Aspect emotion triple extraction method based on multi-channel graph convolutional network

Abstract

The invention discloses an aspect emotion triple extraction method based on a multi-channel graph convolutional network, and relates to a graph convolutional neural network architecture extraction method, and the network architecture comprises three core modules: a feature output module, a multi-channel graph convolutional network module and a table decoding module. The framework effectively utilizes the advantages of BERT, BiLSTM and a prompt learning mechanism to carry out feature extraction. Through parallel three-channel design, word pair relations under positive, negative and neutral sentiments are modeled respectively, so that sentiment information in a text can be extracted and expressed more comprehensively. The multi-channel design not only improves the expression ability of the model, but also enhances the robustness of the model when processing complex emotion information. The method provides core power for developing a user comment sentiment analysis system, and the new technology has far-reaching influence on timely product and service improvement of merchants and personal optimized consumption decision.

Description

Technical Field

[0001] The present invention relates to a graph convolutional neural network architecture extraction method, and in particular to an aspect sentiment triple extraction method based on a multi-channel graph convolutional network. Background Art

[0002] Driven by the rapid technological development of the digital age, the number of comments on social media platforms has surged, generating a vast amount of textual data containing user emotions and opinions, providing rich material for sentiment analysis. However, traditional holistic sentiment analysis methods (e.g., positive, neutral, and negative) are no longer sufficient for analyzing complex data. More fine-grained sentiment analysis methods for textual data are urgently needed, leading to the emergence of the Aspect Sentiment Triplet Extraction (ASTE) task.

[0003] ASTE aims to extract triplets from a given sentence, consisting of aspects, opinions, and emotions, to facilitate more detailed sentiment analysis. While research in this field has made significant progress in recent years, it still faces challenges such as accuracy in processing large datasets, insufficient mining of syntactic and semantic information, lack of interaction between cells in relational tables, and difficulty in fully capturing the complex relationships of multiple coexisting emotions. Summary of the Invention

[0004] The purpose of the present invention is to provide an aspect-sentiment triplet extraction method based on a multi-channel graph convolutional network. The method is based on the aspect-sentiment triplet extraction architecture of a multi-channel graph convolutional network, which introduces BERT, a bidirectional long short-term memory network (BiLSTM) and a prompt learning mechanism, and performs graph convolution operations under different sentiment dimensions. It aims to improve the performance of the ASTE task and the robustness when processing complex sentiment information, thereby advancing the state-of-the-art in fine-grained sentiment analysis.

[0005] The purpose of the present invention is achieved through the following technical solutions:

[0006] A method for extracting aspect sentiment triples based on a multi-channel graph convolutional network, comprising the following components:

[0007] (1) Feature output module

[0008] BERT and BiLSTM are used as dual encoders to capture semantic relationships and local dependencies between words, respectively, to improve the model's ability to understand text. Then, by introducing a hint learning mechanism, emotional hint information is directly embedded into the sentence encoding, enabling the model to adaptively learn emotional relationships without relying on external knowledge. First, the dual encoder is used to extract the representation of the sentence. The encoding process is as follows:

[0009]

[0010] in, is the last hidden state output of the encoder, d represents the dimension of the hidden state, H is the representation of the input sentence X, and H T It is the representation of template T. T Contains 6 representations corresponding to the mask slots in the template, namely Represent aspect terms and opinion terms under positive, negative and neutral emotions respectively;

[0011] After obtaining the joint representation of the sentence and the template, the model then calculates the correlation between each word and each slot in the template; the prompt attention mechanism is introduced to calculate the attention score matrix

[0012] (2) Multi-channel graph convolutional network module

[0013] The Tri-Channel Graph Convolution Network (TCGN) is the core module of this architecture. It aims to capture the sentiment relationship between word pairs in a sentence through graph convolution operations under different sentiment dimensions. This module represents the nodes in the graph structure as the input of multi-layer convolution operations. It processes different sentiment information through a parallel three-channel structure (positive, negative, and neutral sentiment) to ultimately generate a comprehensive sentiment relationship representation for each word pair.

[0014] (3) Table decoding module

[0015] Table decoding is the key step in the model's extraction of sentiment triplets from the node representations generated by the three-channel graph convolutional network. This process is divided into two parts: detection and classification. Through these two stages, the model identifies candidate sentiment regions from the node representations and performs sentiment classification on them, thereby generating the final sentiment triplet.

[0016] The method for extracting aspect emotion triples based on a multi-channel graph convolutional network, wherein each row of the attention score matrix of the feature output module corresponds to the embedding of a template slot, and the column corresponds to each word in the input sentence;

[0017] P=softmax(τWH T )

[0018] in, The six embedding representations in the representation template, represents the embedding matrix of the sentence, is a learnable parameter matrix; through matrix multiplication, the model captures the correlation between each word in the sentence and each slot in the template, and normalizes the result through the softmax function to ensure that the value in the attention score matrix is ​​a probability distribution; each element P of the attention score matrix ij represents the correlation between the i-th slot in the template and the j-th word in the sentence.

[0019] The method for extracting aspect sentiment triples based on a multi-channel graph convolutional network is described. The graph convolutional network (GCN) of the multi-channel graph convolutional network module is a neural network commonly used for graph structured data. It updates the representation of each node by aggregating information about each node and its neighboring nodes. In this module, the core idea of ​​the graph convolutional network is to gradually aggregate sentiment information from neighboring nodes through multi-layer convolution operations, thereby enhancing the model's understanding of the relationship between nodes. Given a graph G, the graph convolution operation is performed using the following formula:

[0020]

[0021] in, is c in layer l ij The expression, is a learnable weight matrix, σ is a nonlinear activation function (such as ReLU), Represents the connection node c ij and its neighbor node g k The edge weights between .

[0022] The method for extracting aspect emotion triples based on a multi-channel graph convolutional network is described. The multi-channel graph convolutional network module has a three-channel parallel convolution structure, each channel corresponds to an emotion graph, and independently performs graph convolution operations. In each channel, the model gradually aggregates emotion information through multiple layers of convolution.

[0023] The method for extracting aspect emotion triples based on a multi-channel graph convolutional network is described. Each channel model gradually aggregates emotion information through multi-layer convolution. The specific process is as follows: the positive emotion channel is responsible for processing the positive emotion graph G pos In this channel, the model aggregates the nodes and edge weights related to positive emotions through graph convolution operations and gradually updates the node representations. The final output of the convolution operation is a node representation matrix C′ containing positive emotion information. pos This matrix provides the basis for the positive sentiment part of the subsequent sentiment triplet; the operation of the negative sentiment channel is similar to the positive sentiment channel, but it processes the negative sentiment graph G neg Through graph convolution, the model aggregates node information related to negative emotions and generates a node representation matrix C′ containing negative emotion information neg ; Neutral emotion channel processes neutral emotion graph Gneu , this channel generates the neutral emotion node representation matrix C′ by aggregating the node information of neutral emotion neu .

[0024] After completing the graph convolution operation on the three emotion channels, the model fuses the results of the three channels together through the splicing operation to generate the final node representation matrix C′. The fusion operation is as follows:

[0025]

[0026] The advantages and effects of the present invention are:

[0027] 1. This paper proposes an innovative and efficient deep learning framework for extracting aspect-sentiment triplets. This framework effectively leverages the strengths of BERT, BiLSTM, and a hint learning mechanism for feature extraction. Through a parallel three-channel design, word-pair relationships under positive, negative, and neutral sentiment are modeled separately, enabling a more comprehensive extraction and representation of sentiment information in text. This multi-channel design not only improves the model's expressive power but also enhances its robustness in processing complex sentiment information.

[0028] 2. This invention pushes the boundaries of the field of fine-grained sentiment analysis and provides the core driving force for the development of user review sentiment analysis systems. This new technology has a profound impact on merchants' timely improvement of products and services and individuals' optimized consumption decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is the overall framework diagram of the model of the present invention. DETAILED DESCRIPTION

[0030] The present invention will be described in detail below with reference to the embodiments shown in the accompanying drawings.

[0031] The present invention provides a method for extracting aspect emotion triples based on a multi-channel graph convolutional network, which is constructed as follows:

[0032] Feature output module

[0033] In natural language processing tasks, it is usually necessary to extract multiple types of information from the text, such as grammatical structure and semantic meaning. However, a single encoder can often only focus on one type of information, and therefore cannot fully utilize all the information in the text. Traditional sentiment analysis tasks often rely on direct modeling of specific words, which often ignores the complex interactions between different emotions. In order to solve these problems, the present invention first uses BERT and BiLSTM as dual encoders, which are used to capture the semantic relationships and local dependencies between words, respectively, to improve the model's ability to understand the text. Then, by introducing a hint learning mechanism, the emotional hint information is directly embedded into the sentence encoding, so that the model can adaptively learn emotional relationships without relying on external knowledge. First, the dual encoder is used to extract the representation of the sentence, and its encoding process is as follows:

[0034]

[0035] in, is the last hidden state output of the encoder, d represents the dimension of the hidden state, H is the representation of the input sentence X, and H T It is the representation of template T. T Contains 6 representations corresponding to the mask slots in the template, namely Represent aspect terms and opinion terms under positive, negative and neutral sentiments respectively.

[0036] After obtaining the joint representation of the sentence and the template, the model then calculates the correlation between each word and each slot in the template. To achieve this, the present invention introduces a prompt attention mechanism to calculate the attention score matrix Each row of this matrix corresponds to the embedding of a template slot, and the columns correspond to each word in the input sentence.

[0037] P=softmax(τWH T )

[0038] in, The six embedding representations in the representation template, represents the embedding matrix of the sentence, is a learnable parameter matrix. Through matrix multiplication, the model can capture the correlation between each word in the sentence and each slot in the template, and normalize the results through the softmax function to ensure that the values ​​in the attention score matrix are probability distributions. Each element P of the attention score matrix ij represents the correlation between the i-th slot in the template and the j-th word in the sentence.

[0039] The advantage of the cued attention score mechanism is that it can weight words separately under different emotional polarities, thereby achieving fine-grained perception of the coexistence of multiple emotions; compared with other models that rely on external emotional dictionaries or rules, the cued attention score uses embedded emotional cues to adaptively adjust the weight of each word, thereby demonstrating stronger generalization capabilities in complex contexts.

[0040] Multi-channel graph convolutional network module

[0041] The Tri-Channel Graph Convolution Network (TCGN) is the core module of this architecture. It aims to capture the sentiment relationships between word pairs in a sentence through graph convolution operations at different sentiment dimensions. This module represents nodes in the graph structure as input to multi-layer convolution operations. It processes different sentiment information through a parallel three-channel structure (positive, negative, and neutral sentiment), ultimately generating a comprehensive sentiment relationship representation for each word pair.

[0042] A graph convolutional network (GCN) is a neural network commonly used for graph-structured data. It updates the representation of each node by aggregating information about its neighbors. In this module, the core idea of ​​the GCN is to gradually aggregate sentiment information from neighboring nodes through multi-layer convolution operations, thereby enhancing the model's understanding of the relationships between nodes. Given a graph G, the graph convolution operation is performed using the following formula:

[0043]

[0044] in, is c in layer l ij The expression, is a learnable weight matrix, σ is a nonlinear activation function (such as ReLU), Represents the connection node c ij and its neighbor node g k The edge weights between nodes. In the lth layer, the nodes The representation of not only depends on the representation of the previous layer It is also updated by aggregating information from its neighboring nodes. This aggregation operation ensures that the model can gradually integrate sentiment information from more contexts in each convolution layer, thereby improving the ability to model node relationships.

[0045] In order to simultaneously model the word pair relationships under the three emotional dimensions of positive, negative, and neutral, the model designs a three-channel parallel convolution structure. Each channel corresponds to a sentiment graph and performs graph convolution operations independently. In each channel, the model gradually aggregates sentiment information through multiple layers of convolution. The specific process is as follows: the positive sentiment channel is responsible for processing the positive sentiment graph G posIn this channel, the model aggregates the nodes and edge weights related to positive emotions through graph convolution operations and gradually updates the node representation. The final output of the convolution operation is a node representation matrix C′ containing positive emotion information. pos This matrix provides the basis for the positive part of the subsequent emotional triple; the operation of the negative emotional channel is similar to the positive emotional channel, but it processes the negative emotional graph G neg Through graph convolution, the model aggregates node information related to negative emotions and generates a node representation matrix C′ containing negative emotion information. neg ; Neutral emotion channel processes neutral emotion graph G neu , this channel generates the neutral emotion node representation matrix C′ by aggregating the node information of neutral emotion neu .

[0046] After completing the graph convolution operation on the three emotion channels, the model fuses the results of the three channels together through the splicing operation to generate the final node representation matrix C′. Specifically, the fusion operation is as follows:

[0047]

[0048] This fusion of multi-channel results ensures that the model can simultaneously consider the positive, negative, and neutral sentiment information in the sentence. By processing different sentiment maps in parallel, the model can better understand the multi-sentiment relationships in complex sentences and ensure that the information under each sentiment dimension is fully utilized.

[0049] Table decoding module

[0050] Table decoding is a key step in extracting sentiment triplets from the node representations generated by the three-channel graph convolutional network. This process is divided into two main parts: detection and classification. Through these two stages, the model is able to identify candidate sentiment regions from the node representations and classify them into sentiment, thereby generating the final sentiment triplet.

[0051] The main goal of table decoding is to extract the correct sentiment triplet t = (a, o, l) from the comprehensive node representation output by the three-channel graph convolutional network. In the detection stage, the task of the model is to identify candidate regions that may contain sentiment relations. In the two-dimensional table C, each cell represents a word pair in the sentence, and the model determines the candidate regions by identifying the boundaries in these word pairs. Specifically, in the detection stage, the model first uses a linear layer and a sigmoid function to predict whether each cell is the boundary of an emotional region from the node representation generated by the three-channel graph convolutional network. To this end, the model generates two score tables: the score table in the upper left corner and the score table in the upper left corner. and the score table in the lower right corner in represents the score of cell (i, j) as the upper left corner of the candidate region, S E Indicates the score of the lower right corner of the cell (i, j) as a candidate region. In order to select the vertices of the candidate region from the score table, the model uses a top-k strategy. This strategy first performs a top-k operation on the score table S. S and S E The scores in are sorted and the top k% positions with the highest scores are selected as candidate vertices. Specifically, the candidate upper left corner set S is S Select the cell positions corresponding to the top k% scores from S, and the candidate lower right corner set E is E The cell positions corresponding to the top k% scores are selected from . In this way, the model determines a set of candidate regions that may contain sentiment triplets. In the classification stage, the model performs sentiment classification on the candidate regions identified in the detection stage. For the detected candidate regions, the model first needs to construct a representation of the region. Each candidate region is determined by two vertices: the upper left vertex (a, b) and the lower right vertex (c, d). The model generates a representation R of the candidate region by splicing the representations of these vertices and the maximum pooling representation of the region:

[0052]

[0053] Among them, C a，b and C c，d The nodes representing the upper left and lower right corners of the candidate region, respectively, C a：c，b：d The set of all cells representing the candidate region. By performing maximum pooling on this set, the model can extract the most significant feature information in the region. Once the representation of the candidate region is generated, the model inputs the representation into a linear layer and classifies the sentiment of the candidate region through the softmax function. Classification results The values ​​of are {0, 1, 2, 3}, where 0 represents padding, meaning the region does not contain a valid sentiment triplet; 1 represents positive sentiment (Pos); 2 represents negative sentiment (Neg); and 3 represents neutral sentiment (Neu). Through this classification step, the model can assign a sentiment label to each candidate region, ultimately generating a sentiment triplet.

Claims

1. A method for extracting aspect sentiment triples based on a multi-channel graph convolutional network, characterized in that: The method It includes the following components: (1) Feature output module BERT and BiLSTM are used as dual encoders to capture the semantic relationship and local dependency between words, respectively, to improve the model's ability to understand text. Then, by introducing a prompt learning mechanism, the emotional prompt information is directly embedded into the sentence encoding, so that the model can adaptively learn emotional relationships without relying on external knowledge. First, the dual encoder is used to extract the representation of the sentence, and the encoding process is as follows: in, is the last hidden state output of the encoder, d represents the dimension of the hidden state, H is the representation of the input sentence X, and H T is the representation of template T; H T Contains 6 representations corresponding to the mask slots in the template, namely Represent aspect terms and opinion terms under positive, negative and neutral sentiments, respectively; After obtaining the joint representation of the sentence and the template, the model then calculates the correlation between each word and each slot in the template; the cued attention mechanism is introduced to calculate the attention score matrix (2) Multi-channel graph convolutional network module The Tri-Channel Graph Convolution Network is the core module in the architecture, which aims to capture the emotional relationship between word pairs in a sentence through graph convolution operations under different emotional dimensions. This module represents the nodes in the graph structure as the input of multi-layer convolution operations, and processes different emotional information separately through a parallel three-channel structure (positive, negative and neutral emotions), and finally generates a comprehensive emotional relationship representation for each word pair. (3) Table decoding module Table decoding is the key step for the model to extract sentiment triplets from the node representation generated by the three-channel graph convolutional network; the process is divided into two parts: detection and classification; through these two stages, the model identifies candidate sentiment regions from the node representation and classifies them sentimentally, thereby generating the final sentiment triplet.

2. According to the method for extracting aspect emotion triples based on a multi-channel graph convolutional network according to claim 1, it is characterized in that: Each row of the feature output module attention score matrix corresponds to the embedding of a template slot, and the column corresponds to each word in the input sentence; P=softmax(τWH T ) in, The six embedding representations in the representation template are: represents the embedding matrix of the sentence, is a learnable parameter matrix; through matrix multiplication, the model captures the correlation between each word in the sentence and each slot in the template, and normalizes the result through the softmax function to ensure that the value in the attention score matrix is ​​a probability distribution; each element P of the attention score matrix ij Represents the correlation between the i-th slot in the template and the j-th word in the sentence.

3. According to the method for extracting aspect emotion triples based on a multi-channel graph convolutional network according to claim 1, it is characterized in that: The graph convolutional network (GCN) of the multi-channel graph convolutional network module is a neural network commonly used for graph structured data, which updates the representation of each node by aggregating the information of each node and its neighboring nodes; In this module, the core idea of ​​the graph convolutional network is to gradually aggregate the sentiment information from neighboring nodes through multi-layer convolution operations, thereby enhancing the model's understanding of the relationship between nodes; given a graph G, the graph convolution operation is performed through the following formula: in, is c in layer l ij The expression, is a learnable weight matrix, σ is a nonlinear activation function (such as ReLU), Represents the connection node c ij and its neighbor node g k The edge weights between .

4. The method for extracting aspect emotion triples based on a multi-channel graph convolutional network according to claim 1, characterized in that: The multi-channel graph convolutional network module has a three-channel parallel convolution structure, each channel corresponds to a sentiment graph, and performs graph convolution operations independently. In each channel, the model gradually aggregates sentiment information through multiple layers of convolution.

5. According to the method for extracting aspect emotion triples based on a multi-channel graph convolutional network according to claim 4, it is characterized in that: Each channel model gradually aggregates emotional information through multi-layer convolution. The specific process is as follows: the positive emotional channel is responsible for processing the positive emotional graph G pos ; In this channel, the model aggregates the nodes and edge weights related to positive emotions through graph convolution operations, and gradually updates the node representation; the final output of the convolution operation is a node representation matrix C′ containing positive emotion information pos This matrix provides the basis for the positive sentiment part of the subsequent sentiment triplet; the operation of the negative sentiment channel is similar to the positive sentiment channel, but it processes the negative sentiment graph G neg ; Through graph convolution, the model aggregates node information related to negative emotions and generates a node representation matrix C' containing negative emotion information neg ; Neutral emotion channel processes neutral emotion graph G neu This channel aggregates the node information of neutral emotions and generates a neutral emotion node representation matrix C' neu ; After completing the graph convolution operation on the three emotion channels respectively, the model fuses the results of the three channels together through the splicing operation to generate the final node representation matrix C'. The fusion operation is as follows:

Images