Text abstract extraction method and device, storage medium and electronic device

Abstract

The invention discloses a text abstract extraction method and device, a storage medium and an electronic device, and relates to the technical field of information processing. The method comprises the steps that a tweet data set is obtained, and the tweet data set comprises tweet data related to a preset event; encoding the tweet data to obtain a plurality of tweet vectors; constructing a target tweet relation graph based on the plurality of tweet vectors; determining a significance score of the target tweet data based on the target tweet relation graph and a target model; and generating a target abstract based on the significance score and the tweet data. Through the abstract generation method and device, the technical problems that a large number of noisy and non-standard short texts on a social media platform cannot be processed due to the problems of information fragmentation, language informity, too large text quantity and the like in the prior art, the quality of generated abstracts is not high, and the requirements of users are difficult to meet are solved.

Description

Technical Field

[0001] The present application relates to the field of information processing technology, and in particular to a text summary extraction method and device, a storage medium and an electronic device. Background Art

[0002] In recent years, with the rapid development of the Internet, social networks have played an increasingly important role in people's daily lives and work. Social platforms such as Sina Weibo and WeChat public accounts have a huge number of posts per day. On these social platforms, users continue to generate various forms of content, making online social networks the main channel for expressing opinions, discussing public events, and sharing life.

[0003] In this context, the amount of information on social media is growing exponentially. Users are facing the problem of information overload and find it difficult to quickly identify and obtain key information. This situation has put forward new requirements and challenges for social media text summarization. There is an urgent need for a technology that can automatically and quickly extract core information from a large amount of short text. The research on social media text summarization, especially the summary extraction for specific topics, can not only help users save time in searching and reading and improve the efficiency of information acquisition, but also has important practical application value for business intelligence analysis, e-commerce strategy formulation and crisis response.

[0004] Traditional text summarization methods face many challenges, including fragmentation of information, informality of language, and excessive amount of text. These methods are often unable to handle the large amount of noisy and non-standard short text on social media platforms, resulting in low-quality summaries that are difficult to meet user needs.

[0005] Therefore, there is an urgent need for an innovative technology that can handle the specific text features of social media platforms to improve the accuracy and efficiency of text summarization. Summary of the invention

[0006] The main purpose of this application is to provide a text summary extraction method and device, a storage medium and an electronic device to solve the technical problem in the related art that due to the fragmentation of information, the informality of language and the excessive amount of text, a large amount of noisy and non-standard short texts on social media platforms cannot be processed, which leads to the low quality of the generated summary and difficulty in meeting the needs of users.

[0007] In order to achieve the above-mentioned purpose, according to one aspect of the present application, a text summary extraction method is provided. The method comprises: obtaining a tweet data set, wherein the tweet data set comprises tweet data related to a preset event; encoding the tweet data to obtain a plurality of tweet vectors; constructing a target tweet relationship graph based on the plurality of tweet vectors, wherein the target tweet relationship graph is used to reflect the data relationship of the tweet data and the tweet information volume of the tweet data; determining a significance score of the target tweet data based on the target tweet relationship graph and the target model, wherein the significance score is used to represent the probability of the tweet data being added to the target summary; generating a target summary based on the significance score and the tweet data, wherein the target summary is used to summarize and describe the preset event.

[0008] Furthermore, encoding the tweet data to obtain multiple tweet vectors includes: performing data processing on the tweet data to obtain first tweet data; performing word segmentation processing on the first tweet data to obtain second tweet data; encoding the second tweet data based on a preset model to determine a preset tag vector of an output layer of the preset model; and determining multiple tweet vectors based on the preset tag vector.

[0009] Furthermore, constructing a target tweet relationship graph based on multiple tweet vectors includes: creating a first tweet relationship graph, wherein the first tweet relationship graph includes multiple nodes, the multiple nodes are used to represent tweet data, and the multiple nodes include a first node and a second node; updating the first tweet relationship graph based on the text similarity between the tweet vectors of the first nodes to obtain a second tweet relationship graph, wherein the first nodes in the second tweet relationship graph are connected by line segments; updating the second tweet relationship graph based on the cosine similarity between the tweet vectors of the second nodes to obtain a third tweet relationship graph, wherein the second nodes in the third tweet relationship graph are connected by line segments; adjusting the third tweet relationship graph based on the semantic term scores of the multiple nodes to obtain a fourth tweet relationship graph; adjusting the fourth tweet relationship graph based on the target social signal to obtain the target tweet relationship graph.

[0010] Furthermore, determining the significance score of the target tweet data based on the target tweet relationship graph and the target model includes: determining the local neighborhood information of the target tweet relationship graph based on multiple graph convolutional layers of the target model; updating the node features of the target tweet relationship graph based on the local neighborhood information; determining the initial significance score of the target tweet data based on a preset scoring strategy and node features; and determining the significance score based on the target loss function and the initial significance score.

[0011] Furthermore, generating a target summary based on the significance score and tweet data includes: arranging the significance scores in descending order to obtain an arrangement result; determining the first tweet data with the highest score in the arrangement result; filtering the remaining tweet data except the first tweet data based on preset filtering rules to obtain second tweet data; and generating a target summary based on the first tweet data and the second tweet data.

[0012] Furthermore, based on preset filtering rules, the remaining tweet data except the first tweet data are filtered to obtain the second tweet data, including: determining the cosine similarity between the tweet data in the remaining tweet data and the first tweet data; determining the similarity percentage of the tweet data in the remaining tweet data in the remaining tweet data; and adding the tweet data whose cosine similarity is less than a first threshold and whose similarity percentage is greater than a second threshold to the second tweet data.

[0013] In order to achieve the above-mentioned purpose, according to another aspect of the present application, a text summary extraction device is provided. The device includes: an acquisition module, which is used to acquire a tweet data set, wherein the tweet data set includes tweet data related to a preset event; an encoding module, which is used to encode the tweet data to obtain multiple tweet vectors; a construction module, which is used to construct a target tweet relationship graph based on multiple tweet vectors, wherein the target tweet relationship graph is used to reflect the data relationship of the tweet data and the tweet information volume of the tweet data; a determination module, which is used to determine the significance score of the target tweet data based on the target tweet relationship graph and the target model, wherein the significance score is used to represent the probability of the tweet data being added to the target summary; and a generation module, which is used to generate a target summary based on the significance score and the tweet data, wherein the target summary is used to summarize and describe the preset event.

[0014] Furthermore, the encoding module is also used to process the tweet data to obtain first tweet data; perform word segmentation processing on the first tweet data to obtain second tweet data; encode the second tweet data based on a preset model to determine a preset tag vector of the output layer of the preset model; and determine multiple tweet vectors based on the preset tag vector.

[0015] Furthermore, the construction module is also used to create a first tweet relationship graph, wherein the first tweet relationship graph includes multiple nodes, the multiple nodes are used to represent tweet data, and the multiple nodes include a first node and a second node; the first tweet relationship graph is updated based on the text similarity between the tweet vectors of the first nodes to obtain a second tweet relationship graph, wherein the first nodes in the second tweet relationship graph are connected by line segments; the second tweet relationship graph is updated based on the cosine similarity between the tweet vectors of the second nodes to obtain a third tweet relationship graph, wherein the second nodes in the third tweet relationship graph are connected by line segments; the third tweet relationship graph is adjusted based on the semantic term scores of the multiple nodes to obtain a fourth tweet relationship graph; the fourth tweet relationship graph is adjusted based on the target social signal to obtain a target tweet relationship graph.

[0016] Furthermore, the determination module is also used to determine the local neighborhood information of the target tweet relationship graph based on multiple graph convolutional layers of the target model; update the node features of the target tweet relationship graph based on the local neighborhood information; determine the initial significance score of the target tweet data based on a preset scoring strategy and node features; determine the significance score based on the target loss function and the initial significance score.

[0017] Furthermore, the generation module is also used to arrange the significance scores in descending order to obtain an arrangement result; determine the first tweet data with the highest score in the arrangement result; filter the remaining tweet data except the first tweet data based on preset filtering rules to obtain second tweet data; and generate a target summary based on the first tweet data and the second tweet data.

[0018] Furthermore, the generation module is also used to determine the cosine similarity between the tweet data in the remaining tweet data and the first tweet data; determine the similarity percentage of the tweet data in the remaining tweet data in the remaining tweet data; and add the tweet data whose cosine similarity is less than a first threshold and whose similarity percentage is greater than a second threshold to the second tweet data.

[0019] In order to achieve the above objective, according to another aspect of the present application, a computer-readable storage medium is provided, wherein the storage medium stores a program, wherein the program executes any one of the above-mentioned text summary extraction methods.

[0020] In order to achieve the above-mentioned purpose, according to another aspect of the present application, an electronic device is provided, which includes one or more processors and a memory, wherein the memory is used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement any one of the above-mentioned text summary extraction methods.

[0021] Through this application, the following steps are adopted: obtaining a tweet dataset, wherein the tweet dataset includes tweet data related to a preset event; encoding the tweet data to obtain multiple tweet vectors; constructing a target tweet relationship graph based on multiple tweet vectors, wherein the target tweet relationship graph is used to reflect the data relationship of the tweet data and the tweet information volume of the tweet data; determining the significance score of the target tweet data based on the target tweet relationship graph and the target model, wherein the significance score is used to represent the probability of the tweet data being added to the target summary; generating a target summary based on the significance score and the tweet data, wherein the target summary is used to summarize and describe the preset event, solving the technical problem that the fragmentation of information, the informality of language, and the excessive amount of text in the related technology lead to the inability to process a large amount of noisy and non-standard short texts on the social media platform, thereby resulting in the generated summary being of low quality and difficult to meet the needs of users. By capturing and integrating the semantic information and social interaction relationships in social media texts, using graph convolutional networks and pre-trained models, information redundancy is reduced, the information of neighbor nodes is efficiently aggregated, and the feature representation of each node is updated, thereby improving the efficiency and quality of summary extraction. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0023] Figure 1 is a flowchart of a text summary extraction method provided according to an embodiment of the present application;

[0024] Figure 2 is a flow chart of a graph convolutional network social media summary extraction method based on tweet relations in this embodiment;

[0025] Figure 3 is a schematic diagram of constructing a tweet relationship graph in this embodiment;

[0026] Figure 4 is a schematic diagram of a text summary extraction device provided according to an embodiment of the present application;

[0027] Figure 5 is a schematic diagram of an electronic device provided according to an embodiment of the present application. DETAILED DESCRIPTION

[0028] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0029] In order to enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this application.

[0030] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequential order. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present application described here. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0031] It should be noted that the relevant information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for display, data for analysis, etc.) involved in this disclosure are all information and data authorized by the user or fully authorized by all parties. For example, an interface is set between this system and the relevant user or organization. Before obtaining relevant information, it is necessary to send an acquisition request to the aforementioned user or organization through the interface, and obtain relevant information after receiving the consent information fed back by the aforementioned user or organization.

[0032] The present invention is described below in conjunction with preferred implementation steps. Figure 1 is a flowchart of a text summary extraction method provided in an embodiment of the present application. Figure 1 As shown, the method comprises the following steps:

[0033] Step S10: Acquire a tweet data set, wherein the tweet data set includes tweet data related to a preset event.

[0034] In the embodiment of the present invention, the tweet dataset can be understood as a set of social media tweets collected and related to a specific event or topic. For example, a set of tweets related to a specific event is defined as an event cluster D, and the event cluster D is the tweet dataset, which is not limited here.

[0035] A preset event can be understood as a defined, specific event or topic, and the preset event can be the focus of social media text summary extraction. For example, the preset event can be any event that attracts public attention, such as news events, social hot spots, sports competitions, product launches, entertainment news, etc., or a specific topic of interest to users, such as technological innovation, environmental protection initiatives, healthy lifestyles, etc., which are not limited here.

[0036] Obtaining a tweet dataset, where the tweet dataset includes tweet data related to a preset event can be understood as first determining a preset event, and then collecting tweets related to the event from a social media platform according to the preset event.

[0037] In the embodiment of the present invention, by defining preset events and collecting relevant tweet data, basic data is provided for subsequent processing steps.

[0038] Step S11, encoding the tweet data to obtain multiple tweet vectors.

[0039] In the embodiment of the present invention, the tweet vector can be understood as a tweet representation, which is used to reflect the semantic similarities and differences between tweets, and is not limited here.

[0040] Encoding the tweet data to obtain multiple tweet vectors can be understood as converting the text content of each tweet into a form of numerical representation, obtaining multiple vectors in a high-dimensional space, i.e., multiple tweet vectors. For example, the bidirectional encoder representation (Bidirectional Encoder Representations from Transformers, BERT) model in the Transformer model can be used to encode the tweet data, i.e., the text content of each tweet is input into the BERT model, and the model outputs a vector representation that contains the semantic information of the tweet and the relevance of the context. Finally, the output of the BERT model is a multi-layer vector representation, i.e., multiple tweet vectors, which are not limited here.

[0041] In an embodiment of the present invention, by converting tweet data into vector representation, numerical input can be provided for subsequent graph construction and graph convolutional network learning, so that the algorithm can perform analysis based on the semantic and structural information of the tweets, and ultimately generate high-quality social media text summaries.

[0042] Step S12: construct a target tweet relationship graph based on the multiple tweet vectors, wherein the target tweet relationship graph is used to reflect the data relationship of the tweet data and the tweet information volume of the tweet data.

[0043] In the embodiment of the present invention, the target tweet relationship graph can be understood as a graph data structure, where nodes represent tweets, edges represent the associations between tweets, and the target tweet relationship graph is also the tweet relationship graph after optimization. Exemplarily, the target tweet relationship graph is not only a static text similarity network, but also includes social interaction information between tweets, such as forwarding, commenting, and replying, which is not limited here.

[0044] The data relationship of tweet data can be understood as the mutual connection and influence between tweets. For example, the data relationship of tweet data is mainly reflected in text similarity, social interaction, and time series relationship, which are not limited here.

[0045] The tweet information volume of the tweet data can be understood as the amount of valuable information contained in each tweet. The tweet information volume of the tweet data is a key indicator for evaluating the representativeness and importance of tweets in event summaries. For example, the tweet information volume of the tweet data can be calculated by semantic term scores or social signals, which are not limited here.

[0046] Constructing a target tweet relationship graph based on multiple tweet vectors can be understood as constructing a graph structure that can reflect the data relationship between tweets (including text similarity and social interaction) and the amount of information of each tweet by analyzing tweet vectors, which is not limited here.

[0047] The target tweet relationship graph is used to reflect the data relationship of tweet data and the amount of tweet information of tweet data. It can be understood as the target tweet relationship graph, which not only contains the text information of tweets, but also reflects the social relationship between tweets and the potential amount of information of each tweet, providing a richer data structure for subsequent tweet saliency evaluation and summary extraction.

[0048] In the embodiment of the present invention, the semantic information of tweets and the structural information of social networks are integrated into a unified graph representation, thereby providing a framework that can simultaneously reflect the content of tweets and social interactions for subsequent model training and summary generation. The node and edge weights in the target tweet relationship graph not only represent the text similarity of tweets, but also integrate the social influence and information density of tweets, so that the graph convolutional network can more comprehensively understand the context and importance of social media texts.

[0049] Step S13, determining a significance score of the target tweet data based on the target tweet relationship graph and the target model, wherein the significance score is used to represent the probability of the tweet data being added to the target summary.

[0050] In the embodiment of the present invention, the target model can be understood as a trained graph convolutional network (GCN) model, which is used to process the target tweet relationship graph and learn the high-level feature representation of tweet nodes therefrom, and then calculate a significance score for each tweet to reflect its importance or probability when generating the target summary, which is not limited here.

[0051] The significance score of the target tweet data can be understood as the learning result of the target model based on the tweet relationship graph, which measures the importance and influence of each tweet in the event cluster.

[0052] Determining the significance score of the target tweet data based on the target tweet relationship graph and the target model can be understood as processing the target tweet relationship graph through the target model (graph convolutional network GCN), thereby calculating a significance score for each tweet data, which is not limited here.

[0053] The significance score is used to indicate the probability of the tweet data being added to the target summary. It can be understood that the significance score reflects the probability of the tweet being added to the target summary, that is, the importance and uniqueness of the tweet in the event, which is not limited here.

[0054] In the embodiment of the present invention, the target model (graph convolutional network GCN) can capture the complex relationships between tweets, and the significance score reflects the importance of these relationships and the characteristics of the tweets themselves, which together provide key data support for generating high-quality social media text summaries.

[0055] Step S14, generating a target summary based on the significance score and the tweet data, wherein the target summary is used to summarize and describe the preset event.

[0056] In the embodiment of the present invention, the target summary can be understood as a refined text that summarizes the preset event by selecting the most representative and informative tweet set based on the significance score and tweet data. For example, a set of representative tweets S is selected from the event cluster D so that S can maximize the coverage of the information in D, that is:

[0057]

[0058] Where S′ is any subset of tweets selected from D, and S is the optimal subset that can maximize the coverage of information in D, that is, S is the target summary. The function Coverage(S', D) is used to measure the degree of information coverage of the selected tweet set S' on the event cluster D. Ideally, S should contain tweets that best represent the core content of the event while minimizing information redundancy, which is not restricted here.

[0059] Generating a target summary based on the saliency score and tweet data can be understood as utilizing the saliency score and tweet data to generate a concise and informative target summary to summarize the preset event.

[0060] The target summary is used to summarize the preset event. It can be understood that the target summary is a highly condensed description of all relevant tweet information of the preset event, which is not limited here.

[0061] In the embodiment of the present invention, by using the significance score and tweet data to generate the target summary, not only the accuracy and efficiency of information extraction are improved, but also the comprehensiveness and user-friendliness of the summary content are ensured, providing a powerful solution for processing the problem of social media text summarization.

[0062] By defining preset events and collecting relevant tweet data, basic data is provided for subsequent processing steps. By converting tweet data into vector representation, numerical input is provided for subsequent graph construction and graph convolutional network learning, so that the algorithm can analyze based on the semantic and structural information of tweets and finally generate high-quality social media text summaries. The semantic information of tweets and the structural information of social networks are integrated into a unified graph representation, thus providing a framework that can simultaneously reflect tweet content and social interaction for subsequent model training and summary generation. The node and edge weights in the target tweet relationship graph not only represent the text similarity of tweets, but also integrate the social influence and information density of tweets, so that the graph convolutional network can more comprehensively understand the context and importance of social media text. The target model (graph convolutional network GCN) can capture the complex relationships between tweets, and the significance score reflects the importance of these relationships and the characteristics of the tweets themselves, which together provide key data support for generating high-quality social media text summaries. By using the significance score and tweet data to generate the target summary, not only the accuracy and efficiency of information extraction are improved, but also the comprehensiveness and user-friendliness of the summary content are ensured, providing a powerful solution for dealing with the problem of social media text summarization.

[0063] Optionally, in step S11, encoding the tweet data to obtain multiple tweet vectors may include the following steps:

[0064] Step S111, processing the tweet data to obtain first tweet data;

[0065] Step S112, performing word segmentation processing on the first tweet data to obtain second tweet data;

[0066] Step S113, encoding the second tweet data respectively based on the preset model, and determining a preset tag vector of an output layer of the preset model;

[0067] Step S114, determining a plurality of tweet vectors based on the preset tag vector.

[0068] In the embodiment of the present invention, data processing can be understood as preprocessing the original tweet data, which may include removing irrelevant information (such as URL links, tags, emoticons, etc.), standardizing text (such as converting all letters to lowercase), and possible noise removal operations, which are not limited here.

[0069] The first tweet data can be understood as text data that has been preliminarily cleaned and formatted, and is not limited here.

[0070] Word segmentation can be understood as a basic step in natural language processing, which is used to segment continuous text into meaningful words or phrases, and is not limited here.

[0071] The second tweet data may be understood as text data that has been decomposed into vocabulary units, which is not limited here.

[0072] The preset model can be understood as a pre-trained deep learning model, such as the BERT model, which is not limited here.

[0073] The preset tag vector of the output layer can be understood as a special tag [CLS] in the BERT model, which usually appears at the beginning of the input sequence. The [CLS] tag vector is a vector representation of the tweet encoded by the BERT model. It captures the semantic features of the tweet and is the basis for the subsequent construction of the tweet relationship graph and graph convolutional network learning. It is not limited here.

[0074] Processing the tweet data to obtain the first tweet data can be understood as performing preliminary data cleaning and preprocessing on the original tweet data to remove unnecessary information and formatted text, thereby obtaining the first tweet data.

[0075] Performing word segmentation processing on the first tweet data to obtain the second tweet data can be understood as dividing the text of the first tweet data into individual words or vocabulary units, and the individual words or vocabulary units are the second tweet data.

[0076] The second tweet data is encoded based on the preset model. Determining the preset tag vector of the output layer of the preset model can be understood as inputting the word-segmented vocabulary unit sequence into the BERT model. The BERT model will encode the tweet data after word segmentation, and then the vector corresponding to the [CLS] tag in the output layer is extracted as the semantic representation of the entire tweet.

[0077] Determining multiple tweet vectors based on preset tag vectors can be understood as each tweet undergoes the same preprocessing, word segmentation and encoding process, and finally generates a unique [CLS] tag vector. All [CLS] tag vectors are collected to form a vector set, thereby determining multiple tweet vectors.

[0078] In the embodiment of the present invention, the quality of data can be significantly improved by cleaning and standardizing data, removing irrelevant information, and standardizing the text format. The text is divided into independent vocabulary units, and the vocabulary units are encoded using a pre-trained deep learning model (such as BERT) and converted into a numerical vector representation. Based on the preset tag vector output by the pre-trained model, the vector representation of each tweet is determined, which helps the model to more accurately understand the text content, capture social interaction relationships, and ultimately generate concise and comprehensive summary information.

[0079] Optionally, in step S12, constructing a target tweet relationship graph based on multiple tweet vectors may include the following steps:

[0080] Step S121, creating a first tweet relationship graph, wherein the first tweet relationship graph includes a plurality of nodes, the plurality of nodes are used to represent tweet data, and the plurality of nodes include a first node and a second node;

[0081] Step S122, updating the first tweet relationship graph based on the text similarity between the tweet vectors of the first nodes to obtain a second tweet relationship graph, wherein the first nodes in the second tweet relationship graph are connected by line segments;

[0082] Step S123, updating the second tweet relationship graph based on the cosine similarity between the tweet vectors of the second nodes to obtain a third tweet relationship graph, wherein the second nodes in the third tweet relationship graph are connected by line segments;

[0083] Step S124, adjusting the third tweet relationship graph based on the semantic term scores of the multiple nodes to obtain a fourth tweet relationship graph;

[0084] Step S125: adjusting the fourth tweet relationship graph based on the target social signal to obtain a target tweet relationship graph.

[0085] In the embodiment of the present invention, the first tweet relationship graph can be understood as the initial state of constructing the tweet relationship graph, and the nodes in the graph represent tweet data on the social media platform, that is, each tweet is represented as a node in the graph. The nodes in the graph have not yet been connected according to the similarity or social interaction relationship between the tweets.

[0086] The first node can be understood as a source tweet node, that is, a starting tweet or a key tweet in an event cluster.

[0087] The second node can be understood as a non-source tweet node, which can respond, forward, or comment on the source tweet, constituting the rest of the social media discussion.

[0088] Text similarity can be understood as a measure of the directional consistency of two vectors obtained by calculating the cosine similarity between tweet vectors.

[0089] The second tweet relationship graph can be understood as a relationship graph that is optimized based on the first tweet relationship graph, in which connections are established between source tweets (first nodes) based on text similarity.

[0090] The third tweet relationship graph can be understood as a more comprehensive graph structure formed by adjusting the connections between non-source tweets (second nodes) based on the second tweet relationship graph according to text similarity.

[0091] The semantic term score can be understood as the score of key semantic information in a tweet (such as entity, location, time expression, etc.), which is used to evaluate the information content and importance of a tweet.

[0092] The fourth tweet relationship graph can be understood as a relationship graph obtained by adjusting the weights of the edges by using the semantic term scores of the nodes on the basis of the third tweet relationship graph.

[0093] The target social signal can be understood as specific social interaction information, such as the number of reposts and replies of a tweet, etc., which is not limited here.

[0094] A first tweet relationship graph is created, wherein the first tweet relationship graph includes multiple nodes, the multiple nodes are used to represent tweet data, and the multiple nodes include a first node and a second node. It can be understood that an empty graph structure, namely the first tweet relationship graph, is created, and the first tweet relationship graph will be used to represent the relationship between tweets. Each tweet will correspond to a node in the graph. The first node usually refers to the source tweet node, that is, the starting tweet or core tweet in the event cluster, which may be the source of the discussion, the proposal of key ideas, etc. The second node is a non-source tweet node, including responses, forwarding, comments, etc. to the first node.

[0095] The first tweet relationship graph is updated based on the text similarity between the tweet vectors of the first nodes to obtain the second tweet relationship graph, where the first nodes in the second tweet relationship graph are connected by line segments, which can be understood as using cosine similarity to calculate the similarity between the tweet vectors of the first nodes (i.e., source tweet nodes). In the first tweet relationship graph, the nodes already exist but have not yet been connected. Based on the calculated text similarity, we will create connections, i.e., edges, between the source tweet nodes. These edges not only represent the association between the source tweets, but also the weights of the edges will be determined according to the similarity of the source tweet vectors. After creating the edges based on text similarity, the second tweet relationship graph is obtained.

[0096] The second tweet relationship graph is updated based on the cosine similarity between the tweet vectors of the second nodes to obtain a third tweet relationship graph, wherein the second nodes in the third tweet relationship graph are connected by line segments, which can be understood as calculating the cosine similarity between the non-source tweet vectors to quantify the semantic similarity between the non-source tweet vectors. The second tweet relationship graph is further updated using the cosine similarity between the non-source tweet vectors, that is, the relationship between the non-source tweet nodes is represented by line segments, and the weight of the edge is determined according to the cosine similarity. The higher the similarity, the greater the weight of the edge. After the update, the third tweet relationship graph is obtained.

[0097] The third tweet relationship graph is adjusted based on the semantic term scores of multiple nodes to obtain the fourth tweet relationship graph. It can be understood that by introducing the semantic term scores and adjusting the edge weights, the edge weights not only include the text similarity information between tweets, but also incorporate the semantic term scores of the tweets. This allows the graph to more accurately reflect the importance of the tweets and their roles in the events. After adjusting the edge weights, the fourth tweet relationship graph is obtained.

[0098] Adjusting the fourth tweet relationship graph based on the target social signal to obtain the target tweet relationship graph can be understood as adjusting the weight of each edge in the fourth tweet relationship graph according to the social signals of the two connected nodes (tweets). For example, if there is a high frequency of forwarding or replying interactions between two tweets connected by an edge, the weight of this edge will be increased, thereby highlighting these high-influence tweet relationships in the graph. By adjusting the fourth tweet relationship graph based on social signals, the target tweet relationship graph is finally obtained.

[0099] In an embodiment of the present invention, a first tweet relationship graph is first created, and the semantic relevance between the source tweet (first node) and the non-source tweet (second node) in the graph is enhanced by calculating the text similarity and cosine similarity between the tweet vectors. Semantic term scores are introduced to quantify the information value of nodes (i.e., tweets). This adjustment enables the graph structure to reflect the uniqueness and importance of tweets in event discussions, avoiding the limitations of text similarity alone. The target social signals, such as the number of reposts and replies, are further considered, and the graph structure is adjusted to finally obtain the target tweet relationship graph, which greatly improves the quality and efficiency of summary generation.

[0100] Optionally, in step S13, determining the significance score of the target tweet data based on the target tweet relationship graph and the target model may include the following steps:

[0101] Step S131, determining local neighborhood information of a target tweet relationship graph based on multiple graph convolutional layers of the target model;

[0102] Step S132, updating the node features of the target tweet relationship graph based on the local neighborhood information;

[0103] Step S133, determining an initial significance score of the target tweet data based on a preset scoring strategy and node features;

[0104] Step S134: determining a significance score based on the target loss function and the initial significance score.

[0105] In the embodiment of the present invention, local neighborhood information can be understood as information about neighboring nodes around each node (tweets) in the target tweet relationship graph. For example, in a graph convolutional network (GCN), when performing information aggregation, each node considers the features of its directly connected neighboring nodes and the feature information of indirect neighboring nodes propagated through multiple hops (multi-layer GCN). The above information together constitutes the local neighborhood information of the node, which is not limited here.

[0106] The node features of the target tweet relationship graph can be understood as the vector information carried by each tweet node in the graph that represents its content and social attributes. For example, local neighborhood information is input into the graph convolution layer, and a more advanced node feature representation is generated through multi-layer information aggregation and updating, which is not limited here.

[0107] The preset scoring strategy can be understood as a rule or function for calculating the significance score of a tweet, which combines node features and specific scoring criteria. The scoring strategy may include different combinations and weightings of node features to quantify the importance and uniqueness of the tweet in the event cluster. For example, the scoring strategy may consider the semantic information content, social signals, and correlation with other tweets of the tweet, which are not limited here.

[0108] The initial saliency score can be understood as a preliminary score obtained by evaluating the node features of the target tweet relationship graph according to a preset scoring strategy. The initial saliency score reflects the importance of the tweet without fully considering the global structure of the social network and the constraints of summary optimization.

[0109] The target loss function can be understood as a function used to evaluate the difference between the predicted saliency score and the true saliency score during model training. During the training process, the target loss function guides the optimization of model parameters to minimize the gap between the predicted score and the true score, thereby improving the accuracy and quality of summary extraction.

[0110] Determining the local neighborhood information of the target tweet relationship graph based on multiple graph convolutional layers of the target model can be understood as processing the target tweet relationship graph using the graph convolutional layer in the Graph Neural Network (GNN), and determining the local neighborhood information of each node by aggregating the features of the direct neighbor nodes of each node (i.e. tweet) and the multi-hop neighbor nodes in the graph. This process includes information propagation and node feature update, so that the representation of each tweet node not only contains its own text and social signal information, but also integrates the information of other tweet nodes associated with it, forming a more comprehensive and in-depth node feature representation.

[0111] Updating the node features of the target tweet relationship graph based on local neighborhood information can be understood as using local neighborhood information to update the feature representation of each node in the target tweet relationship graph. This process is part of the graph convolutional network learning phase. By iteratively updating node features, the model can gradually deepen its understanding of the relationship between tweets and the evaluation of the amount of information in tweets.

[0112] Determining the initial significance score of the target tweet data based on the preset scoring strategy and node features can be understood as applying the preset scoring strategy to calculate the initial significance score of the target tweet data. The scoring strategy may include a linear combination of node features, weight adjustment, or a specific scoring function, the purpose of which is to evaluate the importance and representativeness of each tweet in the event discussion.

[0113] Determining the significance score based on the target loss function and the initial significance score can be understood as using the target loss function to optimize the model parameters to improve the accuracy of the score. The target loss function measures the difference between the significance score predicted by the model and the manually annotated significance score. By minimizing this loss function, the model can learn how to more accurately evaluate the importance of tweets, thereby generating a predicted summary that is closer to the real summary.

[0114] In an embodiment of the present invention, by gradually and deeply analyzing the information in the tweet relationship graph, from the determination of local neighborhood information to the updating of node features, and then to the calculation and optimization of significance scores, a high-quality text summary is finally generated, so that the summary can not only cover the key information of the event, but also reflect the social value and influence of the tweet in the social media environment.

[0115] Optionally, in step S14, generating a target summary based on the significance score and tweet data may include the following steps:

[0116] Step S141, arranging the significance scores in descending order to obtain an arrangement result;

[0117] Step S142, determining the first tweet data with the highest score in the arrangement result;

[0118] Step S143, filtering the remaining tweet data except the first tweet data based on a preset filtering rule to obtain second tweet data;

[0119] Step S144: Generate a target summary based on the first tweet data and the second tweet data.

[0120] In the embodiment of the present invention, the significance scores are arranged in descending order, and the arrangement result can be understood as sorting the significance scores of all tweets from high to low, thereby obtaining an arrangement result. The purpose of descending order is to give priority to tweets with higher significance scores, which are considered to carry more important and unique information in the event discussion. The arrangement result provides a clear priority list for the summary extraction algorithm, guiding it to build a summary starting from the most important tweets.

[0121] Determining the first tweet data with the highest score in the ranking results can be understood as selecting the tweet with the highest score as the first tweet data in the descending ranking results. This tweet is regarded as the most representative and informative tweet in the event cluster, and usually contains key information of the event or the focus of discussion.

[0122] The second tweet data can be obtained by filtering the remaining tweet data except the first tweet data based on the preset filtering rules. It can be understood that after determining the first tweet data, the most representative tweets are further filtered out from the remaining tweets according to the preset filtering rules, i.e., the second tweet data. The preset filtering rules may include requirements for the diversity of tweet content, relevance to the first tweet data, or strategies to avoid redundant information. The filtering process ensures that the summary not only covers the key aspects of the event, but also avoids information duplication and provides complementary perspectives and details.

[0123] Generating a target summary based on the first tweet data and the second tweet data can be understood as combining the first tweet data with the filtered second tweet data to generate a target summary. The target summary contains the core information of the event and supplementary details from multiple angles, which not only reflects the overall picture of the event, but also avoids information redundancy, thereby improving the readability and information density of the summary.

[0124] In the embodiment of the present invention, the significance scores are arranged in descending order to obtain an arrangement result, the first tweet data with the highest score in the arrangement result is determined, the remaining tweet data except the first tweet data is filtered based on a preset filtering rule to obtain the second tweet data, and a high-quality summary that can meet the user's needs is generated based on the first tweet data and the second tweet data. By sorting and filtering the significance scores, it is ensured that the construction of the summary is based on the importance of the tweets while taking into account the diversity and integrity of the content.

[0125] Optionally, in step S143, filtering the remaining tweet data except the first tweet data based on a preset filtering rule to obtain the second tweet data may include the following steps:

[0126] Determine the cosine similarity between the tweet data in the remaining tweet data and the first tweet data;

[0127] determining a similarity percentage of the tweet data in the rest of the tweet data to the rest of the tweet data;

[0128] The tweet data whose cosine similarity is less than the first threshold and whose similarity percentage is greater than the second threshold are added to the second tweet data.

[0129] In the embodiment of the present invention, the first threshold may be understood as a preset cosine similarity threshold, which is used to filter tweets having a lower similarity with the first tweet data.

[0130] The second threshold can be understood as a preset similarity percentage threshold, which is used to measure the representativeness of the tweet and its influence in the event discussion.

[0131] Determining the cosine similarity between the tweet data in the remaining tweet data and the first tweet data can be understood as the algorithm using the cosine similarity metric to calculate the text similarity between all the remaining tweet data except the selected first tweet data and the first tweet data.

[0132] Determining the similarity percentage of a tweet in the rest of the tweet data can be understood as evaluating the distribution of similarities of each remaining tweet in the overall rest of the tweet data. By calculating the similarity percentage, the algorithm can determine whether a tweet is representative of the discussion, that is, whether it generally reflects a certain aspect of the event or has relevant opinions. A high similarity percentage of a tweet indicates that it may capture the common points of most tweets, thus helping the summary cover a wider range of discussion content.

[0133] Adding tweet data with a cosine similarity less than a first threshold and a similarity percentage greater than a second threshold to the second tweet data can be understood as, for each remaining tweet, only when the cosine similarity of each remaining tweet with the first tweet data is lower than the first threshold (i.e., it is semantically significantly different from the information already included in the summary), and the similarity percentage in all remaining tweets is higher than the second threshold (i.e., it is highly representative in the remaining tweets and can reflect the general focus or core information of the event), the tweet will be selected and added to the second tweet data set.

[0134] In the embodiment of the present invention, through the above steps, the purpose is to select tweets that can reflect the core of the event and maintain the freshness and diversity of the summary from the massive information of social media. By using cosine similarity and similarity percentage in combination, the algorithm can more accurately identify tweets that are both unique and representative, thereby creating high-quality summary content.

[0135] The present invention proposes a text summary extraction method, which can be divided into four parts: graph construction, model architecture, model training, and tweet selection. Among them, the construction of the graph includes generating tweet representations and constructing a tweet relationship graph. Through the above steps, a tweet relationship graph reflecting the relationship between tweets and the amount of tweet information is constructed, which will be used for subsequent graph convolutional network processing and tweet saliency estimation. This construction process takes into account the text information and social signals between tweets. The model architecture consists of three key layers: input layer, graph convolution layer, and saliency evaluation layer. The input layer is responsible for receiving the tweet relationship graph designed in the previous article as a sample to provide basic input data for subsequent graph learning. The graph convolution layer is composed of multiple graph convolutional network (GCN) layers, which aims to learn the high-level feature representation of nodes in the tweet relationship graph. Each GCN layer updates its feature representation by considering the neighbor information of the node, thereby effectively capturing the structural relationship between tweets. Based on the graph convolution layer, the saliency evaluation layer calculates a saliency score for each tweet node, which will be used to guide the subsequent tweet selection process. The goal of model training is to optimize the parameters by minimizing the loss function, thereby improving the accuracy of the model in predicting tweets. Through training, the model can learn effective parameters, thus achieving better performance in social media text summarization tasks. Some events may require more than one tweet to be selected as a summary. In this case, each tweet represents a different aspect or development stage of the event. The number of tweets selected as summaries is dynamic, depending on the similarity between tweets and their respective representativeness, aiming to ensure that the generated summary can comprehensively cover the key information of the event while avoiding duplication of content.

[0136] Figure 2 is a flow chart of the graph convolutional network social media summary extraction method based on tweet relations in this embodiment. Figure 2 As shown in the figure, it is mainly divided into generating tweet representations, building a tweet relationship graph, graph convolutional network (GCN) learning stage, and selecting a set of optimal tweets as the target summary. Specifically, the BERT model is used to convert each tweet into a vector form, providing a basis for subsequent graph construction and feature learning. The connection between nodes is determined based on the text similarity and social interaction relationship between tweets, and the weight of the edge is set. The graph convolutional network (GCN) updates the feature representation of each node by aggregating the feature information of neighboring nodes, thereby capturing the structural relationship between tweets. On this basis, the algorithm calculates the significance score of each tweet, reflecting its importance in the summary. The algorithm selects a set of tweets with the highest score as the event summary based on the significance score. The selection process may take into account the diversity between tweets to ensure that the summary covers different aspects of the event.

[0137] In order to better understand the technical solution of the present application, the implementation process of the text summary extraction method described in the present invention is now described.

[0138] First, the main process of generating tweet representations is given. The collected dataset contains 30 social topics, each of which consists of four parts: tweet content, interactive relationships between tweets, user information, and manually created expert summaries. Table 1 shows the detailed statistics of each part of four events.

[0139] Table 1

[0140]

[0141] The content of tweets includes the text information of the tweet and its basic data (such as forwarding, likes, comments, etc.), and the average length of all 30 events is 16.67 words. The first two rows of the table show the number of tweets in each event and the average length of each tweet. Unlike other summary datasets, a unique feature of the dataset of this method is the interactive relationship between tweets. The forwarding, comments and corresponding relationships of tweets in each event are collected to construct social networks, and the statistical characteristics of these generated networks are further analyzed.

[0142] Each tweet is encoded using a pre-trained BERT language model to generate a vector representation of the tweet. The BERT model is based on a multi-layer bidirectional Transformer and is used for specific natural language understanding tasks by training on a large amount of text data. In this model, given a tweet, its text is tokenized using WordPiece after removing the Uniform Resource Locator (URL) and retweet mentions, and is input into the BERT model. The CLS token vector of the final output layer is used as the representation of the tweet.

[0143] Secondly, the steps for constructing the tweet relationship graph are introduced. Figure 3 is a schematic diagram of constructing a tweet relationship graph in this embodiment, such as Figure 3 As shown in the figure, a set of topic event clusters are first input. Each tweet is first encoded through the pre-trained BERT model to generate the initial feature vector of the node. Based on the node features processed by BERT, the text similarity between tweets is calculated. Using the calculated similarity value, a tweet relationship graph that only considers text information is constructed. At the same time, the process also processes social signals related to topic event clusters, such as forwarding, comments, and likes. On the tweet relationship graph based on text information, additional edges are added to the nodes in the graph according to social signals, and the weights of the edges are updated. After the above steps, a graph structure that can simultaneously reflect the content of tweets and social interactions is constructed, which provides rich information for subsequent graph convolutional network learning, helps the algorithm to more accurately evaluate the significance of tweets, and finally generates high-quality social media text summaries.

[0144] Specifically, create an empty graph where each node represents a tweet. Suppose we consider the text similarity information between tweets. Specifically, the i-th tweet is x i , whose corresponding tweet is represented by v i To ensure the connectivity of the graph, we need to set an edge between all source tweets and use the vector representation of the tweets to calculate the cosine similarity between two tweets. This similarity value is used as the initial weight of the edge. The formula for calculating cosine similarity is shown in formula (2):

[0145]

[0146] For all non-source tweet nodes, calculate the cosine similarity sim of the tweet vector representations of the two ij If sim ij is greater than the preset threshold θ, then at node v i and node v j Then, for all nodes in the tweet relationship graph, the semantic term score s(x i ), which reflects the amount of information in a tweet. The above steps are performed by considering the semantic terms contained in the tweet (such as entities, places, time expressions, etc., which are determined by the events corresponding to the event cluster). Specifically, a method such as linear regression is used to calculate a score for each tweet based on these semantic terms. Then, the node score is used to update the edge weight. For the edge from node x i To node x j The directed edge of text (x i , x j ) is calculated as shown in formula (3):

[0147]

[0148] Among them, w(x i , x j ) is the original edge weight (i.e., cosine similarity), s(x i ) is the node x i The denominator is the score of node x i The sum of similar calculations for all outgoing edges is calculated. This update process ensures that the edge weights not only reflect text similarity but also take into account the amount of information in the tweet. Finally, all edges are converted to undirected edges, and the edge weights are updated to the average of the weights in both directions. Then, all edge weights are rescaled to between 0 and 1 for subsequent processing. Finally, a tweet relationship graph that only considers text information is generated.

[0149] For social signals between tweets. For each tweet, refer to its forwarded and replied information, including which tweets are replies to the tweet and which other tweets the tweet forwarded. Add edges to the nodes in the graph based on the collected social signal data. Specifically: if tweet A is a reply to tweet B, add a directed edge between tweet A and tweet B, from A to B; if tweet A forwards tweet B, add a directed edge between tweet A and tweet B, from A to B.

[0150] In social media, forwarding and replying are usually triggered by the behavior of the departure node (i.e., the tweet that initiates the forwarding or replying). Therefore, considering the weight of the edge from the departure node is more in line with the actual situation and can intuitively reflect the influence of the departure node on the target node. And in the information propagation model, information is usually propagated from one node to another. Therefore, when constructing a tweet relationship graph, the weight of the edge usually reflects the ability or strength of information propagation from the departure node to the target node, rather than the reverse influence. The present invention believes that the weight of the edge should be considered from the departure node of the edge. For the edge from node x i To node x j The weight update formula of the directed edge is shown in formula (4) and formula (5):

[0151]

[0152] w new (x i , x j )=β×w social (x i , x j )+(1-β)w text (x i , x j ) Formula (5)

[0153] Among them, n retweet is node x i Represents the number of retweets of the tweet, n reply is node x i Represents the number of replies to a tweet, RT max is the maximum number of retweets among all tweets, RP max is the maximum number of replies in all tweets, and α is the weight coefficient, which is used to adjust the respective roles of the number of forwarding and the number of replies in the edge weight update. The edge weight setting method focuses on the social signals between tweets, and the dimension of the weight is standardized using the normalization method.

[0154] Next, the algorithm enters the graph convolutional network (GCN) learning phase. GCN updates the feature representation of each node by aggregating the feature information of neighboring nodes, thereby capturing the structural relationship between tweets. In order to consider a wider range of contextual information to judge the importance of tweets, a three-layer GCN is used, which means that the feature representation of each node will be affected by its three-hop neighbors (that is, nodes separated by three edges). GCN learns the transformation function f(X, E) of each node by aggregating the features of neighboring nodes to generate hidden features. This function uses the following layer-by-layer propagation rule, namely formula (6):

[0155]

[0156] Among them, H (l) represents the node feature matrix of the lth layer, where each row represents the feature vector of a node. Initially, H (0) Typically, it is the raw feature matrix of the node. For tweet nodes, these features are the tweet embeddings obtained by the BERT model. represents the adjacency matrix with self-connection added. is a diagonal matrix, representing the adjacency matrix with self-connection The degree matrix of is equal to the degree of node i (including self-connections), that is W(l) represents the weight matrix of the lth layer, which is a parameter that needs to be learned through training. σ represents the activation function, which is used to introduce nonlinearity. This method uses the ReLU function.

[0157] Through the above layer-by-layer propagation rules, GCN can update the feature representation of nodes by aggregating the features of neighboring nodes at each layer, thereby capturing the local neighborhood information of nodes in the graph structure. This feature aggregation mechanism enables GCN to effectively learn the representation of graph structured data.

[0158] On the basis of the graph convolution layer, this layer calculates a significance score for each tweet node, which will be used to guide the subsequent tweet selection process. The GCN component learns new tweet embeddings. In order to obtain the global context of the event, a cluster-level embedding C is generated by averaging the embeddings of all tweets in the cluster. Then, the cluster embedding C and the tweet embedding Y are used to extract the global context of the event. i To calculate tweet X i The significance score S i , as shown in formula (7):

[0159] S u =f(x i )=v T tanh(W s Y i +W c C) Formula (7)

[0160] Among them, W s , W C and v are parameters. S i This is then normalized by ranking all tweets in the cluster.

[0161] The loss function used by the model is the cross-entropy loss, which is a loss function commonly used in classification problems to measure the difference between the probability distribution predicted by the model and the true distribution. The goal of the loss function is to minimize the difference between the predicted tweet significance score and the true annotation (here the annotation based on the ROUGE-1 score). Specifically, for each tweet in an event cluster, the model predicts a significance score that represents the probability that the tweet is selected as a summary. These scores are normalized by the softmax function to ensure that their sum is 1, thus forming a probability distribution. Then, the cross-entropy loss function is used to compare this predicted probability distribution with the true distribution (that is, the distribution corresponding to the manually annotated summary).

[0162] The formula of the loss function can be expressed as formula (8):

[0163]

[0164] Where D is the set of all event clusters, d is an event cluster among all event clusters, and x i is a tweet in the corresponding event cluster d, f(x i ) is the ranked and normalized significance score predicted by the model, for each tweet x in the event cluster i Calculate the ROUGE-1 score between it and the manually annotated reference summary as Rouge_1(x i ). This score reflects the i As the quality of the summary. By minimizing this loss function, the model learns how to predict the saliency of tweets based on their content and their relationship in event clusters, thereby generating more accurate event summaries.

[0165] The goal of model training is to optimize the parameters by minimizing the loss function, thereby improving the accuracy of the model in predicting tweets. Specifically, the training process includes the following aspects:

[0166] Optimizer: The Adam optimizer is used to update the model parameters to minimize the loss function. The Adam optimizer combines the advantages of momentum and adaptive learning rate, which can effectively speed up the training process and improve model performance. The parameter update formula is formula (9):

[0167]

[0168] Among them, θ represents the model parameters, η is the learning rate, L is the gradient of the loss function with respect to the parameters θ.

[0169] Learning rate setting: The learning rate is a key hyperparameter that determines the parameter update step size of the model during training. In order to ensure the stability and efficiency of training, an adaptive learning rate adjustment strategy is adopted. Specifically, the learning rate is dynamically adjusted according to the changes in the loss function during training. If the loss does not drop significantly, the learning rate will be appropriately reduced to avoid falling into the local minimum of the loss function. This adaptive adjustment helps the model converge faster and achieve better performance.

[0170] Regularization: In order to prevent the model from overfitting, the training process needs to introduce regularization technology. This method uses dropout. In each iteration, the features of some nodes are randomly discarded with a certain probability, thereby enhancing the generalization ability of the model.

[0171] Iterative update: The model is trained through multiple rounds of iterations, and each round of iteration updates the model parameters according to the gradient of the loss function. The iteration process continues until the loss function converges or the predetermined number of iterations is reached.

[0172] Through the above training process, the model can learn effective parameters and achieve better performance in the social media text summary extraction task.

[0173] Finally, tweet selection is performed. Some events may require more than one tweet to be selected as a summary. In this case, each tweet represents a different aspect or development stage of the event. In this method, whether multiple tweets should be selected is determined in the following way: First, all tweets are sorted in descending order according to their scores. The tweet with the highest score is selected as part of the summary first, and then starting from the second tweet, each tweet is checked whether it meets the following two conditions: Condition 1: Its cosine similarity with any tweet selected as a summary is less than the threshold C1. This condition ensures that the next tweet selected as a summary discusses different aspects of the event than the selected tweets. Condition 2: The percentage of tweets similar to this tweet (measured by cosine similarity, with a threshold of C2) should be greater than the threshold P. This ensures that this tweet is representative enough.

[0174] With this approach, the number of tweets selected as summaries is determined dynamically, depending on the similarity between tweets and their respective representativeness, aiming to ensure that the generated summary can comprehensively cover the key information of the event while avoiding duplication of content.

[0175] The present invention innovatively combines social signals with text information. By constructing a tweet relationship graph, the semantic similarity and social interaction relationship are comprehensively considered. It can not only capture the semantic relationship of tweets, but also reflect the social interaction relationship between tweets, thereby improving the accuracy of summary extraction. The present invention uses a pre-trained BERT model to encode tweets, obtain their vector representations, and then construct a tweet relationship graph. Then, the tweet relationship graph is learned through a graph convolutional network, which can efficiently aggregate the information of neighbor nodes and update the feature representation of each node. In this way, the present invention can better capture the local neighborhood information of nodes in the graph structure, so that the representation of each tweet not only contains its own text information, but also integrates the social interaction information related to it. The significance score of each tweet is calculated based on the feature representation of the tweet, and the most representative text summary is extracted based on the score, so that the importance of each tweet can be accurately evaluated, thereby generating a more accurate and representative summary, and helping users quickly obtain key information. Compared with other existing methods, the present invention has achieved better results. By combining text information and structured social information, it can more effectively capture key information in social media texts and improve the quality and efficiency of summaries.

[0176] The embodiment of the present application also provides a text summary extraction device. It should be noted that the text summary extraction device of the embodiment of the present application can be used to execute the text summary extraction method provided in the embodiment of the present application. The text summary extraction device provided in the embodiment of the present application is introduced below.

[0177] Figure 4 4 is a schematic diagram of a text summary extraction device 400 provided according to an embodiment of the present application. Figure 4 As shown, the device includes: an acquisition module 401, an encoding module 402, a construction module 403, a determination module 404, and a generation module 405.

[0178] Specifically, the acquisition module 401 is used to acquire a tweet data set, wherein the tweet data set includes tweet data related to a preset event;

[0179] An encoding module 402, configured to encode the tweet data to obtain a plurality of tweet vectors;

[0180] A construction module 403 is used to construct a target tweet relationship graph based on multiple tweet vectors, wherein the target tweet relationship graph is used to reflect the data relationship of the tweet data and the tweet information volume of the tweet data;

[0181] A determination module 404, configured to determine a significance score of the target tweet data based on the target tweet relationship graph and the target model, wherein the significance score is used to represent a probability of the tweet data being added to the target summary;

[0182] The generating module 405 is used to generate a target summary based on the significance score and the tweet data, wherein the target summary is used to summarize and describe the preset event.

[0183] In summary, the text summary extraction device provided in the embodiment of the present application obtains a tweet data set through an acquisition module 401, wherein the tweet data set includes tweet data related to a preset event; encodes the tweet data through an encoding module 402 to obtain multiple tweet vectors; constructs a target tweet relationship graph based on multiple tweet vectors through a construction module 403, wherein the target tweet relationship graph is used to reflect the data relationship of the tweet data and the tweet information volume of the tweet data; determines the significance score of the target tweet data based on the target tweet relationship graph and the target model through a determination module 404, wherein the significance score is used to represent the probability of the tweet data being added to the target summary; generates a target summary based on the significance score and the tweet data through a generation module 405, wherein the target summary is used to summarize and describe the preset event, thereby solving the technical problem in the related art that a large amount of noisy and non-standard short texts on the social media platform cannot be processed due to the fragmentation of information, the informality of language, and the excessive amount of text, thereby resulting in the generated summary being of low quality and difficult to meet the needs of users. By capturing and integrating semantic information and social interaction relationships in social media texts, using graph convolutional networks and pre-trained models, we can reduce information redundancy, efficiently aggregate information of neighbor nodes, and update the feature representation of each node, thereby improving the efficiency and quality of summary extraction.

[0184] Furthermore, the encoding module 402 is also used to process the tweet data to obtain first tweet data; perform word segmentation processing on the first tweet data to obtain second tweet data; encode the second tweet data based on a preset model to determine a preset tag vector of an output layer of the preset model; and determine multiple tweet vectors based on the preset tag vector.

[0185] Furthermore, construction module 403 is also used to create a first tweet relationship graph, wherein the first tweet relationship graph includes multiple nodes, the multiple nodes are used to represent tweet data, and the multiple nodes include a first node and a second node; the first tweet relationship graph is updated based on the text similarity between the tweet vectors of the first nodes to obtain a second tweet relationship graph, wherein the first nodes in the second tweet relationship graph are connected by line segments; the second tweet relationship graph is updated based on the cosine similarity between the tweet vectors of the second nodes to obtain a third tweet relationship graph, wherein the second nodes in the third tweet relationship graph are connected by line segments; the third tweet relationship graph is adjusted based on the semantic term scores of the multiple nodes to obtain a fourth tweet relationship graph; the fourth tweet relationship graph is adjusted based on the target social signal to obtain a target tweet relationship graph.

[0186] Furthermore, the determination module 404 is also used to determine the local neighborhood information of the target tweet relationship graph based on multiple graph convolution layers of the target model; update the node features of the target tweet relationship graph based on the local neighborhood information; determine the initial significance score of the target tweet data based on a preset scoring strategy and node features; determine the significance score based on the target loss function and the initial significance score.

[0187] Furthermore, the generation module 405 is also used to arrange the significance scores in descending order to obtain an arrangement result; determine the first tweet data with the highest score in the arrangement result; filter the remaining tweet data except the first tweet data based on preset filtering rules to obtain second tweet data; and generate a target summary based on the first tweet data and the second tweet data.

[0188] Furthermore, the generation module 405 is also used to determine the cosine similarity between the tweet data in the remaining tweet data and the first tweet data; determine the similarity percentage of the tweet data in the remaining tweet data in the remaining tweet data; and add the tweet data whose cosine similarity is less than the first threshold and whose similarity percentage is greater than the second threshold to the second tweet data.

[0189] The text summary extraction device includes a processor and a memory. The above-mentioned acquisition module 401, encoding module 402, construction module 403, determination module 404, generation module 405, etc. are all stored in the memory as program units, and the processor executes the above-mentioned program units stored in the memory to realize corresponding functions.

[0190] The processor includes a kernel, which retrieves the corresponding program unit from the memory. One or more kernels can be set, and the text summary can be generated by adjusting the kernel parameters to meet user needs.

[0191] The memory may include non-permanent memory in a computer-readable medium, random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0192] An embodiment of the present invention provides a computer-readable storage medium on which a program is stored. When the program is executed by a processor, the text summary extraction method is implemented.

[0193] An embodiment of the present invention provides a processor, which is used to run a program, wherein the text summary extraction method is executed when the program is running.

[0194] like Figure 5As shown, an embodiment of the present invention provides an electronic device, the device includes a processor, a memory and a program stored in the memory and executable on the processor, and the processor implements the following steps when executing the program: obtaining a tweet data set, wherein the tweet data set includes tweet data related to a preset event; encoding the tweet data to obtain a plurality of tweet vectors; constructing a target tweet relationship graph based on the plurality of tweet vectors, wherein the target tweet relationship graph is used to reflect the data relationship of the tweet data and the amount of tweet information of the tweet data; determining a significance score of the target tweet data based on the target tweet relationship graph and the target model, wherein the significance score is used to represent the probability of the tweet data being added to the target summary; generating a target summary based on the significance score and the tweet data, wherein the target summary is used to summarize and describe the preset event.

[0195] When the processor executes the program, the following steps are also implemented: data processing is performed on the tweet data to obtain first tweet data; word segmentation is performed on the first tweet data to obtain second tweet data; encoding the second tweet data respectively based on a preset model to determine a preset tag vector of an output layer of the preset model; and multiple tweet vectors are determined based on the preset tag vector.

[0196] When the processor executes the program, the following steps are also implemented: creating a first tweet relationship graph, wherein the first tweet relationship graph includes multiple nodes, the multiple nodes are used to represent tweet data, and the multiple nodes include a first node and a second node; updating the first tweet relationship graph based on the text similarity between the tweet vectors of the first nodes to obtain a second tweet relationship graph, wherein the first nodes in the second tweet relationship graph are connected by line segments; updating the second tweet relationship graph based on the cosine similarity between the tweet vectors of the second nodes to obtain a third tweet relationship graph, wherein the second nodes in the third tweet relationship graph are connected by line segments; adjusting the third tweet relationship graph based on the semantic term scores of the multiple nodes to obtain a fourth tweet relationship graph; adjusting the fourth tweet relationship graph based on the target social signal to obtain a target tweet relationship graph.

[0197] When the processor executes the program, the following steps are also implemented: determining the local neighborhood information of the target tweet relationship graph based on multiple graph convolution layers of the target model; updating the node features of the target tweet relationship graph based on the local neighborhood information; determining the initial significance score of the target tweet data based on a preset scoring strategy and node features; and determining the significance score based on the target loss function and the initial significance score.

[0198] When the processor executes the program, the following steps are also implemented: arranging the significance scores in descending order to obtain an arrangement result; determining the first tweet data with the highest score in the arrangement result; filtering the remaining tweet data except the first tweet data based on preset filtering rules to obtain the second tweet data; and generating a target summary based on the first tweet data and the second tweet data.

[0199] When the processor executes the program, the following steps are also implemented: determining the cosine similarity between the tweet data in the remaining tweet data and the first tweet data; determining the similarity percentage of the tweet data in the remaining tweet data in the remaining tweet data; and adding the tweet data whose cosine similarity is less than a first threshold and whose similarity percentage is greater than a second threshold to the second tweet data.

[0200] The devices in this article can be servers, PCs, PADs, mobile phones, etc.

[0201] The present application also provides a computer program product, which, when executed on a data processing device, is suitable for executing a program that is initialized with the following method steps: obtaining a tweet data set, wherein the tweet data set includes tweet data related to a preset event; encoding the tweet data to obtain multiple tweet vectors; constructing a target tweet relationship graph based on the multiple tweet vectors, wherein the target tweet relationship graph is used to reflect the data relationship of the tweet data and the amount of tweet information of the tweet data; determining a significance score of the target tweet data based on the target tweet relationship graph and the target model, wherein the significance score is used to represent the probability of the tweet data being added to the target summary; generating a target summary based on the significance score and the tweet data, wherein the target summary is used to summarize and describe the preset event.

[0202] When executed on a data processing device, it is also suitable for executing an initialized program having the following method steps: performing data processing on tweet data to obtain first tweet data; performing word segmentation processing on the first tweet data to obtain second tweet data; encoding the second tweet data respectively based on a preset model to determine a preset tag vector of an output layer of the preset model; and determining multiple tweet vectors based on the preset tag vector.

[0203] When executed on a data processing device, it is also suitable for executing an initialized program having the following method steps: creating a first tweet relationship graph, wherein the first tweet relationship graph includes multiple nodes, the multiple nodes are used to represent tweet data, and the multiple nodes include a first node and a second node; updating the first tweet relationship graph based on the text similarity between the tweet vectors of the first nodes to obtain a second tweet relationship graph, wherein the first nodes in the second tweet relationship graph are connected by line segments; updating the second tweet relationship graph based on the cosine similarity between the tweet vectors of the second nodes to obtain a third tweet relationship graph, wherein the second nodes in the third tweet relationship graph are connected by line segments; adjusting the third tweet relationship graph based on the semantic term scores of the multiple nodes to obtain a fourth tweet relationship graph; adjusting the fourth tweet relationship graph based on the target social signal to obtain a target tweet relationship graph.

[0204] When executed on a data processing device, it is also suitable for executing a program that is initialized with the following method steps: determining local neighborhood information of a target tweet relationship graph based on multiple graph convolutional layers of a target model; updating node features of the target tweet relationship graph based on the local neighborhood information; determining an initial significance score of the target tweet data based on a preset scoring strategy and node features; and determining a significance score based on a target loss function and the initial significance score.

[0205] When executed on a data processing device, it is also suitable for executing an initialized program having the following method steps: arranging the significance scores in descending order to obtain an arrangement result; determining the first tweet data with the highest score in the arrangement result; filtering the remaining tweet data except the first tweet data based on preset filtering rules to obtain second tweet data; and generating a target summary based on the first tweet data and the second tweet data.

[0206] When executed on a data processing device, it is also suitable for executing a program initialized with the following method steps: determining the cosine similarity between tweet data in the remaining tweet data and the first tweet data; determining the similarity percentage of the tweet data in the remaining tweet data in the remaining tweet data; and adding tweet data whose cosine similarity is less than a first threshold and whose similarity percentage is greater than a second threshold to the second tweet data.

[0207] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0208] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0209] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0210] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0211] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0212] The memory may include non-permanent memory in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.

[0213] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0214] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0215] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0216] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included within the scope of the claims of the present application.

Claims

1. A text summary extraction method, characterized in that: The method comprises: Acquire a tweet data set, wherein the tweet data set includes tweet data related to a preset event; Encoding the tweet data to obtain multiple tweet vectors; constructing a target tweet relationship graph based on the multiple tweet vectors, wherein the target tweet relationship graph is used to reflect the data relationship of the tweet data and the tweet information volume of the tweet data; Determining a significance score of the target tweet data based on the target tweet relationship graph and the target model, wherein the significance score is used to represent the probability of the tweet data being added to the target summary; The target summary is generated based on the significance score and the tweet data, wherein the target summary is used to summarize and describe the preset event.

2. The method according to claim 1, characterized in that The encoding of the tweet data to obtain a plurality of tweet vectors comprises: Performing data processing on the tweet data to obtain first tweet data; Performing word segmentation processing on the first tweet data to obtain second tweet data; Encode the second tweet data respectively based on a preset model, and determine a preset tag vector of an output layer of the preset model; The plurality of tweet vectors are determined based on the preset tag vector.

3. The method according to claim 1, characterized in that The constructing a target tweet relationship graph based on the multiple tweet vectors includes: Creating a first tweet relationship graph, wherein the first tweet relationship graph includes a plurality of nodes, the plurality of nodes are used to represent the tweet data, and the plurality of nodes include a first node and a second node; The first tweet relationship graph is updated based on the text similarity between the tweet vectors of the first nodes to obtain a second tweet relationship graph, wherein the first nodes in the second tweet relationship graph are connected by line segments; updating the second tweet relationship graph based on the cosine similarity between the tweet vectors of the second nodes to obtain a third tweet relationship graph, wherein the second nodes in the third tweet relationship graph are connected by line segments; adjusting the third tweet relationship graph based on the semantic term scores of the multiple nodes to obtain a fourth tweet relationship graph; The fourth tweet relationship graph is adjusted based on the target social signal to obtain the target tweet relationship graph.

4. The method according to claim 3, characterized in that Determining the significance score of the target tweet data based on the target tweet relationship graph and the target model includes: Determining local neighborhood information of the target tweet relationship graph based on multiple graph convolutional layers of the target model; Updating node features of the target tweet relationship graph based on the local neighborhood information; Determining an initial significance score of the target tweet data based on a preset scoring strategy and the node features; The saliency score is determined based on a target loss function and the initial saliency score.

5. The method according to any one of claims 1 to 4, characterized in that Generating the target summary based on the significance score and the tweet data includes: Arranging the significance scores in descending order to obtain an arrangement result; Determine the first tweet data with the highest score in the arrangement result; Filtering the remaining tweet data except the first tweet data based on a preset filtering rule to obtain second tweet data; The target summary is generated based on the first tweet data and the second tweet data.

6. The method according to claim 5, characterized in that The filtering of the remaining tweet data except the first tweet data based on the preset filtering rule to obtain the second tweet data includes: Determine the cosine similarity between the tweet data in the remaining tweet data and the first tweet data; determining a similarity percentage of tweet data in the remaining tweet data among the remaining tweet data; The tweet data whose cosine similarity is less than the first threshold and whose similarity percentage is greater than the second threshold are added to the second tweet data.

7. A text summary extraction device, characterized in that: The device comprises: An acquisition module, configured to acquire a tweet data set, wherein the tweet data set includes tweet data related to a preset event; An encoding module, used for encoding the tweet data to obtain a plurality of tweet vectors; A construction module, configured to construct a target tweet relationship graph based on the multiple tweet vectors, wherein the target tweet relationship graph is used to reflect the data relationship of the tweet data and the tweet information volume of the tweet data; A determination module, configured to determine a significance score of target tweet data based on the target tweet relationship graph and the target model, wherein the significance score is used to represent a probability of the tweet data being added to the target summary; A generating module is used to generate the target summary based on the significance score and the tweet data, wherein the target summary is used to summarize and describe the preset event.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, wherein the computer program is configured to execute the text summary extraction method described in any one of claims 1 to 6 when running on a computer or a processor.

9. A computer program product, characterized in that The invention comprises a computer program, which implements the text summary extraction method according to any one of claims 1 to 6 when being executed by a processor.

10. An electronic device comprising a memory and a processor, characterized in that: The memory stores a computer program, and the processor is configured to run the computer program to execute the text summary extraction method described in any one of claims 1 to 6.

Images