Multi-dimensional information fusion user stance detection method based on graph attention network

Abstract

This invention discloses a multi-dimensional information fusion method for user stance detection based on graph attention networks, comprising the following steps: preprocessing tweet data, converting user tags to tweet tags using pseudo-labels, and training a text stance detection model; filtering user time-series tweet sequences, obtaining vector representations through the stance detection model, and processing the time-series information using a bidirectional GRU to obtain user speech information features; recognizing text in profile pictures using OCR technology, concatenating it with the user's personal profile, and embedding it through a multilingual pre-trained model to obtain user profile information features; and embedding the user's multi-dimensional features using a graph attention network to obtain user node representations containing neighbor node information. By fusing user speech information features, profile information features, and social relationship features, accurate determination of user stance can be achieved. This invention's detection method has advantages such as rich feature dimensions and the ability to predict "silent users."

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, specifically to a multi-dimensional information fusion method for user stance detection based on graph attention networks. Background Technology

[0002] With the development of internet social media, social networks have become ubiquitous. According to the "2022 Global Digital Atlas" report, active social media users account for 58.4% of the global population, with each person spending more than 2.5 hours daily on different platforms. Despite its global reach, the number of social media users continues to rise, increasing by 10.1% in 2022 compared to the previous year. Different social media networks offer users a variety of functions, such as creating communities for like-minded individuals and providing an easy way to share ideas. These mainstream social media platforms all provide users with the ability to share content and their own opinions. The report's statistics show that 29.5% of people use social media to stay informed about what's being discussed, and 24.5% use it to share and discuss opinions with others, providing a basis for analyzing user stances on social media.

[0003] Social network user stance detection is an intersection of natural language processing (NLP) and computational social science. NLP focuses on enabling computers to understand natural language, while computational social science utilizes big data and online social activity to model and analyze social behavior, relationships, and interactions. User stance, representing users' opinions on certain topics—whether "supporting," "opposing," or "neutral"—is a crucial topic with wide applications in politics, social sciences, user behavior analysis, and social media analysis. In public opinion surveys, stance detection is considered more effective at aligning with public opinion and can be seen as a replacement or supplement to traditional opinion measurement tools. By analyzing stances on specific targets, companies can gather public opinion online to improve and upgrade products, enhancing user experience; regulatory agencies can identify malicious accounts, clean up platform environments, and provide a healthy online environment.

[0004] Most current research uses user-generated content (tweets or blog posts) for stance prediction, relying solely on textual features. Recent studies have incorporated user social relationships and behavioral interactions to aid in stance determination. However, current methods suffer from several drawbacks: relying solely on textual features requires large amounts of labeled datasets, ignores the temporal dimension of user statements, and fails to address the "silent user" problem; using interaction relationship mapping introduces redundancy, and irrelevant interactions may negatively impact detection performance; furthermore, they don't effectively utilize user profile information, such as personal introductions and avatars. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a multi-dimensional information fusion user stance detection method based on graph attention networks. This method acquires and fuses feature information from three dimensions: user statements, user profiles, and user social relationships. It has advantages such as rich feature dimensions and the ability to predict "silent users." The technical solution is as follows:

[0006] A multi-dimensional information fusion user stance detection method based on graph attention networks includes the following steps:

[0007] Step 1: Preprocess the tweet data and use a publicly available text stance prediction model to predict all tweet information to obtain predicted pseudo-labels;

[0008] Step 2: Construct a text stance detection model, compare the predicted pseudo-labels with user labels, and use the data with consistent labels as the dataset to train the tweet-level text stance detection model;

[0009] The LaBSE multilingual pre-trained model maps words from different languages ​​to the same vector space. The tweet is segmented into sentences, and word and sentence embeddings are performed by the LaBSE multilingual pre-trained model to obtain the tweet's representation vector. The target text sequence is encoded by the LaBSE multilingual pre-trained model to obtain its feature vector, and then averaged to obtain the target information feature vector. The words in the sentence are encoded by bidirectional LSTM to obtain word embedding vectors. The target information feature vector and the word embedding vector are concatenated and then input into a linear transformation function to calculate the weight coefficient of each word. Finally, matrix operations are performed to obtain the semantic feature representation vector of each sentence and the global semantic representation feature of the tweet.

[0010] Step 3: Filter the S tweets related to the target from the user comments, sort them by time to obtain the text sequence, use the trained text stance detection model to model the tweet vector hidden layer representation; then obtain the s-th tweet through the bidirectional GRU module. Forward hidden state and backward hidden state The final state output is determined by... and It is pieced together, that is After sequence training is completed, the final user speech information features are obtained. ;

[0011] Step 4: Model the user profile information. The personal introduction content is embedded using the multilingual pre-trained model LaBSE. The avatar information is concatenated with the personal introduction and modeled together to obtain the user profile information features. ;

[0012] Step 5: Use a graph attention network to embed the user's multi-dimensional features to obtain a user position representation that includes the neighbor's position information. ;

[0013] Step 6: Feature the user comments obtained in Step 3 Features of user profile information obtained in step 4 The user's stance representation obtained by aggregating neighbor information in step 5 The final vector representation R is obtained by concatenating the vectors. The result of the entire text stance detection model is output through a feedforward network layer, which performs a linear transformation.

[0014] Furthermore, step 1 specifically includes:

[0015] Step 1.1: Text preprocessing: Remove invalid information from the collected tweets and standardize the information in the tweets; replace the emojis in the tweets with words of the same meaning, and restore the tags to words or phrases;

[0016] Step 1.2: Pseudo-label prediction: The preprocessed tweet data is predicted using a publicly available text stance prediction model. A hierarchical attention network that performs well on the SemEval2016-Task6 dataset is used for prediction to obtain the predicted labels for all tweets.

[0017] Furthermore, step 1.1 specifically includes:

[0018] Step 1.1.1: Emoji Interpretation: Use Python's third-party library emoji to replace the emotions represented by emojis with words that have similar meanings;

[0019] Step 1.1.2: Standardized cleaning: Replace the URLs, dates, and email information in the tweets with categories;

[0020] Step 1.1.3: Tag Segmentation: For tweets containing hashtags composed of multiple words, the words are split and used as part of the tweet's semantics to extract the semantic information of the hashtags.

[0021] Furthermore, step 2 specifically includes:

[0022] Step 2.1: Tweet Sentence segmentation is performed, and word embedding and sentence embedding are conducted using the multilingual pre-trained model LaBSE; the model input is text data. , The tweet indicates the first 1 sentence, of which , The tweet indicates the first The first sentence One word;

[0023] Step 2.2: The target text sequence is represented as follows The feature vector is obtained after encoding by the multilingual pre-trained model LaBSE. The target information feature vector after mean normalization :

[0024] ;

[0025] in, Represents the first of the target text sequence One word; Indicates the first The feature vectors of each word; n represents the length of the target text sequence;

[0026] Step 2.3: Transfer the target information feature vector The word embedding vectors are concatenated with the bidirectional LSTM encoded word embedding vectors and then input into a linear transformation function to calculate the weight coefficients of each word.

[0027] ;

[0028] ;

[0029] ;

[0030] in, Indicates the first The first sentence The feature vector of each word after bidirectional LSTM encoding; The result of concatenating the feature vector of the word with the target feature vector; This is a vector representation of the weighting coefficients; It is a weight matrix. For bias terms; For activation functions; This indicates that the eigenvectors are being normalized. This represents the normalized weight coefficients;

[0031] Step 2.4: Target information attention assigns different weights to each word, and finally, matrix operations are used to obtain the semantic feature representation vector of each sentence. Global semantic feature representation vector of tweets :

[0032] ;

[0033] ;

[0034] in, Let N be the length of the sentence sequence and N be the number of sentences.

[0035] Furthermore, step 5 specifically includes:

[0036] Step 5.1: Construct a user social relationship network, using users as nodes and social relationships as edges, and user speech information features. and user profile information characteristics As node features, they form a directed isomorphic social graph;

[0037] The directed isomorphic social graph is represented as: , It is a collection of user nodes. It is a set of edges; node features are represented as Where H is the number of nodes and F is the dimension of the node features. Features for each node;

[0038] Step 5.2: Identify user speech information features and archival information characteristics The node features of node h are obtained by splicing. :

[0039] ;

[0040] Step 5.3: Obtain node features through linear transformation Advanced expressions :

[0041] ;

[0042] in, It is a weight matrix, composed of user speech information features. and user profile information characteristics obtained by piecing together The feature representation of node h after transformation. This indicates that this is a response to Layer calculation;

[0043] Step 5.4: Calculate the importance of each of node h's first-degree neighbor nodes g, and use... As the activation function, the attention score is obtained, and then... Convert attention scores into probabilities;

[0044] ;

[0045] ;

[0046] in, It is a single-layer feedforward neural network. The attention score represents the importance of node g to node h; This represents the normalized attention coefficient. The feature representation of node g after transformation; Let h be the set of all neighboring nodes of node h;

[0047] Step 5.5: Calculate a weighted average of the attention coefficients and node features, and then pass it through an activation function. Obtain the output features;

[0048] ;

[0049] Step 5.6: After performing a linear transformation on the original features, stack them with the attention values ​​obtained through the network to obtain the output features:

[0050] ;

[0051] in, Indicates the first The normalized weight coefficients are calculated using the group attention mechanism. It is the first The weight matrix for group attention calculation; Indicates a splicing operation; The node characteristics of node g;

[0052] Using a single-layer graph attention network, the hidden layer output for a single user is represented as follows: , For the number of attention heads, For residuals, i.e. The final user position representation is obtained by aggregating neighbor positions through a graph attention network. .

[0053] Compared with the prior art, the beneficial effects of the present invention are:

[0054] 1) This invention proposes a scheme for detecting user stance based on the fusion of three dimensions of information: user-generated content, user profile information, and user social relationship information. This enriches the features used for stance determination and improves the determination effect.

[0055] 2) This invention proposes a solution for tweet-level stance detection when only user tags are available and text tags are lacking: pseudo-tags are used to convert user tags to tweet tags, avoiding the need for extensive manual data annotation. Simultaneously, a multilingual pre-trained model is used for word embedding, effectively addressing the multilingual issue.

[0056] 3) This invention leverages the temporal order of user tweets, embedding them into a time sequence to obtain a sequence showing how tweet stances change over time. A recurrent neural network is then used to predict the user's latest stance. Compared to summarizing all user tweets, this method offers better interpretability.

[0057] 4) This invention uses graph attention networks to break free from the constraints of the Laplacian matrix, making it more suitable for directed graph computation. It incorporates users' social relationships, capturing the influence of neighboring users on their own stance, improving detection accuracy, and also enabling the prediction of "silent users." Attached Figure Description

[0058] Figure 1 This is a user stance detection framework diagram of the present invention.

[0059] Figure 2 This is a specific implementation process for text-level position detection of the present invention.

[0060] Figure 3 This is a specific implementation process for user profile information processing according to the present invention.

[0061] Figure 4 This is a specific implementation process of the present invention based on graph attention network to fuse user speech information, user profile information and user social relationship information.

[0062] Figure 5 This is a structural diagram of the user stance detection model of the present invention. Detailed Implementation

[0063] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more explicit definition of the scope of protection of the present invention.

[0064] like Figure 1As shown, the method of this invention takes user speech information, user profile information, and user social relationship information as inputs and outputs user stance detection results. First, the dataset is classified: all tweet data undergoes text preprocessing, including emoji interpretation, standardization cleaning, and tag segmentation. Then, existing models are used to predict pseudo-tags and compare them with user tags to construct an annotated text dataset for training the text stance detection model. The trained model is then used to model user text sequences. Next, user profile information is processed: Baidu OCR is used to recognize text information in user avatars, which is then concatenated with user personal profile information and modeled using a pre-trained model. Finally, a relationship graph is constructed from user social relationship data, using user speech information and profile information as node features. Information fusion is performed based on a graph attention network to obtain the final vector representation of the user, which is then used to train the user stance detection model. Details are as follows:

[0065] Step 1: Preprocess the tweet data and use a publicly available text stance prediction model to predict all tweet information to obtain pseudo-labels.

[0066] 1) Text Preprocessing; The text preprocessing of this invention mainly includes emoji interpretation, standardization cleaning, and tag segmentation. Invalid information such as special punctuation marks in the collected tweets is removed, and various types of information in the tweets are standardized. Emojis in the tweets are replaced with words of the same meaning, and tags are restored to words or phrases. In addition, URLs, usernames, etc., are uniformly identified by category.

[0067] a) Expression interpretation:

[0068] Tweets often contain a large number of emojis, which contain rich emotional information. Emotional information is of great reference value for detecting the stance of the text. We can use the Python third-party library emoji to replace the emotions represented by the emojis with words of similar meaning.

[0069] b) Standardized cleaning:

[0070] The URLs, dates, and email information in the tweets differ significantly from the text content, which may affect the language. These elements will be uniformly replaced with categories; for example, URLs will be replaced with... <http>Date replacement <datetime>Replace email information with <email>.

[0071] c) Tag segmentation:

[0072] Tweets often contain hashtags composed of multiple words. To better extract semantic information from these hashtags, the words in the hashtags are broken down and incorporated into the tweet's semantics. This is achieved using the Python third-party library ekphrasis.

[0073] 2) Pseudo-label prediction: The preprocessed tweet data is predicted using a publicly available text stance prediction model. A hierarchical attention network that performs well on the SemEval2016-Task6 dataset is used for prediction to obtain the predicted labels for all tweets.

[0074] Step 2: Build a text stance detection model, construct a labeled dataset, and train a tweet-level stance detection model.

[0075] 1) Dataset Construction. The predicted labels obtained in step 1 are compared with the user labels corresponding to the tweets. Data with matching labels are added to the dataset as labeled data for training the text detection model. This solves the problem of converting user labels to tweet labels when only user labels are available and tweet labels are unavailable, avoiding the manual cost of labeling tweet data and improving efficiency. A labeled tweet dataset is constructed and divided into training, validation, and test sets in a 6:2:2 ratio.

[0076] 2) Model Construction and Training. The framework of the text stance detection model consists of three parts: word embedding, target attention, and global semantic representation. For word embedding, the multilingual pre-trained model LaBSE is used to map words from different languages ​​to the same vector space. The process of obtaining the global semantic representation of the text data through modeling tweets is as follows:

[0077] a) Tweet Sentence segmentation is performed, and word embedding and sentence embedding are conducted using a pre-trained model. The model input is text data. , The tweet indicates the first 1 sentence, of which , The tweet indicates the first The first sentence One word.

[0078] b) The target text sequence is represented as The feature vector is obtained after encoding by the pre-trained model. The target information feature vector after mean normalization .

[0079] ;

[0080] c) Transfer the target information feature vector The word embedding vector is concatenated with the bidirectional LSTM encoded word embedding vector and then input into a linear transformation function to calculate the weight coefficient of each word.

[0081] ;

[0082] ;

[0083] ;

[0084] in, Indicates the first The first sentence The feature vector of each word after bidirectional LSTM encoding; These are the weighting coefficients; It is a weight matrix. For bias terms;

[0085] d) Target information attention assigns different weights to each word, and finally, through matrix operations, obtains the semantic feature representation vector of each sentence and the global semantic feature representation vector of the tweet. :

[0086] ;

[0087] ;

[0088] in, Let N be the length of the sentence sequence, and N be the number of sentences. The final sentence sequence is obtained after calculation. The representation vector is The global semantic features of the tweet are represented as .

[0089] e) The model's output layer consists of a fully connected layer and Functional composition, the final representation of a tweet. After the function is computed through the fully connected layer, then... The probability of the output stance category of the layer. .

[0090] ;

[0091] in, These are the weight matrix and bias term, obtained through training. The model is trained using the cross-entropy loss function.

[0092] ;

[0093] in, , This represents the index of the training set data and the position classification index, with the input parameter being the tweet text. Authentic tags in tweets And predictions for different stance categories ,function This represents a text stance detection model. This represents the L2 regularization term.

[0094] f) The multilingual pre-trained model LaBSE freezes all layer parameters and uses them only as the encoder. The hidden layer parameters are initialized using the Xavier normal distribution, and then the model is trained using the tweet dataset.

[0095] The implementation process of text-level stance detection is as follows: Figure 2 As shown, steps 1 and 2 were used to build and train the text stance detection model, which is then used for modeling user speech information features.

[0096] The process of multi-dimensional information fusion based on graph attention networks is as follows: Figure 4 As shown, firstly, step 3 obtains the user's temporal speech information features; then, step 4 obtains the user's profile information features; finally, step 5 constructs the user's social relationship network, and the speech features and profile features are concatenated as node features. Neighbor node information is then aggregated based on a graph attention network to obtain a user node representation containing neighbor node information. The specific steps for obtaining user information in various dimensions are as follows:

[0097] Step 3: Filter the S tweets that are relevant to the user and the target, and sort them by time to obtain the text sequence. The model trained in step 2 is used for modeling to obtain the hidden layer representation of the tweet vector. Then, the s-th tweet is obtained through a bidirectional GRU module. Forward hidden state and backward hidden state :

[0098] ;

[0099] The final state output is determined by and It is pieced together, that is After sequence training is completed, the final user speech information features are obtained. :

[0100] ;

[0101] Step 4: Model the user profile information. For example... Figure 3 As shown, first, the baidu-aip library is installed, and AipOcr is called to perform text recognition on all users' avatars to obtain the text information in the avatars. Then, it is concatenated with the personal profile information. If the recognition result does not contain text information, only the personal profile information is used, and it is embedded using a multilingual pre-trained LaBSE model. All parameters are frozen when using the pre-trained LaBSE model, which is used as the encoder. This yields the user profile information features. .

[0102] Step 5: Embed the user's multi-dimensional features using a graph attention network to obtain a user node representation that includes information about neighboring nodes. Graph attention networks do not rely on the Laplacian matrix; they use an attention mechanism to calculate node weights instead of convolution operations when aggregating node features, making them suitable for graph representations in social networks. During computation, the network traverses the graph each time, calculating the attention score between each node and its neighboring nodes, thus freeing it from the constraints of the Laplacian matrix and making it more suitable for directed graph computation.

[0103] 1) Construct a user social relationship network, using users as nodes, social relationships as edges, and user speech and profile information as node features, forming a directed isomorphic social graph. During construction, deduplication of nodes and relationships is required to ensure the uniqueness of user nodes and edges, and self-loops are added to each node.

[0104] Social graphs can be represented as , It is a collection of user nodes. It is a set of edges. Node characteristics can be represented as... Where H is the number of nodes and F is the dimension of the node features. For each node, the features are derived from the user speech information features obtained in steps 3 and 4. and archival information characteristics The result obtained by piecing together:

[0105] ;

[0106] 2) Aggregate neighbor information using a graph attention network. First, obtain node features through a linear transformation. Advanced expressions :

[0107] ;

[0108] in, It is a weight matrix. This is the transformed feature representation. This indicates that this is a response to Layer calculation.

[0109] Then, for node h, the importance of each of its first-degree neighbor nodes g is calculated, and used... As the activation function, the attention score is obtained, and then... Convert attention scores into probabilities:

[0110] ;

[0111] ;

[0112] in, It is a single-layer feedforward neural network. The attention score represents the importance of node g to node h; This represents the normalized attention coefficient. The feature representation of node g after transformation; Let h be the set of all neighboring nodes of node h.

[0113] After obtaining the attention coefficients, they are used as weights to calculate a weighted average with node features, and then passed through an activation function. Output features are obtained:

[0114] ;

[0115] Multi-head attention mechanisms can capture richer user representations and improve the generalization ability of graph attention networks. This involves processing the input features... To reduce the influence of the network on the original features, the original features are linearly transformed and then stacked with the attention values ​​obtained through the network to obtain the output features.

[0116] ;

[0117] in, Indicates the first The normalized weight coefficients are calculated using the group attention mechanism. It is the first The weight matrix for group attention calculation. A single-layer graph attention network is used, with eight attention heads and a hidden layer size of 768. The output of this layer for a single user is represented as follows. , For the number of attention heads, For residuals, i.e. The final user position representation, which aggregates the positions of neighbors, is obtained through a graph attention network. .

[0118] Step 6: As Figure 4 The diagram shows the architecture of the user stance detection model, which incorporates the user speech information features obtained in step 3. Features of user profile information obtained in step 4 The user's stance representation obtained by aggregating neighbor information in step 5 The final vector representation R is obtained by concatenating the vectors. The feedforward network layer performs a linear transformation and outputs the result of the entire model. This layer contains a ReLU activation function and a linear layer. The calculation method of this layer is as follows:

[0119] ;

[0120] in, The weights of the fully connected layer, This is for bias. The loss function used is the cross-entropy loss function, and the loss is calculated as follows:

[0121] ;

[0122] in, For users' real tags, Predicted labels for users.

[0123] After the user social graph is constructed, the user dataset is divided into training and test sets in a 7:3 ratio. In the actual prediction process, user speech and profile information may be unknown, so relevant information is removed from the test set. During training, its node features are no longer obtained through steps 3 and 4, but are set to default values. During prediction, user speech and profile information features are obtained through steps 3 and 4 to predict nodes. Alternatively, only the profile information obtained in step 4 can be used, omitting speech information, to predict users as "silent users."

[0124] The training process was repeated 10 times independently, and the average result was calculated. In each training iteration, the model had 50 epochs, a batch size of 32, a learning rate of 0.01, and used the Adam optimizer. The F1 score was calculated at the end of each epoch; training was terminated early if the F1 score did not increase after 10 consecutive epochs.< / email> < / datetime> < / http>

Claims

1. A multi-dimensional information fusion user stance detection method based on graph attention networks, characterized in that, Includes the following steps: Step 1: Preprocess the tweet data and use a publicly available text stance prediction model to predict all tweet information to obtain predicted pseudo-labels; Step 2: Construct a text stance detection model, compare the predicted pseudo-labels with user labels, and use the data with consistent labels as the dataset to train the tweet-level text stance detection model; The LaBSE multilingual pre-trained model maps words from different languages ​​to the same vector space. Tweets are segmented into sentences, and word and sentence embeddings are performed using the LaBSE multilingual pre-trained model to obtain the representation vector of the tweet. The target text sequence is encoded by the LaBSE multilingual pre-trained model to obtain its feature vector, and then averaged to obtain the target information feature vector. Words in the sentence are encoded by bidirectional LSTM to obtain word embedding vectors. The target information feature vector and the word embedding vector are concatenated and then input into a linear transformation function to calculate the weight coefficient of each word. Finally, matrix operations are performed to obtain the semantic feature representation vector of each sentence and the global semantic representation feature of the tweet. Step 3: Filter the S tweets related to the target from the user comments, sort them by time to obtain the text sequence, use the trained text stance detection model to model the tweet vector hidden layer representation; then obtain the s-th tweet through the bidirectional GRU module. Forward hidden state and backward hidden state The final state output is determined by... and It is pieced together, that is After sequence training is completed, the final user speech information features are obtained. ; Step 4: Model the user profile information. The personal introduction content is embedded using the multilingual pre-trained model LaBSE. The avatar information is concatenated with the personal introduction and modeled together to obtain the user profile information features. ; Step 5: Use a graph attention network to embed the user's multi-dimensional features to obtain a user position representation that includes the neighbor's position information. ; Step 6: Feature the user comments obtained in Step 3 Features of user profile information obtained in step 4 The user's stance representation obtained by aggregating neighbor information in step 5 The final vector representation R is obtained by concatenating the vectors. The result of the entire text stance detection model is output through a feedforward network layer, which performs a linear transformation.

2. The user stance detection method based on graph attention network for multi-dimensional information fusion according to claim 1, characterized in that, Step 1 specifically includes: Step 1.1: Text preprocessing: Remove invalid information from the collected tweets and standardize the information in the tweets; replace the emojis in the tweets with words of the same meaning, and restore the tags to words or phrases; Step 1.2: Pseudo-label prediction: The preprocessed tweet data is predicted using a publicly available text stance prediction model. A hierarchical attention network that performs well on the SemEval2016-Task6 dataset is used for prediction to obtain the predicted labels for all tweets.

3. The user stance detection method based on graph attention network for multi-dimensional information fusion according to claim 2, characterized in that, Step 1.1 specifically includes: Step 1.1.1: Emoji Interpretation: Use Python's third-party library emoji to replace the emotions represented by emojis with words that have similar meanings; Step 1.1.2: Standardized cleaning: Replace the URLs, dates, and email information in the tweets with categories; Step 1.1.3: Tag Segmentation: For tweets containing hashtags composed of multiple words, the words are split and used as part of the tweet's semantics to extract the semantic information of the hashtags.

4. The user stance detection method based on graph attention network for multi-dimensional information fusion according to claim 1, characterized in that, Step 2 specifically includes: Step 2.1: Tweet Sentence segmentation is performed, and word embedding and sentence embedding are conducted using the multilingual pre-trained model LaBSE; the model input is text data. , The tweet indicates the first 1 sentence, of which , The tweet indicates the first The first sentence One word; Step 2.2: The target text sequence is represented as follows The feature vector is obtained after encoding by the multilingual pre-trained model LaBSE. The target information feature vector after mean normalization : ； in, Represents the first of the target text sequence One word; Indicates the first The feature vectors of each word; n represents the length of the target text sequence; Step 2.3: Transfer the target information feature vector The word embedding vectors are concatenated with the bidirectional LSTM encoded word embedding vectors and then input into a linear transformation function to calculate the weight coefficients of each word. ； ； ； in, This represents the i-th sentence. The feature vector of each word after bidirectional LSTM encoding; The result of concatenating the feature vector of the word with the target feature vector; This is a vector representation of the weighting coefficients; It is a weight matrix. For bias terms; For activation functions; This indicates that the eigenvectors are being normalized. This represents the normalized weight coefficients; Step 2.4: Target information attention assigns different weights to each word, and finally, matrix operations are used to obtain the semantic feature representation vector of each sentence. Global semantic feature representation vector of tweets : ； ； in, Let N be the length of the sentence sequence and N be the number of sentences.

Images