Social media text summarization method considering semantic content, social relationship and multi-aspect attributes of text

Abstract

The invention provides a social media text summarization method considering semantic content, social relationships and multi-aspect attributes of a text. The method comprises the following steps: (1) constructing a multi-relationship graph of a social relationship graph and a semantic relationship graph; (2) a multi-stream encoder is introduced to mine social and semantic relations among posts, and the problem of post information missing is solved by integrating the information; (3) after the social information and the content relevance are fused, a multi-stream reconstruction module is adopted to reconstruct expressions of the posts, including expressions of an original text, a social and semantic relation graph and an abstract; and (4) providing a centrality ranking model based on aspect perception, introducing sentence-document weights to make the model pay more attention to different aspects, and adding the weights into centrality scores of posts to ensure that the abstract covers a wider aspect.

Description

Technical Field

[0001] The present invention relates to the field of natural language processing and social media data analysis, and in particular to a method for extracting social media summaries that comprehensively considers semantic content and social relationships. This method enhances the summary quality of social media content in an innovative multi-dimensional manner by combining the sociological relationships of user interactions in social networks and the semantic relevance between posts. The core of the invention is a model called MultiSum, which uses a graph convolutional network (GCN) and a novel graph-based ranking method to not only capture the deep semantic information of text content, but also comprehensively consider the social relationships between posts, thereby extracting summaries that contain multiple perspectives and information aspects. In addition, the effectiveness of the present invention in different language environments has been experimentally verified, and it can improve the informativeness of summaries and reduce bias. Background Art

[0002] With the rapid development of information technology, social media has become an integral part of daily life. It has not only transformed how people communicate but has also significantly impacted the dissemination and acquisition of information. Users publish, share, and comment on a wide variety of content through social media platforms, forming a vast and vibrant virtual society. The real-time and interactive nature of information is a hallmark of social media. Users can post messages instantly and receive feedback from other users, thus forming a multi-dimensional information exchange network.

[0003] While social media provides a convenient platform for information exchange, it also creates the problem of information overload. Every day, hundreds of millions of users generate massive amounts of text data on social platforms. These texts are short, timely, diverse in style, and imbued with rich social attributes. Users often feel overwhelmed by this vast amount of information and struggle to quickly filter out valuable content. This information overload not only increases users' cognitive burden but also reduces the efficiency of information retrieval and decision-making.

[0004] The development of social media summarization technology aims to address the problem of information overload by automatically extracting key information from social media and producing concise summaries. This technology helps users quickly grasp the main points of large amounts of text, saving time and improving the efficiency of information consumption. However, due to the unique characteristics of social media text, such as informal language, colloquial expressions, the use of abbreviations and emoticons, as well as the shortness and contextual dependence of text, traditional text summarization techniques are difficult to directly apply to the social media domain.

[0005] Furthermore, text in social media isn't just independent information; it also encompasses complex social relationships. For example, user interactions, discussion participation, and information dissemination paths all constitute important social dimensions beyond the textual content. These social dimensions are crucial for evaluating the importance of a post, but traditional summarization techniques often overlook this aspect.

[0006] Current social media summarization techniques primarily focus on the textual content of posts, employing various content-based ranking algorithms to select key information. These methods, such as manually defined scoring functions, graph-based ranking algorithms, and clustering-based approaches, can extract key content from text to a certain extent. However, they treat each post as an independent unit of information, ignoring the interrelationships between posts and the network structure of social media.

[0007] Faced with these challenges in social media text, researchers have begun exploring new approaches. Some studies have attempted to incorporate social signals (such as the number of retweets, likes, and comments) into the summary process, hoping to better reflect the social influence of a post and user engagement. While these social signals can indeed help determine a post's popularity and importance to a certain extent, they do not always accurately reflect the quality of the post's content. A post may receive high attention due to a compelling title or the publisher's high social status, but this does not necessarily mean that the content is high-quality or highly relevant to a specific query. To more comprehensively consider social media content and produce high-quality summaries, some studies have begun to view social media posts as a holistic network, exploring how to leverage the relationships between posts to improve the accuracy and information coverage of summaries. For example, by analyzing semantic connections between posts and social relationships between users, it is possible to better identify important posts and reflect the diversity and comprehensiveness of information on social media in summaries. This approach recognizes that social media summarization is not just a technical issue of extracting key information, but also a sociological issue of considering and interpreting the structure of social networks.

[0008] However, these studies fail to consider the diverse relationships between social media texts, such as social network relationships and semantic relationships within text. Leveraging these relationships can offset the information deficit of individual social media texts. Furthermore, when extracting summaries, mainstream graph-based methods tend to select similar content. While clustering can increase the diversity of summaries, it requires manual feature extraction from posts and a defined number of facets, which requires expert experience and limited flexibility.

[0009] To address these issues, we propose MultiSum, a social media text summarization method that considers semantic content, social relationships, and multiple aspects of textual attributes. This model not only comprehensively considers the textual content and semantic information of a post, but also integrates the social relationships between posts and user behavior patterns. Using a graph convolutional network (GCN) and a graph-based ranking method, MultiSum effectively processes and analyzes complex social network data, resulting in more accurate, comprehensive, and less biased social media summaries.

[0010] In summary, this invention represents a significant innovation in the field of social media summarization technology. It not only addresses the limitations of traditional summarization techniques in social media applications but also provides new perspectives and tools for considering social network structure and user behavior. This invention enables users to quickly access valuable summaries in an information-overloaded environment, thereby improving the efficiency and quality of information consumption. Summary of the Invention

[0011] This paper proposes a social media text summarization method, named MultiSum, that considers semantic content, social relationships, and multifaceted text attributes. This method improves social media summarization by considering multiple relationships between posts and diverse perspectives in summaries. First, the paper utilizes a graph convolutional network (GCN) to collect information about a post's neighbors based on both social and semantic relationships, which helps alleviate the problem of insufficient information about a single post. Second, during the selection phase, an improved ranking algorithm is introduced to reduce aspect bias caused by traditional centrality calculation methods and improve the diversity of perspectives in summaries.

[0012] The present invention comprises the following key components:

[0013] 1. Multi-relationship graph construction: Based on a sociological perspective, this paper creates a social relationship graph and a semantic relationship graph to analyze the social and semantic connections of posts. This process involves first obtaining a representation of the posts and then using a graph convolutional network (GCN) to integrate this information.

[0014] 2. Multi-stream Encoder: We introduce a novel multi-stream encoder designed to mine the social and semantic connections between posts. This encoder consists of two GCN encoders: one for social relationships and one for semantic relationships. These encoders address the issue of missing post information by integrating social and textual information.

[0015] 3. Multi-stream reconstruction: After integrating social information and content relevance, a multi-stream reconstruction module is used to reconstruct the representation of the post, including the representation of the original text, social and semantic relationship graphs, and summaries.

[0016] 4. Multifaceted Summarizer: The core of this invention is to apply a graph-based ranking algorithm to evaluate the centrality score of each post, introduce a directed graph variant method to achieve wide coverage of summaries, and use a model to enhance the similarity between candidate summaries and all posts to evaluate their relevance.

[0017] The effectiveness of the present invention has been verified through tests on two real social media datasets, and evaluations by human evaluation and three large language models have shown that the extracted summaries can cover more perspectives and have lower redundancy.

[0018] Furthermore, step (1) specifically includes:

[0019] (101) Building a social relationship network at the post level

[0020] The post-level social relationship network consists of a node set and an edge set. Each node in the node set represents a post, and each edge in the edge set represents the social relationship between corresponding posts. There are two types of social relationships between posts: expression consistency relationship and expression contagion relationship. Expression consistency relationship refers to the relationship between posts published by the same user. When constructing the post-level social relationship network, an edge is established between post nodes with expression consistency relationship. Expression contagion relationship refers to the relationship between posts published by users with direct interaction relationship, among which direct interaction relationship refers to the interaction relationship of following, forwarding, and commenting between users. When constructing the post-level social relationship network, an edge is established between post nodes with expression contagion relationship.

[0021] The post-level social relationship network is formally described as follows: Represents a post set, N is the number of posts, where s i (1≤i≤N) represents the i-th post; Represents a user set, which contains M users in total, where u i (1≤i≤M) represents the i-th user; for user u i ,make Represents user u i Neighbor users, that is, users u i A collection of users with direct social relationships; Indicates all the i A collection of published posts; a post-level social relationship network is constructed according to the following rules in Represents a set of nodes, each node corresponds to a post, ε represents the set of edges between nodes, and each edge corresponds to the set of social relations between posts; in expressing consistent social relations: if a post where u krepresents the kth user, then the post s i With posts j Create an edge e between ij ∈ε; in expressing infectious social relations: if the post And there are or Then for posts i With s j Create an edge e between ij ∈ε; Based on these two rules, a post-level social relationship network is constructed It only contains the post node collection And the relationship between nodes, that is, the edge set ε={e 11 ,e 12 ,…,e NN}; Constructed post-level social relationship network The corresponding adjacency matrix is ​​recorded as Among them A ij >0 indicates post node s i With s j There is a social relationship between them, otherwise A ij =0;

[0022] (102) Constructing a post-level semantic relationship network

[0023] In addition, the semantic relationship graph is used to explore the semantic relationship between posts. Specifically, for each post s in the set S i , input the pre-trained BERT model, and extract the [CLS] output of the last layer as the post representation. In this invention, the BERTWEET model is used, which is more suitable for processing social media posts. Posts i The representation of is given by the following formula:

[0024] x i =BERTWEET(s i ) (1)

[0025] Among them, x i Indicates posts i The initial content representation of .

[0026] Next, we use the cosine similarity between posts as the weight of the edge between them. Based on this method, we construct E sem , and use the adjacency matrix A sem ∈R N×N To represent it. In this way, the semantic relationship graph It is constructed based on the semantic similarity between posts.

[0027] Combining these two components, we first construct a post-level social relationship network based on expression consistency and expression contagion, and then build a semantic relationship graph based on semantic similarity between posts. Together, these two graphs provide the model with rich social context and semantic information, providing a more comprehensive and in-depth basis for subsequent summary extraction.

[0028] Step (2) of the multi-stream encoder includes the following:

[0029] (201) Aggregating multi-relational graph information using graph convolutional networks

[0030] The graph convolutional network GCN is used to aggregate the features of neighbor nodes to learn the representation of nodes in graph structure data. In order to use the graph convolutional network GCN to aggregate the features of neighbor nodes and learn the representation of nodes in graph structure data, we take the social relationship graph as an example, where Represents the node content features extracted by the content encoder, D is the dimension of each node content feature, Represents the adjacency matrix corresponding to the constructed post-level social relationship network. The graph convolutional network combines the node feature X with the adjacency matrix A of the post-level social relationship network. soc As input, information propagation is performed according to the following rules:

[0031]

[0032]

[0033] Among them H (0) =X, is the matrix obtained by adding the adjacency matrix to the unit matrix and then normalizing it; I is the unit matrix, It's A soc +I degree matrix; W (l) with b (l) is the learning parameter of the lth layer; in each layer of the graph convolutional network, each node is normalized by the adjacency matrix To aggregate the features of its adjacent nodes, use W (l) with b (l) The features of neighbor nodes are linearly mapped and then passed through the activation function to obtain the aggregated neighbor node features. Each layer of the graph convolutional network is replaced with a residual connection. This residual connection allows the input data to directly transmit information to the output. Therefore, the specific propagation rules become as follows:

[0034]

[0035] here is the node representation after encoding, Z is the dimension of the node representation; σ(·) represents the activation function, Is a mapping function responsible for mapping node features to semantic space Here, a fully connected layer with ReLU activation function is used as the mapping function.

[0036] Similarly, we can use the same method to obtain the representation of the fused semantic relationship graph By concatenating the outputs of the two encoders, the final representation includes both social relationship information and semantic information.

[0037]

[0038] Step (3) The multi-stream reconstruction module reconstructs the social relationship graph, semantic relationship graph, original text and summary of the post respectively. Among them, the social relationship decoder is responsible for capturing the social relationship information in the social relationship network at the post level, as follows:

[0039] (301) Social Relationship Decoder

[0040] The decoder takes the node representation obtained after graph convolutional network encoding as input and is used to predict the social relationship between nodes. Formally expressed, given the node representation H = [h1,…,h N ] T , the social relationship decoder predicts the social relationship between nodes according to the following formula:

[0041]

[0042] Among them, Dec soc (·) is the social relationship decoder function, is the post-level social relationship network predicted by the decoder, Indicates post p i With post p j The probability of having a social relationship connection, h i With h j Corresponding to post p i With p j The vector representation of ; the vector inner product is used as the social relationship decoder function, which predicts the probability of a social relationship connection between two nodes based on the inner product of the vector representation between the nodes:

[0043] Dec soc (h i ,h j )=P(A i,j |h i ,h j )=σ(h i ·h j ) (7)

[0044] Where σ(·) represents the activation function; during the training process, the prediction result is minimized The binary cross entropy between the social relationship network structure A at the standard post level is:

[0045]

[0046] Where N has the same meaning as above, representing the number of posts, I[·] is the indicator function, when A i,j >0, I[A i,j ]=1, otherwise I[A i,j ]=0;

[0047] The reconstruction of the semantic relationship graph is similar to the reconstruction of the social relationship graph. The difference is that during the training process, the mean absolute error is used as the loss function:

[0048]

[0049] (302) Raw Text Decoder

[0050] In order to consider the post content more deeply and capture the relationship between different semantic units in the text, we built a raw text decoder. The purpose of this decoder is to consider the content of the post text from different semantic levels. We constructed a semantic relationship bipartite graph in Represents N post nodes, Corresponding to V words in the vocabulary; ε b ={e 11 ,…,e NV} represents the semantic relationship between posts and words. If e ij =1, it means post p i Contains the word w j , otherwise e ij =0;

[0051] Given the above semantic relationship bipartite graph, the semantic relationship bipartite graph of the post is reconstructed according to the vector representation of the post node; formally, the vector representation of the post node obtained by the graph convolutional network is H = [h1,h2,…,h N ] T As input to the semantic relation decoder, the decoder predicts the semantic relation between the node and the word based on the node representation:

[0052]

[0053] Here, Dec se (·) is the semantic relation decoder function; is the prediction result output by the decoder, where V is the vocabulary size, eik (1≤k≤V)∈[0,1] represents the post p i Contains the word w k The probability of ; a fully connected layer with a sigmoid activation function is used as the semantic relationship decoder function:

[0054] Dec se (h i )=σ(h i W d +b d ) (11)

[0055] Where W d with b d is a learning parameter; during the training process, the real semantic relationship bipartite graph is used as supervision and the predicted semantic relationship bipartite graph is minimized and the true semantic relationship bipartite graph ε b The binary cross entropy loss between:

[0056]

[0057] (303) Reconstruction of Abstract Representation

[0058] The last important part in our model is the reconstruction loss of the summary. To measure the quality and accuracy of the summary, we use the average cosine distance (denoted as d cos ) as a metric. This metric is used to calculate each encoded post (h j ) and the encoded summary (h s ). In the model, we optimize the summary reconstruction loss to minimize the mean cosine distance, so that the summary is semantically closer to the collection of original posts. The specific steps are as follows:

[0059] For each encoded post h j and the encoded summary h s , calculate the cosine distance between them

[0060] Calculate the average cosine distance between all posts and summaries to get the average cosine distance d cos (h s ,h j )

[0061] Optimize the model to minimize the mean cosine distance, so that the summary is semantically closer to the collection of original posts

[0062]

[0063] In this way, our model is able to not only extract summaries, but also ensure that the content of the summaries is closely related to the original posts, thereby improving the accuracy and usability of the summaries. This reconstruction process is a key part of the entire summary extraction process, which directly affects the quality and effectiveness of the summaries.

[0064] To achieve this goal, we ultimately learn to simultaneously reconstruct representations of the post-level social relation network, semantic network, semantic relation bipartite graph, and summary. The training loss function is as follows:

[0065] min L=L se +L soc +L sem +L sim (14)

[0066] During the training process, the initial representation X of the post is iteratively updated according to the loss function; after the training is completed, the encoder part is used to encode the post to obtain the post representation enhanced with social context, which is applied to the subsequent summary task.

[0067] Furthermore, step (4) the multi-aspect summarizer includes the following:

[0068] (401) In order to achieve more comprehensive and diverse social media content summaries, this paper proposes an aspect-aware centrality-based ranking model. By introducing sentence-document weights, the model pays more attention to different aspects. The weights are added to the centrality scores of posts, allowing the summary to cover a wider range of aspects. i When calculating the centrality score, we use the following formula:

[0069]

[0070] Among them, e ij There are two posts The dot product of ∈=β·(max(e ij )-min(e ij )), β is a hyperparameter. We rank the posts according to this formula and select the top k ranked posts as summaries.

[0071] In addition, we introduce a method to model different aspects of the candidate summary by computing The relevance is measured by the similarity between d and all posts d. is from the one with the highest DC(s i ) score. We combine summary-document similarity with sentence centrality to obtain the best candidate summary:

[0072] summary=topK({DC(si )} i＝1,..,n ) (16)

[0073]

[0074] Where topK is the K posts with the highest scores, candidate summaries Is to use DC(s i ) is the subset of the top K sentences after sorting, and α is a hyperparameter that controls the influence of directional centrality. refers to the sentence-document similarity, where d is the document representation, is a candidate summary representation. get This is the average representation of the summary sentences. We choose the cosine similarity of sim(·). For the document representation, we first collect all the sentence representations h1,h2,...,h n In order to compress all valuable information in the document, we apply the Maxpooling function to the sentence representation. The document representation d is calculated as:

[0075] d=Maxpooling{h1,h2,…,h n} (18)

[0076] Using the aforementioned multi-aspect summarizer, our model demonstrates its effectiveness in experiments on two real-world social media datasets. The experimental results show that our summaries cover more perspectives and are less redundant, as evaluated by humans and three large language models.

[0077] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0078] (1) Dual consideration of structure and content: This paper integrates two graphs for the social media text summarization task: the social network graph and the social text semantic graph. This method considers both the structure and content of posts, effectively capturing the complex relationships and deep semantics in social media texts, thereby improving the accuracy and comprehensiveness of summary extraction.

[0079] (2) Multifaceted Summary Extraction: In light of the inherent multifaceted nature of social media content, this paper uses a multifaceted summarizer to extract summaries covering multiple aspects. This approach can present social media content from different perspectives, ensuring the diversity and comprehensiveness of summaries and reducing information omissions and bias.

[0080] (3) Experimental Validation on Real-World Datasets: We conducted extensive experiments on real-world social media datasets in English and Chinese. The experimental results demonstrate the effectiveness of our model. Our model is not only applicable to different language environments but also can effectively process and summarize large amounts of social media information, demonstrating its broad applicability and superior performance in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Figure 1 This paper describes the overall architecture of the proposed "MultiSum" model, which is specifically designed for social media summarization tasks and considers multiple relationships between posts and different aspects in summaries. Figure 1 The following key components are presented: multi-stream encoder, multi-stream reconstruction module, and multi-aspect summarizer. DETAILED DESCRIPTION

[0082] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0083] This paper proposes an unsupervised extractive social media summarization method, named MultiSum, which improves the extraction of social media summaries by considering multiple relationships between posts and various perspectives in summaries. The method includes the following:

[0084] (1) Construction of post-level social relationship network

[0085] This example uses data from mainstream social media platforms, both domestic and international, for experimental validation. Twitter and Sina Weibo were selected for experimental validation. The data was compiled to ensure that posts under each topic were published within five days. For Twitter data, the primary language of text content is English. The data contains 44,034 posts and 11,240 users, each of whom publishes at least one post and has at least one social connection. Four standard reference summaries were used for evaluating the results under each topic. Links, usernames, and other special characters were removed from the posts. Stemming and stop word removal were performed, and posts with less than three words were filtered out. For Weibo data, Sina Weibo is one of the most popular social media platforms in China. Therefore, this example uses data collected from Sina Weibo, which contains 130k posts and 126k users, encompassing a total of 10 different topics. Posts are organized in a tree structure based on interaction relationships (e.g., replies and forwards). Three standard reference summaries were used for evaluating the results under each topic. The statistical details of the two datasets after preprocessing are shown in Table 1. In the experiment, the ROUGE evaluation standard is used, and the four evaluation results of ROUGE-1, ROUGE-2, ROUGE-L and ROUGE-SU* are mainly reported.

[0086] Formally, let Represents a post set, N is the number of posts, where s i (1≤i≤N) represents the i-th post; Represents a user set, which contains M users in total, where u i (1≤i≤M) represents the i-th user; for user u i ,make Represents user u i Neighbor users, that is, users u i A collection of users with direct social relationships; Indicates all the i A collection of published posts; a post-level social relationship network is constructed according to the following rules in Represents a set of nodes, each node corresponds to a post, ε represents the set of edges between nodes, and each edge corresponds to the set of social relations between posts; in expressing consistent social relations: if a post

[0087] where u k represents the kth user, then the post s i With posts j Create an edge e between ij∈ε; in expressing infectious social relations: if the post And there are or Then for posts i With s j Create an edge e between ij ∈ε; Based on these two rules, a post-level social relationship network is constructed It only contains the post node collection And the relationship between nodes, that is, the edge set ε={e 11 ,e 12 ,…,e NN}; Constructed post-level social relationship network The corresponding adjacency matrix is ​​recorded as Among them A ij >0 indicates post node s i With s j There is a social relationship between them, otherwise A ij =0;

[0088] The semantic relationship graph explores the semantic relationship between posts. First, we need to obtain the representation of the posts. Specifically, for each post s in the set S i , we input it into the pre-trained BERT model. Then, we take the output of the sentence start token ([CLS]) in the last layer of the model as the embedding of the post. In this paper, we use the BERTWEET model, which has a similar structure to BERT but is trained on a large amount of tweet data, making it more suitable for processing social media posts. The representation of each post is obtained by the following formula:

[0089] x i =BERTWEET(s i ) (19)

[0090] Among them, x i Indicates posts i The initial content representation of .

[0091] Next, we use the cosine similarity between posts as the weight of the edge between them. Based on this method, we construct E sem , and use the adjacency matrix A sem ∈R N×N To represent it. In this way, the semantic relationship graph It is constructed based on the semantic similarity between posts.

[0092] Table 1. Details of social media dataset

[0093]

[0094] (2) Aggregating multi-relational graph information using graph convolutional networks

[0095] The graph convolutional network GCN is used to aggregate the features of neighbor nodes to learn the representation of nodes in graph structure data; taking the social relationship graph as an example, represents the content features of the node extracted by the content encoder, D is the dimension of the content features of each node, It represents the adjacency matrix corresponding to the constructed post-level social relationship network; the graph convolutional network combines the node feature X with the adjacency matrix A of the post-level social relationship network soc As input, the information is propagated according to the following rules:

[0096]

[0097]

[0098] Among them H (0) =X, is the matrix obtained by adding the adjacency matrix to the unit matrix and then normalizing it; I is the unit matrix, It's A soc +I degree matrix; W (l) with b (l) is the learning parameter of the lth layer; in each layer of the graph convolutional network, each node is normalized by the adjacency matrix To aggregate the features of its adjacent nodes, use W (l) with b (l) The features of neighbor nodes are linearly mapped and then passed through the activation function to obtain the aggregated neighbor node features. Each layer of the graph convolutional network is replaced with a residual connection. This residual connection allows the input data to directly transmit information to the output. Therefore, the specific propagation rules become as follows:

[0099]

[0100] here is the node representation after encoding, Z is the dimension of the node representation; σ(·) represents the activation function, Is a mapping function responsible for mapping node features to semantic space Here, a fully connected layer with ReLU activation function is used as the mapping function.

[0101] Similarly, we can use the same method to obtain the representation of the fused semantic relationship graph By concatenating the outputs of the two encoders, the final representation includes both social relationship information and semantic information.

[0102]

[0103] (3) Decoding using a multi-stream reconstruction decoder

[0104] Multi-stream reconstruction reconstructs the social and semantic relationship graphs of posts, original text, and summary representations. The social relationship decoder is used to capture the social relationship information of posts in the post-level social relationship network; the social relationship decoder takes the node representation obtained by graph convolutional network encoding as input and predicts the social relationship between input nodes; formally speaking, given the node representation H = [h1,…,h N ] T , the social relationship decoder predicts the social relationship between nodes according to the following formula:

[0105]

[0106] Among them, Dec soc (·) is the social relationship decoder function, is the post-level social relationship network predicted by the decoder, Indicates post p i With post p j The probability of having a social relationship connection, h i With h j Corresponding to post p i With p j The vector representation of ; the vector inner product is used as the social relationship decoder function, which predicts the probability of a social relationship connection between two nodes based on the inner product of the vector representation between the nodes:

[0107] Dec soc (h i ,h j )=P(A i,j |h i ,h j )=σ(h i ·h j ) (25)

[0108] Where σ(·) represents the activation function; during the training process, the prediction result is minimized The binary cross entropy between the social relationship network structure A at the standard post level is:

[0109]

[0110] Where N has the same meaning as above, representing the number of posts, I[·] is the indicator function, when A i,j >0, I[A i,j ]=1, otherwise I[A i,j ]=0;

[0111] The reconstruction of the semantic relationship graph is similar to the reconstruction of the social relationship graph. The difference is that during the training process, the mean absolute error is used as the loss function:

[0112]

[0113] In order to take the post content into consideration more deeply and capture the relationship between different semantic units in the text, we built a raw text decoder. The purpose of this decoder is to understand the content of the post text from different semantic levels. We constructed a semantic relationship bipartite graph in Represents N post nodes, Corresponding to V words in the vocabulary; ε b ={e 11 ,…,e NV} represents the semantic relationship between posts and words. If e ij =1, it means post p i Contains the word w j , otherwise e ij =0;

[0114] Given the above semantic relationship bipartite graph, the semantic relationship bipartite graph of the post is reconstructed according to the vector representation of the post node; formally, the vector representation of the post node obtained by the graph convolutional network is H = [h1,h2,…,h N ] T As input to the semantic relation decoder, the decoder predicts the semantic relation between the node and the word based on the node representation:

[0115]

[0116] Here, Dec se (·) is the semantic relation decoder function; is the prediction result output by the decoder, where V is the vocabulary size, e ik (1≤k≤V)∈[0,1] represents the post p i Contains the word w k The probability of ; a fully connected layer with a sigmoid activation function is used as the semantic relationship decoder function:

[0117] Dec se (h i )=σ(h i W d +b d ) (29)

[0118] Where W d with b dis a learning parameter; during the training process, the real semantic relationship bipartite graph is used as supervision and the predicted semantic relationship bipartite graph is minimized and the true semantic relationship bipartite graph ε b The binary cross entropy loss between:

[0119]

[0120] The last important part in our model is the reconstruction loss of the summary. To measure the quality and accuracy of the summary, we use the average cosine distance (denoted as d cos ), this metric is used to measure the j ) and the encoded summary (h s ) between the hidden states.

[0121] Specifically, the goal of summary reconstruction is to minimize the difference between the encoded post and summary representations. This is achieved by calculating the cosine distance between them. The average cosine distance is an effective metric because it reflects the proximity between the post and summary in semantic space. A lower average cosine distance means that the encoded summary is closer to the encoded representation of the actual post, indicating a higher quality summary.

[0122] The reconstruction process of the summary representation can be formalized into the following steps:

[0123] For each encoded post h j and the encoded summary h s , calculate the cosine distance between them

[0124] Calculate the average cosine distance between all posts and summaries to get the average cosine distance d cos (h s ,h j )

[0125] Optimize the model to minimize the mean cosine distance, so that the summary is semantically closer to the collection of original posts

[0126]

[0127] In this way, our model is able to not only extract summaries, but also ensure that the content of the summaries is closely related to the original posts, thereby improving the accuracy and usability of the summaries. This reconstruction process is a key part of the entire summary extraction process, which directly affects the quality and effectiveness of the summaries.

[0128] The ultimate goal is to learn to simultaneously reconstruct the representation of the post-level social relationship network, semantic network, semantic relationship bipartite graph and summary, and the final training loss is as follows:

[0129] minL=L se +L soc +L sem +L sim (32)

[0130] During the training process, the initial representation X of the post is iteratively updated according to the loss function; after the training is completed, the encoder part is used to encode the post to obtain the post representation enhanced with social context, which is applied to the subsequent summary task.

[0131] (4) Multifaceted Summarizer

[0132] The Multi-Aspect Summarizer aims to achieve graph-based post ranking by calculating a centrality score for each post. It is an aspect-aware centrality ranking model that introduces sentence-document weights to increase the model's focus on different aspects. These weights are added to the post centrality scores, allowing summaries to cover a wider range of aspects, thereby improving their diversity and comprehensiveness.

[0133] Calculating a post i When calculating the centrality score of a post, we set a threshold ∈ to filter out posts with a large semantic gap with the post, thereby reducing the impact of noise. The connection between each post is calculated by the dot product of the two posts, and the centrality score is determined according to the following formula:

[0134]

[0135] Among them, e ij There are two posts The dot product of ∈=β·(max(e ij )-min(e ij )), β is a hyperparameter. We rank the posts according to this formula and select the top K ranked posts as summaries.

[0136] In addition, we introduce a method to model different aspects of the candidate summary by computing The relevance is measured by the similarity between d and all posts d. is from the one with the highest DC(s i ) score. We combine summary-document similarity with sentence centrality to obtain the best candidate summary:

[0137] summary=topK({DC(s i )} i＝1,..,n ) (34)

[0138]

[0139] Where topK is the K posts with the highest scores, candidate summaries Is to use DC(s i ) is the subset of the top K sentences after sorting, and α is a hyperparameter that controls the influence of directional centrality. refers to the sentence-document similarity, where d is the document representation, is a candidate summary representation. get This is the average representation of the summary sentences. We choose the cosine similarity of sim(·). For the document representation, we first collect all the sentence representations h1,h2,...,h n In order to compress all valuable information in the document, we apply the Maxpooling function to the sentence representation. The document representation d is calculated as:

[0140] d=Maxpooling{h1,h2,…,h n} (36)

[0141] In our experiments, we set the following hyperparameter combination: α = 1, β = 0.1, λ1 = 0.4, λ2 = 0.6. By leveraging the aforementioned multi-aspect summarizer, our model not only improves summary performance, but also extracts summaries that cover more aspects and reduce bias, thereby achieving more comprehensive and diverse summaries of social media content. Experiments on two real-world social media datasets validate the effectiveness of our approach. Evaluations with humans and three large language models demonstrate that our summaries cover more perspectives and are less redundant.

[0142] In order to verify the effectiveness of the method of the present invention, the method of the present invention is compared with the current advanced and representative models (LSA, LexRank, DSDR, MDS-Sparse, SNSR, PacSum, Spectral, SCMGR).

[0143] LSA [1] Use singular value decomposition to decompose the feature matrix of the posts, and select the posts with the highest singular value to add to the summary.

[0144] LexRank [2] An affinity graph is constructed between posts and a graph ranking algorithm similar to PageRank is used to extract nodes with higher importance in the graph as summaries.

[0145] DSDR [3] Using a linear reconstruction-based method, the model is trained by minimizing the reconstruction error and the summary is extracted based on the reconstruction coefficient.

[0146] MDS-Sparse[4] A method based on reconstructing the original corpus is also used to extract the summary, and sparse coding technology is also integrated to extract the summary.

[0147] SNSR [5] The social relations between posts are integrated with the sparse reconstruction framework to extract summaries, which improves the diversity of summaries.

[0148] PacSum [6] It is a graph-based extractive summarization method that uses BERT as a feature extractor for sentences, while considering the directionality of the relationship between sentences, sorting them according to their centrality, and selecting posts with higher centrality as summaries.

[0149] Spectral [7] It is an unsupervised multi-document summary extraction model based on the spectral method. It calculates the spectral importance of each post and selects posts with higher spectral importance as summaries.

[0150] SCMGR [8] A graph convolutional network is used to extract social post representations in the original social network, and a sparse reconstruction framework is used to extract summary posts.

[0151] MTGNN [9] We use finer-grained semantic units in the document graph to capture the complex relationships between words and sentences, and construct a heterogeneous graph neural network for summarization.

[0152] Experimental performance was evaluated using the ROUGE evaluation criteria, which include ROUGE-1, ROUGE-2, ROUGE-L, and ROUGE-SU*. ROUGE-N measures the overlap of N-grams between the output summary and the standard summary; ROUGE-L measures the longest common subsequence between the output summary and the standard summary; and ROUGE-SU* measures the match between 1-grams and 2-gram phrases between the output summary and the standard summary, while allowing for discontinuities between words. In subsequent experiments, these four metrics will be denoted as R-1, R-2, RL, and R-SU*, respectively.

[0153] Table 2 shows the experimental results of the model and all the comparison methods on the two datasets. The higher the ROUGE score, the better the performance of the model. Table 3 shows the model on Twitter.

[10] The degradation experimental results on , where MultiSum is the experimental result of the complete model; -social means removing the social relationship information between posts, -semantic means removing the semantic relationship information; -multi steam encoder removes the multi-stream encoder, thereby prohibiting the aggregation of social context information and semantic information between posts.

[0154] Table 2 Performance of the proposed method and other methods on Twitter and Weibo datasets

[0155]

[0156]

[0157] Table 3 Degradation experimental results of the proposed method on Twitter data

[0158] Twitter data R-1 R-2 RL R-SU* MultiSum 45.95 14.24 43.80 20.62 -social 44.98 13.13 43.08 19.89 -semantic 45.12 13.84 43.23 20.18 -multi steam encoder 44.91 13.01 42.91 19.79

[0159] From the evaluation results in Table 2, it can be seen that by aggregating the social context information of the posts and introducing social relationship information and semantic relationship information, the method provided by the present invention achieves a significant performance improvement. We observed that the performance score of the model dropped after deleting any module, verifying the effectiveness of these modules. Specifically, after deleting the social relationship encoder, the R-2 score dropped by more than 1%, which shows that the inclusion of social relations enables the model to capture more keyword combinations. This phenomenon may indicate that there is a high correlation between the content of the text and its social relations. Deleting the semantic relationship encoder causes a significant drop in R-1 and RL scores. This confirms that integrating semantic relations enables the model to learn more accurate word representations and sequential relations between words. In addition, when we do not use the multi-stream encoder, the ROUGE score is lower. This shows that by combining social and semantic relations, the multi-stream encoder improves the post representation.

[0160] The above content is intended to illustrate the technical solution of the present invention, and the present invention is not limited to the embodiments described above. Without departing from the scope of the present invention and the scope of protection of the claims, those skilled in the art can make many forms of specific changes under the guidance of the present invention, all of which fall within the scope of protection of the present invention.

[0161] References:

[0162] [1]Yihong Gong and Xin Liu.2001.Generic Text Summarization UsingRelevance Measure and Latent Semantic Analysis.In Proceedings of the 24thAnnual International ACM SIGIR Conference on Research and Development inInformation Retrieval.19–25

[0163] [2]Gunes Erkan and Dragomir Radev.2011.LexRank:Graph-based LexicalCentrality As Salience in Text Summarization.Journal of ArtifcialIntelligence Research 22(Sept.2011),457–479

[0164] [3]Z.He,C.Chen,J.Bu,C.Wang,L.Zhang,D.Cai,and X.He.2012.Documentsummarization based on data reconstruction.In Twenty-sixth AAAI Conference onArtifcial Intelligence.620–626

[0165] [4]He Liu,Hongliang Yu,and Zhi-Hong Deng.2015.Multi-DocumentSummarization Based on Two-Level Sparse Representation Model.In Proceedingsof the Twenty-Ninth AAAI Conference on Artifcial Intelligence.196–202

[0166] [5]Ruifang He and Xingyi Duan.2018.Twitter Summarization Based onSocial Network and Sparse Reconstruction.In Proceedings of the Thirty-SecondAAAI Conference on Artifcial Intelligence.5787–5794

[0167] [6]Hao Zheng and Mirella Lapata.2019.Sentence Centrality Revisitedfor Unsupervised Summarization.In Proceedings of the 57th Annual Meeting ofthe Association for Computational Linguistics.6236–6247

[0168] [7]Baobao Chang Kexiang Wang and Zhifang Sui.2020.A Spectral Methodfor Unsupervised Multi-Document Summarization.In Proceedings of the 2020Conference on Empirical Methods in Natural Language Processing.435–445

[0169] [8]Huanyu Liu,Ruifang He,Liangliang Zhao,Haocheng Wang,and RuifangWang.2021.SCMGR:Using Social Context and Multi-Granularity Relations forUnsupervised Social Summarization.In Proceedings of the 30th ACMInternational Conference on Information&Knowledge Management.1058–1068

[0170] [9]Xuan-Dung Doan,Le-Minh Nguyen,and Khac-Hoai Nam Bui.2022.MultiGraph Neural Network for Extractive Long Document Summarization.InProceedings of the 29th International Conference on ComputationalLinguistics.5870–5875

[0171]

[10] Ruifang He,Liangliang Zhao,and Huanyu Liu.2020.TWEETSUM:Eventoriented Social Summarization Dataset.In Proceedings of the 28thInternational Conference on Computational Linguistics.5731–5736。

Claims

1. A social media text summarization method that comprehensively considers semantic content, social relationships, and multiple text attributes, characterized by: The following steps are involved: (1) Multi-relationship graph construction: Based on sociological theory, social relationship graphs and semantic relationship graphs are constructed. These graphs explore Social and semantic relations between posts; to build these graphs, we first obtain the representation of the posts and then use GCN aggregates this information; (2) Multi-stream encoder: We propose a multi-stream encoder that exploits the sociological and semantic relationships between posts, where It includes two independent GCN encoders: a social relation encoder and a semantic relation encoder. The encoder helps alleviate the information deficiency of individual posts by integrating social signals and textual relationships; (3) Multi-stream reconstruction: After obtaining the post representation that integrates social information and content relevance, the model is allowed to The reconstruction module reconstructs the representation of the post. We reconstruct the original text, social and semantic relationship graphs, and summary representations of the post respectively. (4) Multi-aspect Summarizer: A centrality ranking model based on aspect perception is proposed by introducing sentence-document Weights allow the model to focus more on different aspects. Weights are added to the centrality score of the post, allowing the summary Covering a wider range of aspects.

2. A social media text summarization method that considers semantic content, social relationships, and multiple text attributes according to claim 1, characterized in that: Step (1) specifically includes: (101) Building a social relationship network at the post level The post-level social relationship network consists of a node set and an edge set. Each node in the node set represents a post, and each edge in the edge set represents the social relationship between corresponding posts. There are two types of social relationships between posts: expression consistency relationship and expression contagion relationship. Expression consistency relationship refers to the relationship between posts published by the same user. When constructing the post-level social relationship network, an edge is established between post nodes with expression consistency relationship. Expression contagion relationship refers to the relationship between posts published by users with direct interaction relationship, among which direct interaction relationship refers to the interaction relationship of following, forwarding, and commenting between users. When constructing the post-level social relationship network, an edge is established between post nodes with expression contagion relationship. The post-level social relationship network is formally described as follows: Represents a post set, N is the number of posts, where s i (1≤i≤N) represents the i-th post; Represents a user set, which contains M users in total, where u i (1≤i≤M) represents the i-th user; for user u i ,make Represents user u i Neighbor users, that is, users u i A collection of users with direct social relationships; Indicates all the i A collection of published posts; a post-level social relationship network is constructed according to the following rules in Represents a set of nodes, each node corresponds to a post, ε represents the set of edges between nodes, and each edge corresponds to the set of social relations between posts; in expressing consistent social relations: if a post where u k represents the kth user, then the post s i With posts j Create an edge e between ij ∈ε; in expressing infectious social relations: if the post And there are or Then for posts i With s j Create an edge e between ij ∈ε; Based on these two rules, a post-level social relationship network is constructed It only contains the post node collection And the relationship between nodes, that is, the edge set ε={e 11 ,e 12 ,…,e NN }; Constructed post-level social relationship network The corresponding adjacency matrix is ​​recorded as Among them A ij >0 indicates post node s i With s j There is a social relationship between them, otherwise A ij =0; (102) Constructing a post-level semantic relationship network The semantic relationship graph explores the semantic relationship between posts. First, we need to obtain the representation of the posts. Specifically, for each post s in the set S i , we input it into the pre-trained BERT model. Then, we take the output of the sentence start token ([CLS]) in the last layer of the model as the embedding of the post. In this paper, we use the BERTWEET model, which has a similar structure to BERT but is trained on a large amount of tweet data, making it more suitable for processing social media posts. The representation of each post is obtained by the following formula: x i =BERTWEET(s i ) (1) Among them, x i Indicates posts i The initial content representation of . Next, we use the cosine similarity between posts as the weight of the edge between them. Based on this method, we construct E sem , and use the adjacency matrix A sem ∈R N×N To represent it. In this way, the semantic relationship graph It is constructed based on the semantic similarity between posts.

3. The social media text summarization method considering semantic content, social relationships and multiple text attributes according to claim 1 is characterized in that: The multi-stream encoder in step (2) is used to aggregate the features of the nodes to learn the representation of the nodes in the graph structure data; specifically, the graph convolutional network GCN is used to aggregate the features of the neighboring nodes to learn the representation of the nodes in the graph structure data. (201) Aggregating multi-relational graph information using graph convolutional networks Taking the social relationship graph as an example, represents the content features of the node extracted by the content encoder, D is the dimension of the content features of each node, It represents the adjacency matrix corresponding to the constructed post-level social relationship network; the graph convolutional network combines the node feature X with the adjacency matrix A of the post-level social relationship network soc As input, the information is propagated according to the following rules: Among them H (0) =X, It is the matrix obtained by adding the adjacency matrix to the unit matrix and then normalizing it; I is the unit matrix, It's A soc +I degree matrix; W (l) with b (l) is the learning parameter of the lth layer; in each layer of the graph convolutional network, each node is normalized by the adjacency matrix To aggregate the features of its adjacent nodes, use W (l) with b (l) The features of neighbor nodes are linearly mapped and then passed through the activation function to obtain the aggregated neighbor node features. Each layer of the graph convolutional network is replaced with a residual connection. This residual connection allows the input data to directly transmit information to the output. Therefore, the specific propagation rules become as follows: here is the node representation after encoding, Z is the dimension of the node representation; σ(·) represents the activation function, Is a mapping function responsible for mapping node features to semantic space Here, a fully connected layer with ReLU activation function is used as the mapping function. Similarly, the same method can be used to obtain the representation of the fusion semantic relationship graph By concatenating the outputs of the two encoders, the final representation includes both social relationship information and semantic information.

4. The social media text summarization method according to claim 1, which considers semantic content, social relationships and multiple text attributes, is characterized in that: In step (3), the multi-stream reconstructor reconstructs the social and semantic relationship graphs of the post, the original text, and the summary representation respectively, as follows: (301) Social Relationship Decoder The social relationship decoder is used to capture the social relationship information of posts in the post-level social relationship network. The social relationship decoder takes the representation of the nodes obtained by the graph convolutional network encoding as input and predicts the social relationship between the input nodes. Formally speaking, the representation of a given node H = [h1,…,h N ] T , the social relationship decoder predicts the social relationship between nodes according to the following formula: Among them, Dec soc (·) is the social relationship decoder function, is the post-level social relationship network predicted by the decoder, Indicates post p i With post p j The probability of having a social relationship connection, h i With h j Corresponding to post p i With p j The vector representation of ; the vector inner product is used as the social relationship decoder function, which predicts the probability of a social relationship connection between two nodes based on the inner product of the vector representation between the nodes: Dec soc (h i ,h j )=P(A i,j |h i ,h j )=σ(h i ·h j ) (7) Where σ(·) represents the activation function; during the training process, the prediction result is minimized The binary cross entropy between the social relationship network structure A at the standard post level is: Where N has the same meaning as above, representing the number of posts, I[·] is the indicator function, when A i,j >0, I[A i,j ]=1, otherwise I[A i,j ]=0; The reconstruction of the semantic relationship graph is similar to the reconstruction of the social relationship graph. The difference is that during the training process, the mean absolute error is used as the loss function: (302) Raw Text Decoder In order to take the post content into consideration more deeply and capture the relationship between different semantic units in the text, we built a raw text decoder. The purpose of this decoder is to understand the content of the post text from different semantic levels. We constructed a semantic relationship bipartite graph in Represents N post nodes, Corresponding to V words in the vocabulary; ε b ={e 11 ,…,e NV } represents the semantic relationship between posts and words. If e ij =1, it means post p i Contains the word w j , otherwise e ij =0; Given the above semantic relationship bipartite graph, the semantic relationship bipartite graph of the post is reconstructed according to the vector representation of the post node; formally, the vector representation of the post node obtained by the graph convolutional network is H = [h1,h2,…,h N ] T As input to the semantic relation decoder, the decoder predicts the semantic relation between the node and the word based on the node representation: Here, Dec se (·) is the semantic relation decoder function; is the prediction result output by the decoder, where V is the vocabulary size, e ik (1≤k≤V)∈[0,1] represents the post p i Contains the word w k The probability of ; a fully connected layer with a sigmoid activation function is used as the semantic relationship decoder function: Dec se (h i )=σ(h i W d +b d ) (11) Where W d with b d is a learning parameter; during the training process, the real semantic relationship bipartite graph is used as supervision and the predicted semantic relationship bipartite graph is minimized and the true semantic relationship bipartite graph ε b The binary cross entropy loss between: (303) Reconstruction of Abstract Representation The last important part in our model is the reconstruction loss of the summary. To measure the quality and accuracy of the summary, we use the average cosine distance (denoted as d cos ), this metric is used to measure the j ) and the encoded summary (h s ) between the hidden states. Specifically, the goal of summary reconstruction is to minimize the difference between the encoded post and summary representations. This is achieved by calculating the cosine distance between them. The average cosine distance is an effective metric because it reflects the proximity between the post and summary in semantic space. A lower average cosine distance means that the encoded summary is closer to the encoded representation of the actual post, indicating a higher quality summary. The reconstruction process of the summary representation can be formalized into the following steps: For each encoded post h j and the encoded summary h s , calculate the cosine distance between them Calculate the average cosine distance between all posts and summaries to get the average cosine distance d cos (h s ,h j ) Optimize the model to minimize the mean cosine distance, so that the summary is semantically closer to the collection of original posts In this way, our model is able to not only extract summaries, but also ensure that the content of the summaries is closely related to the original posts, thereby improving the accuracy and usability of the summaries. This reconstruction process is a key part of the entire summary extraction process, which directly affects the quality and effectiveness of the summaries. The ultimate goal is to learn to simultaneously reconstruct the representation of the post-level social relationship network, semantic network, semantic relationship bipartite graph and summary, and the final training loss is as follows: min L=L se +L soc +L sem +L sim (14) During the training process, the initial representation X of the post is iteratively updated according to the loss function; after the training is completed, the encoder part is used to encode the post to obtain the post representation enhanced with social context, which is applied to the subsequent summary task.

5. The social media text summarization method according to claim 1, which considers semantic content, social relationships and multiple text attributes, is characterized in that: This paper proposes a ranking model based on aspect-aware centrality. By introducing sentence-document weights, the model pays more attention to different aspects. The weights are added to the centrality scores of posts, allowing the summary to cover a wider range of aspects. Step (4) specifically includes: (401) Multifaceted Summarizer Calculating a post i When calculating the centrality score of a post, a threshold ∈ is set to filter out posts with a large semantic gap with the post, thereby reducing the impact of noise. The connection between each post is calculated by the dot product of the two posts, and the centrality score is determined according to the following formula: Among them, e ij There are two posts The dot product of ∈=β·(max(e ij )-min(e ij )), β is a hyperparameter. We rank the posts according to this formula and select the top k ranked posts as summaries. In addition, a different aspect of the model is introduced to measure the relevance by calculating the similarity between the candidate summary c and all posts d. c is selected from the candidate with the highest DC(s i ) score. Combine summary-document similarity with sentence centrality to get the best candidate summary: summary =topK({DC(s i )} i＝1,..,n ) (16) Where topK is the K posts with the highest scores, candidate summaries Is to use DC(s i ) is the subset of the top K sentences after sorting, and α is a hyperparameter that controls the influence of directional centrality. refers to the sentence-document similarity, where g is the document representation, is a candidate summary representation. get This is the average representation of the summary sentences. We choose the cosine similarity of sim(·). For the document representation, we first collect all the sentence representations h1,h2,...,h n In order to compress all valuable information in the document, we apply the Maxpooling function to the sentence representation. The document representation g is calculated as: g=Maxpooling{h1,h2,…,h n } (18) Through the aforementioned multi-aspect summarizer, our model not only improves the performance score of summarization, but also extracts summaries that cover more aspects and reduce bias, thereby achieving more comprehensive and diverse summaries of social media content.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the long-tail word sense disambiguation method integrating decoupled representations as described in any one of claims 1 to 7 are implemented.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the long-tail word sense disambiguation method integrating decoupled representations as described in any one of claims 1 to 7 are implemented.

Images