claim 1, characterized in that, In step S1, the domain of the fusion transfer learning domain adaptation and conditional random field and bidirectional long short-term memory network is used to extract text keywords.

3. The multi-document digest extraction method based on Capsule-BiGRU network and automatic event classification according to claim 1, characterized in that, The specific steps for extracting text keywords from each document are as follows: collect articles and keywords from relevant fields in a specific network as target domain data, and use the unsupervised text set from which the abstract is to be extracted as source domain data; then perform word vector encoding, and complete the keyword extraction task of the text set by minimizing the similarity features and keyword classification cross-entropy between the target domain data and the source domain data.

4. The multi-document digest extraction method based on Capsule-BiGRU network and automatic event classification according to claim 1, characterized in that, In step S1, the initial clustering criteria are as follows: extract keywords from each document, with the number of keywords in each document ranging from 3 to 7, and automatically cluster if the overlap of keywords between documents is ≥0.

8.

5. The multi-document digest extraction method based on Capsule-BiGRU network and automatic event classification according to claim 1, characterized in that, In step S2, the specific steps for extracting the local feature matrix are as follows: input the word vector matrix into the capsule network for convolution operation, perform capsule convolution operation through the main capsule layer, and use the result of the squeezing function operation as the output of the main capsule layer. After the operation through the dynamic routing protocol mechanism, it is connected to the classification capsule layer. The output result of the classification capsule layer is expanded into the local feature vector of the text.

6. The multi-document digest extraction method based on Capsule-BiGRU network and automatic event classification according to claim 1, characterized in that, In step S2, the specific steps for extracting the global feature matrix are as follows: input the word vector matrix into the BiGRU network, and use the bidirectional GRU network to extract the information of the text from two directions to obtain the global feature vector of the text.

7. The multi-document digest extraction method based on Capsule-BiGRU network and automatic event classification according to claim 1, characterized in that, In step S4, the specific steps for similarity determination are as follows: perform similarity analysis on the local and global features of each text to obtain a similarity matrix, use the similarity matrix as input to the fully connected network, and use the classifier in the last layer of the fully connected network to determine whether the texts are similar.

8. The multi-document digest extraction method based on Capsule-BiGRU network and automatic event classification according to claim 1, characterized in that, In step S4, if the similarity of documents does not reach the similarity threshold of ≥0.7, precise clustering is performed; if the similarity threshold of ≥0.7 is reached, it is assumed that the accuracy of the preliminary clustering result has met the requirements, and the result can be used as the precise clustering result to directly enter step S5.

9. A multi-document summarization extraction system based on Capsule-BiGRU network and automatic event classification, characterized in that, The multi-document summarization extraction system based on Capsule-BiGRU network and automatic event classification is used to implement the multi-document summarization extraction method based on Capsule-BiGRU network and automatic event classification as described in any one of claims 1 to 8, including: The keyword extraction module is used to extract text keywords from each document and perform preliminary clustering based on the keyword extraction results; The matrix extraction module is used to perform the initial classification of all documents obtained by clustering after the keyword extraction module. It uses a capsule network to extract the local feature matrix of the text and a bidirectional gated recurrent unit network to extract the global feature matrix of the text. The fusion analysis module is used to perform text similarity fusion analysis based on the local feature matrix and global feature matrix extracted by the matrix extraction module, and obtain a multi-level similarity vector of the text. The judgment module is used to determine the similarity of texts by fusing the multi-level similarity vectors obtained from the analysis module, and to perform accurate clustering based on the text similarity judgment results; The abstract extraction module is used to calculate the minimum dominance set for each document in the text clustering results obtained by the judgment module, and to integrate the topic and semantics to obtain multi-document abstract extraction results.

Images