Abstract extraction method, system, device, and storage medium

Abstract

The application discloses an abstract extraction method, system, device and storage medium, which comprises the following steps: constructing a text relation network, and dividing the text relation network into multiple sub-topic communities, wherein the text relation network is composed of multiple sentence nodes and connecting edges between the sentence nodes; obtaining a scoring result of each sentence according to a sentence weight of each sentence and a contribution value of a word in each sentence to the sentence in each sub-topic community; selecting a target sentence from the text based on the scoring result, and generating a text abstract according to the target sentence. The accuracy of the abstract extraction result is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and particularly relates to a summary extraction method, system, device and storage medium. BACKGROUND

[0002] In recent years, with the continuous expansion of business and the increase in the number of users, the number of customer complaints received by enterprises also increases, and customer complaint problems gradually attract attention. To quickly and accurately handle user complaints, the focus is on timely finding hot issues from a large number of complaints, so that the relevant departments can develop solutions as soon as possible to cope with them, and avoid the adverse effects caused by untimely problem handling. Among them, the hot issues include ordinary hot issues and sudden hot issues. Compared with ordinary hot issues, sudden hot issues are generally not easy to find, and the detection of sudden hot issues is crucial in actual application scenarios.

[0003] In related technologies, a text summary extraction algorithm is used to extract a text summary. However, when the text summary extraction algorithm extracts a text summary, only the importance of a sentence is considered, and the mutual association of words and sentences is ignored, resulting in an inaccurate summary extraction result. SUMMARY

[0004] Embodiments of the present application provide a summary extraction method, system, device and storage medium, aiming to improve the accuracy of the summary extraction result.

[0005] The summary extraction method provided by the embodiments of the present application comprises:

[0006] constructing a text relationship network, and dividing the text relationship network into a plurality of sub-topic communities, wherein the text relationship network is composed of a plurality of sentence nodes and connection edges between the sentence nodes;

[0007] obtaining a scoring result of each sentence according to a sentence weight of each sentence in each sub-topic community and a contribution value of a word in each sentence to the sentence;

[0008] selecting a target sentence from the text based on the scoring result, and generating a text summary according to the target sentence.

[0009] Optionally, the step of obtaining a scoring result of each sentence according to a sentence weight of each sentence in each sub-topic community and a contribution value of a word in each sentence to the sentence comprises:

[0010] obtaining a sentence weight of each sentence after last iteration, a damping coefficient, a number of nodes of the text relationship network and a weight of the connection edges between the sentence nodes in each sub-topic community;

[0011] determining a sentence weight of each sentence after the current iteration according to the sentence weight of each sentence after the last iteration, the damping coefficient, a number of nodes of the text relation network and a weight of a connecting edge between the sentence nodes;

[0012] obtaining a complex network comprehensive feature value of each word in each sentence;

[0013] obtaining a contribution value of a word in each sentence to the sentence according to the sentence weight of each sentence after the current iteration and the complex network comprehensive feature value of each word in each sentence;

[0014] fusing the sentence weight of each sentence after the current iteration and the contribution value of the word in each sentence to the sentence to obtain a scoring result of each sentence.

[0015] Optionally, the step of selecting a target sentence from the text based on the scoring result and generating a text summary according to the target sentence comprises:

[0016] sorting each sentence in the text based on the scoring result to obtain an importance degree corresponding to each sentence, and determining a sentence with an importance degree greater than a preset degree as the target sentence;

[0017] connecting each target sentence based on an initial position of each target sentence in the text to obtain the text summary.

[0018] Optionally, the step of constructing the text relation network comprises:

[0019] determining a sentence vector corresponding to each sentence according to a word vector of each sentence in the text;

[0020] determining a cosine similarity between any two sentence vectors, and determining a weight of a connecting edge between each sentence according to the cosine similarity, wherein each sentence represents a sentence node;

[0021] generating the text relation network according to each sentence node and the weight of the corresponding connecting edge.

[0022] Optionally, the step of dividing the text relation network into a plurality of sub-topic communities comprises:

[0023] initializing the text relation network to obtain a plurality of initial sub-communities and a first modularity of the text relation network, wherein a number of the initial sub-communities is equal to a number of sentence nodes in the text relation network;

[0024] obtaining adjacent communities of each sentence node in each initial sub-community, and predicting a second modularity of the text relation network after each sentence node is merged into a corresponding adjacent community.

[0025] determining a module increment of the text relation network before and after merging of each sentence node according to the first module and the second module;

[0026] when the module increment is greater than a preset value, merging the sentence node corresponding to the module increment greater than the preset value into a corresponding adjacent community, and updating the text relation network;

[0027] when the module of the text relation network is unchanged, determining a community division result according to the updated text relation network.

[0028] Optionally, before the steps of constructing the text relation network and dividing the text relation network into a plurality of sub-topic communities, the method further comprises:

[0029] detecting a hot spot type corresponding to the text;

[0030] when the hot spot type is a sudden hot spot type, performing the steps of constructing the text relation network and dividing the text relation network into a plurality of sub-topic communities;

[0031] when the hot spot type is a common hot spot type, constructing a BTM topic model, detecting topic information of the text by using the BTM topic model, obtaining a number of topics in the text and a set of word pairs in the text, and determining a topic distribution of a current word pair according to the number of topics in the text and the set of word pairs in the text.

[0032] Optionally, before the steps of constructing the BTM topic model, detecting topic information of the text by using the BTM topic model, obtaining a number of topics in the text and a set of word pairs in the text, and determining a topic distribution of a current word pair according to the number of topics in the text and the set of word pairs in the text, the method further comprises:

[0033] initializing a number of topics, and generating a topic by using a topic model modeling;

[0034] calculating a cosine similarity and an average topic similarity of any two topics, and comparing a size of a historical topic similarity and the average topic similarity;

[0035] when the historical topic similarity is less than the average topic similarity, reversing a topic direction in a next iteration process relative to a current round topic direction;

[0036] determining a number of noise topics according to a number of topics whose historical topic similarity is less than the average topic similarity;

[0037] updating the topic number, and determining the topic number in the next iteration process according to the updated topic number, the noise topic number and the topic direction;

[0038] repeating the above steps until the topic number is unchanged or a preset iteration number is reached, and outputting the topic number in the text.

[0039] In addition, to achieve the above-mentioned purpose, the present application also provides an abstract extraction system comprising:

[0040] a network construction module, configured to construct a text relationship network and divide the text relationship network into a plurality of sub-topic communities, the text relationship network being composed of a plurality of sentence nodes and connection edges between the sentence nodes;

[0041] a score result determination module, configured to obtain a score result of each sentence according to a sentence weight of each sentence in each sub-topic community and a contribution value of a word in each sentence to the sentence;

[0042] a text abstract generation module, configured to select a target sentence from the text based on the score result and generate a text abstract according to the target sentence.

[0043] In addition, to achieve the above-mentioned purpose, the present application also provides an abstract extraction device comprising a memory, a processor and an abstract extraction program stored on the memory and executable on the processor, the abstract extraction program realizing the steps of the abstract extraction method when executed by the processor.

[0044] In addition, to achieve the above-mentioned purpose, the present application also provides a computer readable storage medium, the computer readable storage medium storing an abstract extraction program, the abstract extraction program realizing the steps of the abstract extraction method when executed by a processor.

[0045] The abstract extraction method, system, device and storage medium provided in the embodiments of the present application have the technical scheme that, compared with the prior art which only considers the importance of sentences in the abstract extraction process, the present application constructs a text relationship network, divides the text relationship network into a plurality of sub-topic communities, the text relationship network being composed of a plurality of sentence nodes and connection edges between the sentence nodes, obtains a score result of each sentence according to a sentence weight of each sentence in each sub-topic community and a contribution value of a word in each sentence to the sentence, selects a target sentence from the text based on the score result, and generates a text abstract according to the target sentence, so that in the abstract extraction process, not only the importance of sentences is considered, but also the mutual correlation between words and sentences is considered, thereby improving the accuracy of the abstract extraction result. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 A flowchart of a first embodiment of the abstract extraction method of the present application is shown in the figure;

[0047] Figure 2 A function module diagram of the abstract extraction system of the present application is shown in the figure;

[0048] Figure 3 A structure diagram of the abstract extraction device of the present application is shown in the figure.

[0049] The implementation, functional features and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings, which are only an embodiment diagram but not the whole of the invention. DETAILED DESCRIPTION

[0050] In order to better understand the above technical solutions, the exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0051] As shown in the figure, in the first embodiment of the present application, the abstract extraction method of the present application comprises the following steps: Figure 1 Step S110, constructing a text relationship network and dividing the text relationship network into a plurality of sub-topic communities, the text relationship network being composed of a plurality of sentence nodes and connecting edges between the sentence nodes.

[0052] In the present embodiment, each sentence in the text is taken as a sentence node, and the text similarity between each sentence is calculated according to the distance between each sentence, and then the text relationship network is constructed according to the sentence node, the text similarity and the text connection relationship. For example, the text relationship network is

[0053] wherein, , represents a set of sentences in the text, E is a set of text relationships, and W is a set of text similarities between the sentence nodes in the text relationship network. Among them, a sentence node represents a text unit, and the text relationship network composed of text units as nodes, the contact of the text units with high relevance in the network is also more closely expressed, which can better reflect similar complaint hotspots. And the traditional clustering algorithm also needs to determine the number of clusters, which has a certain influence on the final result whether too large or too small, therefore, the present application constructs a weighted undirected text relationship network, and directly divides the network into communities by using the community division algorithm in complex network, which saves the related cumbersome steps of setting K value and quickly obtains the result of community division.

[0054] ​In the embodiment, the text relation network is divided into a plurality of sub-topic communities by using a community division algorithm. Each vertex or each node in the text relation network is expressed as a community isolated from each other, i.e., the number of sub-topic communities is equal to the total number of vertices or nodes in the text relation network. For example, it is assumed that a text is divided into three sentences, and each sentence is taken as a sentence node, i.e., the text relation network is composed of three sentence nodes. When the community division is performed, the number of sub-topic communities obtained by the division is three.

[0055] In the embodiment, in the process of dividing the text relation network into a plurality of sub-topic communities, the input of the community division algorithm is the text relation network, specifically including a sentence set in the text, a text relation set, and a text similarity set between the sentence nodes in the text relation network. The output of the community division algorithm is a community division result, specifically including which sub-topic communities. Since the similarity between the sentences in the initially constructed text relation network is low, the relation between the sentences is not close, and the number of divided communities is large. It can be understood that the more the number of communities is, the less close the relation between the sentences is, and the complaint hotspots cannot be well reflected. In order to reflect the complaint hotspots, the closeness between the sentences needs to be improved. Therefore, in the process of community division, the community to which each sentence node belongs is iteratively updated, and the text relation network is iteratively updated, so that the relation between each text unit in the finally obtained community division result is closer, and the complaint hotspots can be better reflected.

[0056] In the embodiment, in the process of dividing the text relation network by using the community division algorithm, the process mainly includes three parts of initialization, community merging, and text relation network reconstruction, and the finally obtained community division result is obtained by iteratively updating. In the finally obtained community division result, there can be one or more sentences in each sub-topic community.

[0057] In step S120, the score result of each sentence is obtained according to the sentence weight of each sentence in each sub-topic community and the contribution value of each word in each sentence to the sentence.

[0058] In the embodiment, after the community division is completed, the abstract extraction is performed.

[0059] At present, the abstract extraction is completed by taking the sentence as a text unit, and the influence of the word on the sentence is mostly ignored. However, the sentence is composed of words, and the importance of the word has a great influence on the importance of the sentence. Therefore, in the abstract extraction stage, the contribution of the word in the sentence to the sentence is considered in the sentence importance sorting, so that the text abstract extraction can be more accurately completed.

[0060] In the embodiment, the application scores each sentence in each sub-topic community to obtain a score result of each sentence. In the scoring process, not only the sentence weight of each sentence is considered, but also the contribution value of each word in each sentence to the sentence is considered, so that the summary extraction result is more in line with the actual demand.

[0061] Optionally, the score result of the sentence can be obtained by superimposing the sentence weight of the sentence and the contribution value of the word in the sentence to the sentence. Optionally, a corresponding first proportion coefficient can be set for the sentence weight of the sentence and a corresponding second proportion coefficient can be set for the contribution value of the word to the sentence, the score result of the sentence can be obtained by superimposing the product of the sentence weight of the sentence and the first proportion coefficient and the product of the contribution value of the word to the sentence and the second proportion coefficient, wherein the sum of the first proportion coefficient and the second proportion coefficient is 1.

[0062] In step S130, the target sentence is selected from the text based on the score result, and the text summary is generated according to the target sentence.

[0063] In the embodiment, the target sentence is a sentence for generating a text summary, that is, a sentence with a high degree of importance in the text, which can reflect a key sentence of a sudden hot issue. After the community division is completed, each sub-topic community is taken as a research object, the nodes in each topic community are arranged based on the score result, and the key nodes are selected from all nodes according to the sorting result. These key nodes correspond to the target sentences with a high degree of importance in the text, and are arranged in sequence according to the positions of the target sentences in the text to form a text summary.

[0064] In the embodiment, according to the above technical solution, the text relationship network is constructed, and the text relationship network is divided into a plurality of sub-topic communities, the text relationship network is composed of a plurality of sentence nodes and connection edges between the sentence nodes; the score result of each sentence is obtained according to the sentence weight of each sentence in each sub-topic community and the contribution value of each word in each sentence to the sentence; the target sentence is selected from the text based on the score result, and the text summary is generated according to the target sentence. In the summary extraction process, not only the importance of the sentence is considered, but also the mutual correlation between the word and the sentence is considered, and the accuracy of the summary extraction result is improved.

[0065] Further, based on step S110 of the first embodiment, the step of constructing the text relationship network comprises:

[0066] In step S111, the sentence vector corresponding to each sentence is determined according to the word vector of each sentence in the text.

[0067] In this embodiment, the current text is obtained, which can be a complaint text. The text can be preprocessed, and then the word vectors in each sentence of the preprocessed text are calculated, the sentence is represented as a vector, and the feature word weight is taken as a component of the vector.

[0068] Optionally, the word vector of a sentence can be obtained by weighted averaging or concatenating the word vectors of each word in the sentence to represent the entire sentence. Some common methods are as follows:

[0069] Word2Vec average method: add the word vectors of each word in the sentence, then divide by the total number of words to get the average vector as the representation of the sentence.

[0070] TF-IDF weighted average method: for each word, calculate its TF-IDF weight in the sentence, then multiply the word vector by the corresponding weight, add all the weighted word vectors, and divide by the total number of words to get the weighted average vector as the representation of the sentence.

[0071] Doc2Vec method: Doc2Vec is an extension of Word2Vec, which can directly learn document vectors with fixed length. By training a Doc2Vec model, the vector representation of the sentence can be obtained.

[0072] Use pre-trained language models: use pre-trained language models such as BERT or GPT to encode the sentence and obtain the context-related representation of the sentence. These models usually output the representation of each word in the sentence, which can be weighted averaged or further processed to obtain the representation of the entire sentence.

[0073] It should be noted that different calculation methods may be suitable for different tasks or corpora. Moreover, when using pre-trained models, some additional processing steps may be required, such as tokenization, stop word removal, etc., to ensure that the input format meets the requirements of the model.

[0074] In addition, a word vector model can also be trained by oneself, and the word vectors can be trained by a large amount of corpus, and these self-trained word vectors can be used when calculating the sentence vector. This method requires more data and computing resources, but can better adapt to the specific context of the task.

[0075] In summary, the calculation of the word vector of the sentence can be selected according to the specific requirements, and the common methods include average method, weighted average method, pre-trained model and self-trained word vector model.

[0076] Step S112, determine the cosine similarity between any two sentence vectors, and determine the weight of the connection edge between each sentence according to the cosine similarity, wherein each sentence represents a sentence node.

[0077] In the embodiment, after the sentence vectors of each sentence are calculated, the cosine similarity between any two sentences is calculated. Specifically, the cosine similarity between any two sentences in the text is calculated according to the cosine similarity formula. Assuming that there are a sentence vector A and a sentence vector B, the calculation process is as follows: first, the dot product of the sentence vector A and the sentence vector B is calculated, that is, the elements at the corresponding positions of the two vectors are multiplied and then added. Then, the modulus of the sentence vector A and the modulus of the sentence vector B are calculated. Finally, the dot product is divided by the product of the two vector moduli. The cosine similarity ranges from -1 to 1, where 1 represents complete similarity, -1 represents complete dissimilarity, and 0 represents no correlation.

[0078] In step S113, the text relationship network is generated according to the sentence nodes and the weights of the corresponding connection edges.

[0079] In the embodiment, the similarity matrix is output and the text relationship network is constructed. The sentence is taken as a network node, and it is determined whether there is a connection edge between two nodes through the calculation of the similarity. The cosine similarity between the sentence vectors is taken as the weight of the connection edge between the nodes. The text relationship network is generated according to the sentence nodes and the weights of the corresponding connection edges. In the text relationship network, the sentence nodes with higher cosine similarity are more closely connected.

[0080] In the embodiment, the cosine similarity between two sentence vectors is calculated according to the cosine similarity algorithm, the weight of the connection edge between each sentence is determined, and the text relationship network is generated according to the sentence nodes and the weights of the corresponding connection edges. Therefore, the construction of the text relationship network is realized.

[0081] Further, based on the steps S110 or S112-S113 of the first embodiment, the step of dividing the text relationship network into a plurality of sub-topic communities includes:

[0082] In step S114, the text relationship network is initialized to obtain a plurality of initial sub-communities and a first modularity of the text relationship network, wherein the number of the initial sub-communities is equal to the number of sentence nodes in the text relationship network.

[0083] In step S115, the adjacent communities of each sentence node in each initial sub-community are obtained, and a second modularity of the text relationship network after each sentence node is merged into the corresponding adjacent community is predicted.

[0084] In step S116, the modularity increment of the text relationship network before and after each sentence node is merged is determined according to the first modularity and the second modularity.

[0085] In the embodiment, the community division algorithm is used to divide the communities. The principle of community division is that each node takes itself as its community label. Each node traverses all its neighbor nodes, tries to update its community label to the community label of the neighbor node, selects the community label with the maximum modularity increment (greedy idea), and stops until all nodes cannot increase the modularity by changing the community label.

[0086] In the embodiment, the community division algorithm is used to divide the text relation network into multiple sub-topic communities. The input of the community division algorithm is the text relation network: , wherein, , represents a sentence in the text, E is a text relation set, and W is a set of text similarities between sentence nodes in the text relation network. The output of the community division algorithm is a community division result: .

[0087] In the embodiment, in the process of community division, the first step needs to be initialized, that is, the text relation network is initialized into multiple initial sub-communities. Specifically, the vertices in the text relation network are all initialized as initial sub-communities isolated from each other , and the number of initial sub-communities is equal to the total number of sentence nodes in the text relation network.

[0088] In the embodiment, after the multiple initial sub-communities are obtained by initialization, community merging is performed. After the community merging, the first modularity of the text relation network is calculated, which is used to represent the closeness within the community before the community merging.

[0089] In the embodiment, for each sentence node , it is tried to be merged into its adjacent community, and the second modularity after merging is calculated. The calculation methods of the first modularity and the second modularity are conventional methods, which are not described here.

[0090] Specifically, first, each node is assigned to a unique community, and then the nodes are moved between these communities in order. Assuming that there is a node i, it has three neighbor nodes j1, j2, j3, it is tried to move the node i to the communities where j1, j2, j3 are located, and the corresponding modularity increment ΔQ is calculated. The node i is moved to the corresponding community with the maximum change value. According to this method, iteration is repeated until the movement of any node in the network cannot improve the total modularity.

[0091] The modularity gain obtained by moving to a community C can be easily calculated as:

[0092]

[0093] where,∑in is the sum of weights of links within the community C (equal to the degree when the weight is 1), if it is an initial condition, i.e. a node as a community, it is the connection of the node itself to itself, at this time the starting point weight and the end point weight are still needed (even if the starting point and the end point are the same node at this time, the weight is 0 for the acyclic graph).∑tot is the sum of weights of links associated to the nodes in C. ki is the sum of weights of links associated to node i. Ki,in is the sum of links from node i to the nodes in C. m is the sum of weights of all links in the network.

[0094] Step S117, when the module degree increment is greater than the preset value, merging the sentence node corresponding to the module degree increment greater than the preset value into the corresponding adjacent community, and updating the text relation network;

[0095] In this embodiment, each community is merged into a new super node, and the edge weight of the super node is the sum of the edge weights of all nodes in the original community, forming a new text relation network. If the module degree increment is less than the preset value, the corresponding sentence node is continued to be retained in the community in which it is located before, so as to obtain the community division result. The preset value can be set to 0.

[0096] Step S118, when the module degree of the text relation network is unchanged, determining the community division result according to the updated text relation network.

[0097] In this embodiment, according to the above technical scheme, since the community division algorithm is used to perform community division and mining on the sudden hot issue, the accuracy of community division is improved, and the calculation complexity is reduced.

[0098] Further, based on step S120 of the first embodiment, the step of obtaining the scoring result of each sentence according to the sentence weight of each sentence in each sub-topic community and the contribution value of the word in each sentence to the sentence includes:

[0099] Step S121, obtaining the sentence weight of each sentence after the last iteration, the damping coefficient, the number of nodes of the text relation network and the weight of the connection edge between the sentence nodes in each sub-topic community;

[0100] Step S122, determining the sentence weight of each sentence after this iteration according to the sentence weight of each sentence after the last iteration, the damping coefficient, the number of nodes of the text relation network and the weight of the connection edge between the sentence nodes.

[0101] In this embodiment, the semantic information of the text is considered, the cosine similarity between the sentence vectors generated by Word2vec is used as the weight of the connection edge, and it is substituted into the TextRank algorithm to complete the sentence weight calculation.

[0102] wherein the TextRank algorithm is a graph-based text summarization and keyword extraction algorithm, which is widely used in natural language processing field. The algorithm is based on PageRank algorithm, by taking the sentences or words in the text as nodes, a graph containing the relationship between nodes is constructed, and the importance is calculated by using the weight between nodes. The main idea of TextRank algorithm is as follows:

[0103] (1) Divide the text into sentences or words: first, pre-process the text, such as word segmentation, stop word removal, etc., and then divide the text into sentences or words as nodes.

[0104] (2) Build a graph: according to the semantic relationship in the text, a directed and weighted graph is constructed. Each node in the graph represents a sentence or a word, and the edge represents the relationship between nodes. Usually, word co-occurrence, sentence similarity, etc. can be used as the weight of the edge.

[0105] (3) Calculate the importance score of the node: use iterative calculation to give each node an initial importance score. Then, update the score of the node through multiple iterations until convergence. The score of the node can be calculated using the principle of PageRank algorithm, that is, the score of the node is determined by the score of the nodes connected to it and the weight of the edge.

[0106] (4) Extract keywords or generate summary: according to the score of the node, the nodes with higher score can be used as keywords or important sentences for keyword extraction or summary generation.

[0107] The advantages of TextRank algorithm include:

[0108] (1) Unsupervised learning: TextRank algorithm is an unsupervised learning method, which does not need to rely on labeled data. It can automatically extract keywords and generate summaries from original text, with strong universality and applicability.

[0109] (2) Consider context information: TextRank algorithm uses the method of constructing a graph to fully consider the context information between words and sentences in the text, so as to better understand the semantic relationship of the text.

[0110] (3) Scalability: TextRank algorithm can be applied to large-scale text processing tasks, and can be customized and optimized for specific fields or tasks.

[0111] In summary, the TextRank algorithm is a graph-based text processing method that can be used for tasks such as keyword extraction and text summarization. It has advantages such as unsupervised learning, consideration of contextual information, and scalability.

[0112] In this embodiment, the sentence weight of each sentence in the current iteration is determined based on the sentence weight of each sentence after the previous iteration, the damping coefficient, the number of nodes in the text relation network, and the weights of the connection edges between the sentence nodes. Specifically, the sentence weight of each sentence in the current iteration is determined according to the following formula:

[0113] .

[0114] in, Represents sentence nodes and The weight values ​​of the edges connecting them. represents the TextRank value, which is the sentence weight after this iteration, i.e., the sentence score of the i-th sentence after this round of iterations. d is the damping coefficient, typically 0.85, and N is the total number of sentence nodes in the text relation network. This represents the sentence weight of sentence j after the last iteration.

[0115] Step S123: Obtain the complex network integrated feature value of each word in each sentence.

[0116] In this embodiment, the complex network comprehensive feature value of each word in the sentence can be obtained by weighted summation or comprehensive evaluation using indicators such as network connectivity index, node degree distribution, structural features, modularity, and advanced features.

[0117] Step S124: Based on the sentence weight of each sentence after this iteration and the complex network comprehensive feature value of each word in each sentence, obtain the contribution value of each word in the sentence to the sentence;

[0118] Step S125: Combine the sentence weight of each sentence after this iteration with the contribution value of each word in each sentence to the sentence to obtain the score result of each sentence.

[0119] In this embodiment, existing methods for summarizing text using sentences as text units mostly neglect the influence of words on sentences. Since sentences are composed of words, and the importance of words significantly impacts the importance of sentences, this application incorporates the contribution of words to sentences into the sentence importance ranking in the final summarization stage to more accurately extract text summaries. The specific method is as follows:

[0120] Calculate the complex network comprehensive feature values ​​of words to obtain the weight of each word in the text, and represent the text as a matrix. , is the feature value of the jth word in the ith sentence, and the word-sentence relationship expression is obtained by counting the weight of the sentence containing the word j. The word-sentence relationship expression measures the contribution value of a word to a sentence in a text and considers the influence of word information on a sentence.

[0121]

[0122] .

[0123] wherein, is the score of the ith sentence after the iteration, is the complex network comprehensive feature value of each word.

[0124] Then, the contribution value of each word in a sentence to the sentence is calculated according to the following formula to measure the word-sentence relationship, that is, to derive the expression of the part of the final score of the sentence participating in the evaluation of the sentence:

[0125] .

[0126] Finally, the sentence weight of each sentence after the iteration and the contribution value of each word in the sentence to the sentence are fused to form a text sentence scoring function of the following formula. Alternatively, a corresponding first proportion coefficient can also be set for the sentence weight of the sentence and a corresponding second proportion coefficient can also be set for the contribution value of the word to the sentence, and the product of the sentence weight of the sentence and the first proportion coefficient is superimposed with the product of the contribution value of the word to the sentence and the second proportion coefficient, so as to obtain the scoring result of the sentence, wherein the sum of the first proportion coefficient and the second proportion coefficient is 1. The formed text sentence scoring function is as follows:

[0127] .

[0128] wherein, . is the first proportion coefficient, is the second proportion coefficient.

[0129] According to the technical solution described above, in the abstract extraction process, not only the importance of the sentence is considered, but also the mutual correlation between the word and the sentence is considered, so that the accuracy of the abstract extraction result is improved.

[0130] Further, based on the first embodiment, after step S130 and steps S121-S125, the step of obtaining the scoring result of each sentence according to the sentence weight of each sentence and the contribution value of each word in the sentence to the sentence in each sub-topic community includes:

[0131] Step S131, ranking each sentence in the text based on the scoring result to obtain the importance degree of each sentence, and determining the sentence with an importance degree greater than a preset degree as the target sentence;

[0132] Step S132, connecting each target sentence based on the initial position of each target sentence in the text to obtain the text summary.

[0133] In this embodiment, after the community division is completed, each sub-community after division is taken as a research object, the nodes in each community are arranged in reverse order, and the top K nodes are extracted as key nodes, which correspond to sentences with high importance in the text. According to the position of the sentence in the text, the text summary is arranged in order. Alternatively, a preset degree can be set, and the sentence with an importance degree greater than the preset degree is determined as the target sentence. Since only the target key sentence is extracted to form the text summary, the text summary is more concise.

[0134] Based on step S110 of the first embodiment, in the second embodiment of the present application, the summary extraction method of the present application includes the following steps:

[0135] Step S210, detecting the hot topic type corresponding to the text.

[0136] In this embodiment, the hot topic type includes a sudden hot topic and a general hot topic. Other types can also be included.

[0137] Step S220, when the hot topic type is a sudden hot topic type, performing the steps of constructing a text relationship network and dividing the text relationship network into a plurality of sub-topic communities.

[0138] In this embodiment, when the hot topic type is a sudden hot topic type, the related steps of the first embodiment are performed.

[0139] Step S230, when the hot topic type is a general hot topic type, constructing a BTM topic model, detecting the topic information of the text using the BTM topic model to obtain the number of topics in the text and the set of word pairs in the text, and determining the topic distribution of the current word pair according to the number of topics in the text and the set of word pairs in the text.

[0140] In this embodiment, the general hot topic is usually a problem frequently encountered in the customer service system, and the user's feedback problem is often concentrated in some aspects. For this type of problem, it is generally not necessary to browse all the content, and the overall situation of this type of problem can be obtained through some key information. Therefore, the present application uses a topic model to mine and analyze the complaint text.

[0141] The traditional topic analysis uses the LDA topic model, but the complaint texts collected in the customer service system are mainly short texts, and due to the relatively small amount of words of these short texts, the LDA model may face problems such as data sparsity, therefore, the BTM topic model is used, the data model is directly modeled for text data when a word pair is constructed, is not limited by the length of the short text, and the quality of the text topics mined on the short text data is also good, the BTM improves the problem of sparsity of the short text data by constructing a word pair model for the document, and the model principle is as follows:

[0142] The word pair refers to a co-occurring unordered word pair in the text, for a document containing 3 different words, the word pair generation process formula is as follows. After preprocessing, a short text document corpus with n words can generate word pairs.

[0143]

[0144] Given a corpus containing documents , each document can be represented as , and an implicit theme set is introduced as . The BTM generates a graph model as shown in the following figure, and the specific process is as follows:

[0145] (1) Assume that the global theme distribution of the entire word pair set is ;

[0146] (2) For each theme z: assume the word distribution under the theme ;

[0147] (3) For each word in the word pair set B, perform the following operations;

[0148] (4) extract the theme ;

[0149] (5) extract the word from the word distribution corresponding to the theme z;

[0150] extract the word from the word distribution corresponding to the theme z;

[0151] ;

[0152] According to the above process, the joint probability calculation formula of the word pair is:

[0153] ;

[0154] The generation probability of the whole BTM corpus word pair set B is as follows:

[0155] ;

[0156] In the BTM model, the Gibbs sampling method is used to estimate the "word pair set-topic" distribution and the "topic-word" distribution . The topic distribution of the current word pair is calculated by the following formula:

[0157] ;

[0158] In the formula, K is the number of topics in the text, B is the word pair set in the text, are the hyperparameters of the Dirichlet prior distribution, respectively.

[0159] According to the technical scheme, the improved topic mining algorithm is used to extract the content of the complaint text, the common hot issues in the complaint text are displayed, the intuitive and readable topic description is formed, and a method suitable for learning the number of topics is proposed, so that the number of topics of the model is calculated, and the quality of topic discovery is improved.

[0160] Further, based on the step S230 of the second embodiment, that is, constructing the BTM topic model, using the BTM topic model to detect the topic information of the text, obtaining the number of topics in the text and the word pair set in the text, before determining the topic distribution of the current word pair according to the number of topics in the text and the word pair set in the text, the method further comprises the following steps:

[0161] Step S310, initializing the number of topics, and generating topics by using the topic model;

[0162] Step S320, calculating the cosine similarity of any two topics and the average topic similarity, and comparing the size of the historical topic similarity and the average topic similarity;

[0163] Step S330, when the historical topic similarity is less than the average topic similarity, the topic direction is opposite to the topic direction of the current round in the next iteration process;

[0164] Step S340, determining the number of noise topics according to the number of topics whose historical topic similarity is less than the average topic similarity;

[0165] Step S350, when the topic direction changes, updating the number of topics, and determining the number of topics in the next iteration process according to the updated number of topics, the number of noise topics and the topic direction;

[0166] Step S360, repeating the above steps until the number of topics is constant or a preset iteration number is reached, and outputting the number of topics in the text.

[0167] In the embodiment, the selection of the number of topics has a great influence on the classification effect of the model during topic mining, and determines the result quality of topic mining of the topic model. The application proposes a method for adapting to learn the number of topics, which does not need to artificially specify the number of topics, and makes the model calculate the number of topics K most suitable for the text, thereby improving the topic discovery quality. The specific method is as follows:

[0168] (1) Randomly initialize a number of topics K;

[0169] (2) Use the topic model to generate topics and topic words, calculate the cosine similarity of any two topics and the average topic similarity. Compare the size of the historical topic similarity and the average topic similarity. If the historical topic similarity is less than the average topic similarity, the direction of the topic in the next iteration is opposite to the direction of the topic in the current round;

[0170] (3) Calculate the topic density of each topic, that is, the number of topics whose topic similarity is less than the average topic similarity, and calculate the noise topic number, that is, the number of topics whose topic density is less than or equal to K / 3. When the direction of the topic changes, update the number of topics. The number of topics in the next iteration process is determined according to the updated number of topics, the noise topic number and the topic direction, including: determining the product of the noise topic number and the topic direction, adding the number of topics and the product to obtain the number of topics in the next iteration process.

[0171] (4) Repeat the above steps until the number of topics is constant or a preset iteration number is reached. The preset iteration number can be set according to actual conditions.

[0172] According to the above technical solution, the embodiment improves the result quality of topic mining of the topic model.

[0173] In other embodiments, the application needs to perform text feature analysis and text data collection before constructing a text relationship network and dividing the text relationship network into a plurality of sub-topic communities, and pre-processes the text data sample. Then, according to the pre-processed text data sample, subsequent ordinary hot issue mining analysis and sudden hot issue mining analysis are performed. Specifically, it includes two parts: complaint text feature analysis and data collection and sample data preprocessing. Each part will be introduced in detail as follows:

[0174] (I) Complaint text feature analysis and data collection

[0175] 1. Complaint information feature analysis

[0176] Complaints refer to the collection of opinions, attitudes and emotions of customers on abnormal phenomena or problems in the use of enterprise products, with certain concentration, consistency and continuity, which will have a certain degree of impact on the development process of business. Customer complaints are different from general text information, with the characteristics of quick expression, interactive mode and information diversification. Therefore, the traditional processing method is difficult to analyze the complaint information, and the occurrence of hot issues often has the characteristics of not easy to find and uncontrollable, which puts forward higher requirements for complaint analysis.

[0177] According to the analysis of complaint data by the complaint department, the complaint text is divided into ordinary hot issues and sudden hot issues. Compared with ordinary problems in complaints, sudden hot issues are not easy to be found and cannot be identified by using traditional statistical methods. Ordinary problems usually have a higher probability of occurrence and a more stable time of outbreak.

[0178] 2. Complaint sample data collection

[0179] Data collection is the basic work of the present application, and the quality and quantity of data will have different degrees of noise effect on the effectiveness of the present application. The data used in the present application is from the complaint data of the customer service system of the enterprise, which is collected from the system database to generate data samples.

[0180] The usual sample collection method is to directly export all the data within a certain time span. This method can ensure that the training set and test set of data are sufficient, and there is no need to worry about the problem of poor model training effect caused by insufficient data. However, this collection method is too simple and rough, and there are not small differences between different months and different services. If it is used directly without screening, it may affect the subsequent analysis effect. Therefore, the present application collects samples based on different dimensions, and finally combines the data collected in different dimensions. The specific dimensions are:

[0181] (1) Time dimension

[0182] The complaint situation in different months of a year and different times of a day will be different. In summer, the experience of a certain product may be poor or the use of a certain service at night may not be good, so it is necessary to collect the corresponding data from the time dimension. The present application divides the time dimension into season dimension and day dimension, the season dimension is divided into four dimensions of spring, summer, autumn and winter, and the day dimension is divided into three dimensions of morning, afternoon and night, and the data with corresponding proportion is selected for each dimension.

[0183] (2) Product / service dimension

[0184] The quality difference of different products of an enterprise or the pros and cons of different services will affect the user complaints. If not filtered directly, the sample data may be unbalanced. For example, the design of a product of an enterprise may not be perfect, and the user complaints are mainly concentrated on this product, which will lead to more complaints about this product in the collected data, and finally the analysis result will also be mainly biased towards this product, and the analysis will be one-sided. Therefore, the data is collected based on the product / service dimension, and the data proportion of each product / service is basically balanced.

[0185] (3) User dimension

[0186] Different users of different identities have different concerns about products / services, and experienced users and new users have different focuses, which ultimately leads to different complaint contents. The user level recorded in the customer service system is used to filter the data.

[0187] (II) Sample data preprocessing

[0188] Before mining the complaint text, the collected data samples need to be preprocessed. Data preprocessing is the process of discovering and solving data quality problems, which usually includes cleaning of similar repeated data, abnormal data, inconsistent data, etc. Through data cleaning, the data quality is improved, so that subsequent analysis can make scientific judgments through data mining and data analysis. The original data content saved in the customer service system database is complex and has many features, which is not suitable for direct use as subsequent analysis data. Some data may not be related to subsequent analysis operations or contain interference terms, so it is necessary to preprocess the data before model construction. Preprocessing mainly includes:

[0189] 1. Noise data elimination

[0190] URL addresses, special symbols, pictures, and other contents or meaningless characters such as zero-width characters \u200b, \u200c may exist in complaint texts. These useless data not only greatly increase the workload of text processing, but also directly affect a series of operations in the later stage, thereby reducing the efficiency of text processing. Therefore, removing these noise data is the first task of data preprocessing.

[0191] 2. Chinese word segmentation

[0192] Complaint texts are too long and have a lot of interference data, so it is necessary to perform word segmentation on complaint texts. Chinese text uses Chinese characters as the basic unit, and there is no obvious morphological boundary between words like English. Moreover, there are many punctuation marks, and computers have difficulty identifying the relationship between words and sentences. Therefore, when processing Chinese text, the first thing to do is to separate words from words, i.e. word segmentation.

[0193] 3. Remove stop words

[0194] After word segmentation, some words in the key words will not help to understand the actual meaning of the reaction document, such as "of" and "of", which are called stop words. Such words have the characteristics of extremely common, high density of appearance, and often appear in any category of articles, but they have no meaning for representing the text. In the feature extraction process, stop words will cause errors in the results, so it is necessary to remove stop words in the text preprocessing stage. This not only eliminates interference, but also reduces the dimension of the text.

[0195] 4. Part-of-speech tagging

[0196] In addition to word segmentation and stop word processing, part-of-speech tagging is also an important part of text preprocessing, which is beneficial to key word screening after word segmentation. When processing complaint texts, since it is a special domain text mining, the word segmentation results need to be further screened according to the particularity of the complaint text. The mood words, interjections and other virtual words in the complaint text have no actual meaning for mining hotspots, and the important ones include nouns, verbs, names, place names and other real words, so after removing stop words, the part-of-speech tagging of the word segmentation results is needed. Extract the words of related parts of speech as the result.

[0197] The application improves the accuracy of the mining analysis result by analyzing the characteristics of the complaint text, collecting data and preprocessing sample data, and performing subsequent ordinary hot topic mining analysis and sudden hot topic mining analysis based on the preprocessed text data samples.

[0198] Embodiments of the abstract extraction method are provided. It should be noted that although a logical sequence is shown in the flowchart, in some cases, the steps shown or described can be performed in an order different from that shown.

[0199] As shown in Figure 2 The abstract extraction system provided by the application comprises:

[0200] The network construction module 10 is configured to construct a text relationship network and divide the text relationship network into a plurality of sub-topic communities, the text relationship network being composed of a plurality of sentence nodes and connection edges between the sentence nodes.

[0201] The scoring result determination module 20 is configured to obtain a scoring result of each sentence according to a sentence weight of each sentence in each sub-topic community and a contribution value of a word in each sentence to the sentence.

[0202] The text abstract generation module 30 is configured to select a target sentence from the text based on the scoring result, and generate a text abstract according to the target sentence.

[0203] The embodiment of the application abstract extraction system is basically the same as the above-mentioned abstract extraction method, and will not be repeated here.

[0204] As Figure 3 shown, Figure 3 is a structural schematic diagram of a hardware operating environment of an abstract extraction device involved in an embodiment of the application. The abstract extraction device can include a processor 1001, such as a CPU, a memory 1005, a user interface 1003, a network interface 1004, and a communication bus 1002. The communication bus 1002 is used to realize the connection and communication between the components. The user interface 1003 can include a display screen, an input unit such as a keyboard, and the optional user interface 1003 can further include a standard wired interface and a wireless interface. The network interface 1004 can optionally include a standard wired interface and a wireless interface (such as a WI-FI interface). The memory 1005 can be a high-speed RAM memory or a stable memory such as a disk memory. The memory 1005 can also be a storage device independent of the aforementioned processor 1001.

[0205] Those skilled in the art can understand that Figure 3 the structure of the abstract extraction device shown in the above embodiment does not constitute a limitation on the abstract extraction device, and can include more or fewer components than those shown, or combine certain components, or different component arrangements.

[0206] As Figure 3 shown, the memory 1005 as a storage medium can include an operating system, a network communication module, a user interface module, and an abstract extraction program. The operating system is a program that manages and controls the hardware and software resources of the abstract extraction device, and the running of the abstract extraction program and other software or programs.

[0207] In Figure 3 the abstract extraction device shown, the user interface 1003 is mainly used to connect a terminal and communicate data with the terminal; the network interface 1004 is mainly used for a background server and communicates data with the background server; and the processor 1001 can be used to call the abstract extraction program stored in the memory 1005.

[0208] In this embodiment, the abstract extraction device includes a memory 1005, a processor 1001, and an abstract extraction program stored in the memory and executable on the processor, wherein:

[0209] When the processor 1001 calls the abstract extraction program stored in the memory 1005, the following operations are performed:

[0210] construct a text relation network, and divide the text relation network into a plurality of sub-topic communities, the text relation network being composed of a plurality of sentence nodes and connecting edges between the sentence nodes;

[0211] obtain a scoring result of each sentence according to a sentence weight of each sentence in each sub-topic community and a contribution value of a word in each sentence to the sentence;

[0212] select a target sentence from the text based on the scoring result, and generate a text summary according to the target sentence.

[0213] Based on the same inventive concept, the embodiments of the present application further provide a computer readable storage medium, which stores a summary extraction program, the summary extraction program being executed by a processor to implement each step of the summary extraction method and achieve the same technical effects. To avoid repetition, no further description is given here.

[0214] The storage medium provided by the embodiments of the present application is the storage medium used for implementing the method of the embodiments of the present application. Therefore, based on the method introduced in the embodiments of the present application, the specific structure and variations of the storage medium can be understood by those skilled in the art, and thus no further description is given here. Any storage medium used by the method of the embodiments of the present application belongs to the scope of the present application.

[0215] It should be noted that in this document, the terms "include", "contain" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method, article or system including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or system. Without more limitations, the element defined by the statement "including a" does not exclude the presence of another identical element in the process, method, article or system including the element.

[0216] The above sequence numbers of the embodiments of the present application are only for description, and do not represent the advantages and disadvantages of the embodiments.

[0217] Those skilled in the art can clearly understand the above-mentioned embodiment method can be realized by means of software and the necessary general hardware platform, of course, can also be through hardware, but in many cases the former is a better embodiment. Based on such understanding, the technical solutions of the present application essentially or say the part of the prior art contribution can be embodied in the form of software products, the computer software product is stored in a storage medium (such as ROM / RAM, magnetic disc, optical disc) as described above, including a number of instructions to make a terminal device (may be a mobile phone, computer, server, television, or network equipment, etc.) executes the method described in various embodiments of the present application.

[0218] The above is only the preferred embodiment of the present application, not therefore limit the patent scope of the present application, all use the contents of the present application specification and drawings of equivalent structure or equivalent process transformation, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. An abstract extraction method characterized by, The abstract extraction method comprises: detecting a hot spot type corresponding to the text; when the hot spot type is a common hot spot type, constructing a BTM topic model, detecting topic information of the text by using the BTM topic model, obtaining a topic quantity in the text and a word pair set in the text, and determining a topic distribution of a current word pair according to the topic quantity in the text and the word pair set in the text, wherein the detecting of the topic information of the text by using the BTM topic model to obtain the topic quantity in the text comprises: initializing the topic quantity, and generating a topic by using a topic model modeling; calculating a cosine similarity of any two topics and an average topic similarity, and comparing a historical topic similarity with the average topic similarity; when the historical topic similarity is less than the average topic similarity, reversing a topic direction in a next iteration process from a topic direction in a current iteration process; determining a noise topic quantity according to a topic quantity for which the historical topic similarity is less than the average topic similarity; when the topic direction changes, updating the topic quantity, and determining a topic quantity in the next iteration process according to the updated topic quantity, the noise topic quantity and the topic direction; repeating the above steps until the topic quantity is constant or a preset iteration number is reached, and outputting the topic quantity in the text; when the hot spot type is a sudden hot spot type, constructing a text relationship network, and dividing the text relationship network into a plurality of sub-topic communities, the text relationship network being composed of a plurality of sentence nodes and connection edges between the sentence nodes; obtaining a scoring result of each sentence according to a sentence weight of each sentence in each sub-topic community and a contribution value of a word pair to a sentence in each sentence; selecting a target sentence from the text based on the scoring result, and generating a text abstract according to the target sentence.

2. The abstract extraction method of claim 1, wherein, The step of obtaining the scoring result of each sentence according to the sentence weight of each sentence in each sub-topic community and the contribution value of the word pair to the sentence in each sentence comprises: obtaining a sentence weight of each sentence after a last iteration, a damping coefficient, a node quantity of the text relationship network and a weight of a connection edge between the sentence nodes in each sub-topic community; determining a sentence weight of each sentence after a current iteration according to the sentence weight of each sentence after the last iteration, the damping coefficient, the node quantity of the text relationship network and the weight of the connection edge between the sentence nodes; obtaining a complex network comprehensive feature value of each word in each sentence; obtaining a contribution value of a word pair to a sentence in each sentence according to the sentence weight of each sentence after the current iteration and the complex network comprehensive feature value of each word in the sentence; fusing the sentence weight of each sentence after the current iteration and the contribution value of the word pair to the sentence in the sentence to obtain a scoring result of each sentence.

3. The abstract extraction method of claim 1 or 2, wherein, The step of selecting the target sentence from the text based on the scoring result and generating the text abstract according to the target sentence comprises: sequencing each sentence in the text based on the scoring result to obtain an importance degree corresponding to each sentence, and determining a sentence with an importance degree greater than a preset degree as a target sentence; connecting each target sentence based on an initial position of each target sentence in the text to obtain the text summary.

4. The abstract extraction method of claim 1, wherein, The step of constructing the text relation network comprises: determining a sentence vector corresponding to each sentence according to a word vector of each sentence in the text; determining a cosine similarity between any two sentence vectors, and determining a weight of a connection edge between each sentence according to the cosine similarity, wherein each sentence represents a sentence node; generating the text relation network according to each sentence node and the weight of the corresponding connection edge.

5. The abstract extraction method of claim 1 or 4, wherein, The step of dividing the text relation network into a plurality of sub-topic communities comprises: initializing the text relation network to obtain a plurality of initial sub-communities and a first modularity of the text relation network, wherein the number of initial sub-communities is equal to the number of sentence nodes in the text relation network; obtaining adjacent communities of each sentence node in each initial sub-community, and predicting a second modularity of the text relation network after each sentence node is merged into the corresponding adjacent community; determining a modularity increment of the text relation network before and after each sentence node is merged according to the first modularity and the second modularity; when the modularity increment is greater than a preset value, merging the sentence node corresponding to the modularity increment greater than the preset value into the corresponding adjacent community, and updating the text relation network; when the modularity of the text relation network is unchanged, determining a community division result according to the updated text relation network.

6. An abstract extraction system characterized by, The summary extraction system comprises: detecting a hot spot type corresponding to the text; when the hot spot type is a common hot spot type, constructing a BTM topic model, detecting topic information of the text using the BTM topic model to obtain a number of topics in the text and a set of word pairs in the text, and determining a topic distribution of a current word pair according to the number of topics in the text and the set of word pairs in the text, wherein the detection of the topic information of the text using the BTM topic model to obtain the number of topics in the text comprises: initializing the number of topics, and generating a topic using a topic model; calculating a cosine similarity between any two topics and an average topic similarity, and comparing the size of the historical topic similarity and the average topic similarity; when the historical topic similarity is less than the average topic similarity, reversing the topic direction in the next iteration process; determining a number of noise topics according to the number of topics whose historical topic similarity is less than the average topic similarity; when the topic direction changes, updating the number of topics, and determining the number of topics in the next iteration process according to the updated number of topics, the number of noise topics and the topic direction; repeating the above steps until the number of topics is unchanged or a preset number of iterations is reached, and outputting the number of topics in the text; The network constructing module is configured to construct a text relation network when the hotspot type is a bursty hotspot type, and divide the text relation network into a plurality of subtopic communities, the text relation network being composed of a plurality of sentence nodes and connecting edges between the sentence nodes; The scoring result determining module is configured to obtain a scoring result of each sentence according to a sentence weight of each sentence in each subtopic community and a contribution value of a word in each sentence to the sentence; The text abstract generating module is configured to select a target sentence from the text based on the scoring result, and generate a text abstract according to the target sentence.

7. An abstract extraction apparatus characterized by comprising: The abstract extraction device comprises a memory, a processor, and an abstract extraction program stored in the memory and running on the processor, and the abstract extraction program implements the steps of the abstract extraction method according to any one of claims 1-5 when executed by the processor.

8. A computer-readable storage medium, characterized in that, The computer readable storage medium stores an abstract extraction program, and the abstract extraction program implements the steps of the abstract extraction method according to any one of claims 1-5 when executed by the processor.

Images