A Method for Aspect Sentiment Triple Extraction Based on Multi-channel Graph Convolutional Networks

Abstract

A method for aspect sentiment triple extraction based on a multi-channel graph convolutional network is disclosed. This method utilizes a graph convolutional neural network architecture comprising three core modules: a feature output module, a multi-channel graph convolutional network module, and a table decoding module. This framework effectively leverages the advantages of BERT, BiLSTM, and cue learning mechanisms 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 handling complex sentiment information. This invention provides a core impetus for developing user review sentiment analysis systems, and this new technology has a profound impact on businesses' timely improvement of products and services and on individuals' optimized consumption decisions.

Description

Technical Field

[0001] This invention relates to a graph convolutional neural network architecture extraction method, and more particularly to an aspect sentiment triple extraction method based on a multi-channel graph convolutional network. Background Technology

[0002] Driven by the rapid development of technology in the digital age, the number of comments on social media platforms has surged, generating a large amount of text data containing user emotions and opinions, providing rich material for sentiment analysis. However, traditional holistic sentiment analysis methods (such as positive, neutral, and negative reviews) are no longer sufficient to meet the needs of analyzing complex data, and there is an urgent need for more granular text data sentiment analysis methods. Therefore, the Aspect Sentiment Triplet Extraction (ASTE) task has emerged.

[0003] ASTE aims to extract triples from given sentences to aid in more nuanced sentiment analysis. These triples consist of aspect, opinion, and sentiment. While significant progress has been made in this field in recent years, challenges remain, including accuracy in handling large datasets, insufficient mining of syntactic and semantic information, lack of interaction between relation table cells, and difficulty in fully capturing complex relationships involving multiple coexisting sentiments. Summary of the Invention

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

[0005] The objective of this invention is achieved through the following technical solution:

[0006] An aspect sentiment triple extraction method based on a multi-channel graph convolutional network, the method 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, improving the model's ability to understand text. Then, by introducing a cue learning mechanism, sentiment cue information is directly embedded into the sentence encoding, enabling the model to adaptively learn sentiment relationships without relying on external knowledge. First, the dual encoder is used to extract the sentence representation, and its encoding process is as follows:

[0009]

[0010] in, This is the output of the last hidden state of the encoder, where d represents the dimension of the hidden state, and H is the representation of the input sentence X. T This represents template T. H T It contains 6 representations corresponding to the mask slots in the template, namely These are terms and viewpoints representing 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; a cue 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 is the core module of this architecture, designed to capture the sentiment relationships between word pairs in a sentence through graph convolution operations under different sentiment dimensions. This module represents the nodes in the graph structure as inputs to multi-layer convolution operations, and processes different sentiment information through a parallel three-channel structure (positive, negative and neutral sentiment) to finally generate a comprehensive sentiment relationship representation for each word pair.

[0014] (3) Table Decoding Module

[0015] Table decoding is a key step in the model to extract sentiment triples from the node representations generated by a three-channel graph convolutional network. The process consists of two parts: detection and classification. Through these two stages, the model identifies candidate sentiment regions from the node representations and classifies them according to their sentiment, thereby generating the final sentiment triples.

[0016] The aspect sentiment triple extraction method based on a multi-channel graph convolutional network is described above, wherein each row of the attention score matrix of the feature output module corresponds to the embedding of a template slot, and the columns correspond to each word in the input sentence.

[0017] P = softmax(τWH) T )

[0018] in, This represents the six embedded representations in the template. The embedding matrix represents the sentence. It 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 results using a softmax function to ensure that the values ​​in the attention score matrix are a probability distribution; each element P of the attention score matrix ij This represents the correlation between the i-th slot in the template and the j-th word in the sentence.

[0019] The method for aspect sentiment triple extraction based on a multi-channel graph convolutional network (GCN) module utilizes a graph convolutional network (GCN), a type of neural network commonly used for graph-structured data. It updates node representations by aggregating information from each node and its neighbors. The core idea of ​​this module is to progressively aggregate sentiment information from neighboring nodes through multiple convolutional operations, thereby enhancing the model's understanding of relationships between nodes. Given a graph G, the graph convolutional operation is performed using the following formula:

[0020]

[0021] in, It is layer l in c ij The expression, It is a learnable weight matrix, and σ is a non-linear activation function (such as ReLU). Indicates connection node c ij and its neighboring node g k Edge weights between them.

[0022] The method for extracting aspect sentiment triples based on a multi-channel graph convolutional network describes a three-channel parallel convolutional network module. Each channel corresponds to a sentiment graph and performs graph convolution operations independently. In each channel, the model aggregates sentiment information step by step through multiple convolutions.

[0023] The method described above is an aspect sentiment triple extraction method based on multi-channel graph convolutional networks. Each channel model aggregates sentiment information step-by-step through multiple convolutional layers. Specifically, the positive sentiment channel is responsible for processing the positive sentiment graph G. pos In this channel, the model aggregates nodes and edge weights related to positive sentiment 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 sentiment information. pos This matrix provides the foundation for the positive sentiment component of the subsequent sentiment triplet; the operation of the negative sentiment channel is similar to that of the positive sentiment channel, but it deals with the negative sentiment graph G. neg Through graph convolution, the model aggregates node information related to negative sentiment, generating a node representation matrix C′ containing negative sentiment information. neg Neutral Emotional Channel Processing Neutral Emotional Map Gneu This channel generates a neutral sentiment node representation matrix C′ by aggregating node information of neutral sentiment. neu .

[0024] After performing graph convolution operations on the three emotion channels respectively, the model merges the results of the three channels together through a concatenation operation to generate the final node representation matrix C′. The fusion operation is as follows:

[0025]

[0026] The advantages and effects of this invention are:

[0027] 1. This invention proposes an innovative and efficient deep learning framework for extracting aspect-based sentiment triples. This framework effectively leverages the advantages of BERT, BiLSTM, and cue learning mechanisms for feature extraction. Through a parallel three-channel design, word pair relationships under positive, negative, and neutral sentiment are modeled separately, thereby 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 handling complex sentiment information.

[0028] 2. This invention pushes the boundaries of fine-grained sentiment analysis and provides a core impetus for the development of user review sentiment analysis systems. Furthermore, this new technology has a profound impact on businesses' timely improvement of products and services and on individuals' optimized consumption decisions. Attached Figure Description

[0029] Figure 1 This is a diagram of the overall framework of the model of this invention. Detailed Implementation

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

[0031] This invention discloses a method for extracting aspect sentiment triples based on a multi-channel graph convolutional network, which is structured as follows:

[0032] Feature output module

[0033] In natural language processing tasks, it is often necessary to extract multiple types of information from text, such as grammatical structure and semantic meaning. However, a single encoder can often only focus on one type of information, thus failing to fully utilize all the information in the text. Traditional sentiment analysis tasks often rely on direct modeling of specific words, which frequently ignores the complex interactions between different sentiments. To address these issues, this invention first employs BERT and BiLSTM 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 cue learning mechanism, sentiment cue information is directly embedded into the sentence encoding, enabling the model to adaptively learn sentiment relationships without relying on external knowledge. The sentence representation is first extracted using dual encoders, and the encoding process is as follows:

[0034]

[0035] in, This is the output of the last hidden state of the encoder, where d represents the dimension of the hidden state, and H is the representation of the input sentence X. T This represents template T. H T It contains 6 representations corresponding to the mask slots in the template, namely These are terms and viewpoints representing positive, negative, and neutral emotions, respectively.

[0036] After obtaining the joint representation of the sentence and the template, the model next calculates the correlation between each word and each slot in the template. To achieve this, this invention introduces a cue attention mechanism for calculating the attention score matrix. Each row of the 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, This represents the six embedded representations in the template. The embedding matrix represents the sentence. It is a learnable parameter matrix. Through matrix multiplication, the model can capture the relevance of each word in the sentence to each slot in the template, and normalize the results using a softmax function to ensure that the values ​​in the attention score matrix are a probability distribution. Each element P of the attention score matrix... ij This represents the correlation between the i-th slot in the template and the j-th word in the sentence.

[0039] The advantage of the cue attention scoring mechanism is that it can weight words separately under different emotional polarities, thereby achieving fine-grained perception of multiple coexisting emotions. Compared with other models that rely on external sentiment dictionaries or rules, cue attention scoring adaptively adjusts the weight of each word through embedded sentiment cues, thus exhibiting stronger generalization ability in complex contexts.

[0040] Multi-channel graph convolutional network module

[0041] The Tri-Channel Graph Convolution Network is the core module of this architecture, designed 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 inputs to multiple convolutional operations, processing 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.

[0042] Graph Convolutional Networks (GCNs) are a type of neural network commonly used for graph-structured data. They update node representations by aggregating information about each node and its neighbors. In this module, the core idea of ​​GCNs is to progressively aggregate sentiment information from neighboring nodes through multiple convolutional operations, thereby enhancing the model's understanding of relationships between nodes. Given a graph G, the graph convolution operation is performed using the following formula:

[0043]

[0044] in, It is layer l in c ij The expression, It is a learnable weight matrix, and σ is a non-linear activation function (such as ReLU). Indicates connection node c ij and its neighboring node g k The edge weights between nodes. In the l-th layer, nodes The representation depends not only on the representation of its higher layer. It also updates by aggregating information from its neighboring nodes. This aggregation operation ensures that the model can gradually integrate sentiment information from more context in each convolutional layer, thereby improving its ability to model node relationships.

[0045] To simultaneously model word pair relationships across three sentiment dimensions—positive, negative, and neutral—a three-channel parallel convolutional structure was designed. Each channel corresponds to a sentiment graph and performs graph convolution operations independently. Within each channel, the model progressively aggregates sentiment information through multiple convolutional layers. Specifically, the positive sentiment channel is responsible for processing the positive sentiment graph G. posIn this channel, the model aggregates nodes and edge weights related to positive sentiment through graph convolution operations, progressively updating the node representations. The final output of the convolution operation is a node representation matrix C′ containing positive sentiment information. pos This matrix provides the foundation for the positive emotional component of the subsequent emotional triplet; the operation of the negative emotional channel is similar to that of the positive emotional channel, but it deals with the negative emotional graph G. neg By using graph convolution, the model aggregates node information related to negative sentiment, generating a node representation matrix C′ containing negative sentiment information. neg Neutral Emotional Channel Processing Neutral Emotional Map G neu This channel generates a neutral sentiment node representation matrix C′ by aggregating node information of neutral sentiment. neu .

[0046] After performing graph convolution operations on the three emotion channels separately, the model merges the results of the three channels together through a concatenation 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 positive, negative, and neutral sentiment information within a sentence. By processing different sentiment maps in parallel, the model can better understand the multi-sentiment relationships in complex sentences and ensure that information from each sentiment dimension is fully utilized.

[0049] Table Decoding Module

[0050] Table decoding is a crucial step in the model's extraction of sentiment triples from node representations generated by a three-channel graph convolutional network. This process mainly consists of two parts: detection and classification. Through these two stages, the model can identify candidate sentiment regions from the node representations and classify them according to their sentiment, thereby generating the final sentiment triples.

[0051] The primary goal of table decoding is to extract the correct sentiment triples t = (a, o, l) from the comprehensive node representation output by the three-channel graph convolutional network. In the detection phase, the model's task is to identify candidate regions that may contain sentiment relationships. In a two-dimensional table C, each cell represents a word pair in a sentence, and the model determines candidate regions by identifying the boundaries within these word pairs. Specifically, in the detection phase, the model first uses a linear layer and a sigmoid function to predict whether each cell represents the boundary of a sentiment region from the node representation generated by the three-channel graph convolutional network. To this end, the model generates two score tables: the top-left score table... and the score table in the bottom right corner in S represents the score of cell (i, j) as the top left corner of the candidate region. E Let S represent the score of cell (i, j) as the bottom right corner of a candidate region. To select the vertices of candidate regions from the score table, the model uses a top-k strategy. This strategy first processes the score table S... S and S E The scores are sorted, and the top k% positions with the highest scores are selected as candidate vertices. Specifically, the candidate top-left corner set S is formed from S... S Select the cell positions corresponding to the top k% of scores, and the candidate bottom right corner set E is from S. E The model selects the cell positions corresponding to the top k% of scores. Using this method, the model identifies a set of candidate regions that may contain sentiment triples. In the classification phase, the model performs sentiment classification on the candidate regions identified in the detection phase. For each detected candidate region, the model first needs to construct a representation of that region. Each candidate region is determined by two vertices: the top-left vertex (a, b) and the bottom-right vertex (c, d). The model generates the representation R of the candidate region by concatenating the representations of these vertices with the max-pooled representation of the region.

[0052]

[0053] Among them, C a，b and C c，d C represents the nodes at the top left and bottom right corners of the candidate region, respectively. a：c，b：d The set represents all cells of a candidate region. By performing max pooling on this set, the model can extract the most salient features within that region. Once the candidate region representation is generated, the model feeds this representation into a linear layer and uses a softmax function to classify the sentiment of the candidate regions. Classification results. The value of takes the form of {0, 1, 2, 3}, where 0 represents Padding, meaning the region does not contain a valid sentiment triple; 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 sentiment triples.

Claims

1. A method for aspect sentiment triple extraction based on multi-channel graph convolutional networks, characterized in that, The method It includes the following components: (1) Feature output module BERT and BiLSTM are used as dual encoders to capture semantic relationships and local dependencies between words, respectively, improving the model's ability to understand text. Then, by introducing a cue learning mechanism, sentiment cue information is directly embedded into the sentence encoding, enabling the model to adaptively learn sentiment relationships without relying on external knowledge. First, the dual encoder is used to extract the sentence representation, and its encoding process is as follows: in, This is the output of the last hidden state of the encoder, where d represents the dimension of the hidden state, and H is the representation of the input sentence X. T This represents template T; H T It contains 6 representations corresponding to the mask slots in the template, namely These are terms and viewpoints representing positive, negative, and neutral emotions, 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; a cue 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 of the architecture, designed to capture the sentiment relationships between word pairs in a sentence through graph convolution operations under different sentiment dimensions. This module represents the nodes in the graph structure as inputs to multi-layer convolution operations. Through a parallel three-channel structure, including positive, negative and neutral sentiment, it processes different sentiment information respectively, and finally generates a comprehensive sentiment relationship representation for each word pair. (3) Table Decoding Module Table decoding is a key step in the model to extract sentiment triples from the node representations generated by a three-channel graph convolutional network. The process consists of two parts: detection and classification. Through these two stages, the model identifies candidate sentiment regions from the node representations and classifies them according to their sentiment, thereby generating the final sentiment triples.

2. The aspect sentiment triple extraction method based on a multi-channel graph convolutional network according to claim 1, characterized in that, Each row of the attention score matrix of the feature output module corresponds to the embedding of a template slot, and the columns correspond to each word in the input sentence; P=softmax(τWH T ) in, This represents the six embedded representations in the template. The embedding matrix represents the sentence. It 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 results using a softmax function to ensure that the values ​​in the attention score matrix are a probability distribution; each element P of the attention score matrix ij This represents the correlation between the i-th slot in the template and the j-th word in the sentence.

3. The aspect sentiment triple extraction method based on a multi-channel graph convolutional network according to claim 1, characterized in that, The multi-channel graph convolutional network module's graph convolutional network (GCN) is a neural network commonly used for graph-structured data. It updates the node representation by aggregating information about each node and its neighboring nodes. In this module, the core idea of ​​graph convolutional networks is to gradually aggregate sentiment information from neighboring nodes through multiple layers of convolutional operations, thereby enhancing the model's understanding of the relationships between nodes. Given a graph G, the graph convolutional operation is performed using the following formula: in, It is layer l in c ij The expression, It is a learnable weight matrix, and σ is a non-linear activation function. Indicates connection node c ij and its neighboring node g k Edge weights between them.

4. The aspect sentiment triple extraction method 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 convolutional structure, with each channel corresponding to a sentiment graph and performing graph convolution operations independently. In each channel, the model gradually aggregates sentiment information through multiple layers of convolution.

5. The aspect sentiment triple extraction method based on a multi-channel graph convolutional network according to claim 4, characterized in that, Each channel model aggregates sentiment information step by step through multiple convolutions. Specifically, the positive sentiment channel is responsible for processing the positive sentiment map G. pos In this channel, the model aggregates nodes and edge weights related to positive sentiment through graph convolution operations, progressively updating the node representations; the final output of the convolution operation is a node representation matrix C' containing positive sentiment information. pos This matrix provides the foundation for the positive sentiment component of the subsequent sentiment triplet; the operation of the negative sentiment channel is similar to that of the positive sentiment channel, but it deals with the negative sentiment graph G. neg Through graph convolution, the model aggregates node information related to negative sentiment, generating a node representation matrix C' containing negative sentiment information. neg Neutral Emotional Channel Processing Neutral Emotional Map G neu This channel generates a neutral sentiment node representation matrix C' by aggregating node information of neutral sentiment. neu ; After performing graph convolution operations on the three emotion channels respectively, the model merges the results of the three channels through a concatenation operation to generate the final node representation matrix C'. The fusion operation is as follows:

Images