Aspect-level sentiment classification method and system based on dependency-aware graph convolutional network

By constructing a graph of dependency marker edges on the dependency tree and using multi-layer dependency-aware graph convolution network layer to integrate dependency type information, the problem of lack of dependency type information processing in the existing technology is solved, and the aspect-level emotional classification is improved Accuracy and stability.

CN113987167BInactive Publication Date: 2024-11-22SHANDONG NORMAL UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111181805.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-10-11
Publication Date
2024-11-22
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The existing graph convolution network lacks integration of dependency type information in aspect-level emotion classification, resulting in the inability to effectively distinguish unimportant word relationships, limiting the accuracy and stability of emotional polarity discrimination.

Method used

By constructing a graph of dependency marker edges on the dependency tree, the integration of dependency type information is increased, and the convolution network layer uses a multi-layer dependency perception graph to process edges between words and words, distinguish important context information and relationships, and enhance emotional extremes Sexual judgment.

Benefits of technology

It improves the accuracy and stability of emotional polarity judgments on different data sets, and enhances the ability to distinguish emotional polarity from specific aspects.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113987167B_ABST
    Figure CN113987167B_ABST
Patent Text Reader

Abstract

The present disclosure provides an aspect-level sentiment classification method and system based on a dependency-aware graph convolutional network, including: obtaining text information: obtaining a sentiment classification result based on the obtained text information and a preset aspect-level sentiment classification model; wherein, in the aspect-level sentiment classification model, the context representation obtained based on the text information is input into a multi-head attention layer to realize the interaction between the context sequence and the aspect words, and the integration of dependency type information is added to the graph neural network, different attention is paid to the processing of the edges between words according to different dependency information, and comprehensive learning is performed using a multi-layer dependency-aware graph convolutional network layer; the present disclosure pays different attention to the processing of the edges between words according to different dependency information, distinguishes important context information from these relationships to enhance ASC, and improves the accuracy and stability of sentiment polarity judgment on different data sets.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present disclosure belongs to the technical field of sentiment classification, and in particular, relates to an aspect-level sentiment classification method and system based on a dependency-aware graph convolutional network. Background Art

[0002] With the rapid development of social media, a large number of short texts with subjective user emotions, short content and rich semantic information have emerged on the Internet. These massive data are a comprehensive presentation and important embodiment of user consciousness and opinions, affecting netizens' views, attitudes and judgments on things. Specifically, when netizens choose products on shopping platforms, they often refer to the comments provided by the product buyers before deciding whether to buy the product. How to accurately and efficiently use computer technology to automatically analyze sentiment information from massive short texts is of great significance for product analysis, topic monitoring, public opinion monitoring, user modeling and opinion analysis. Sentiment classification is the core of sentiment analysis technology, and its task is to judge the sentiment orientation of the text.

[0003] Aspect-level Sentiment Classification (ASC) belongs to text sentiment classification. ASC is different from sentence-level or document-level sentiment classification. It aims to perform more fine-grained task analysis at the sentence level to determine the sentiment polarity (positive, neutral, and negative) for specific aspects. For example, in a restaurant review sentence "The price is acceptable, although its service is not good", "price" and "service" are two aspects, and their sentiment polarities are positive and negative, respectively. That is to say, in the task of aspect-level sentiment classification, it is necessary to distinguish the sentiment polarity of specific aspects based on contextual information; many current methods use graph convolution networks (GCNs) to capture dependencies and discontinuous context words to discriminate the sentiment polarity of specific aspects.

[0004] The inventors of the present disclosure have discovered that in the method of using GCN to capture dependency relationships and discontinuous context words to discriminate sentiment polarity of specific aspects, the use of the GCN model is merely that GCN links grammatically related context words to the target aspect, lacks the information contained in the dependency type, treats all word relationships in the graph equally, is insufficient to handle unimportant relationships, and limits the discrimination of ASC. Summary of the invention

[0005] In order to solve the above problems, the present invention proposes an aspect-level sentiment classification method and system based on a dependency-aware graph convolutional network. The present invention marks edges and constructs a graph according to the dependency types corresponding to the dependency relationships on the dependency tree, increases the integration of dependency type information, pays different attention to the processing of edges between words according to different dependency information, distinguishes important contextual information from these relationships to enhance ASC, and improves the accuracy and stability of sentiment polarity judgment on different data sets.

[0006] In order to achieve the above object, the present invention is implemented through the following technical solutions:

[0007] In a first aspect, the present disclosure provides an aspect-level sentiment classification method based on a dependency-aware graph convolutional network, comprising:

[0008] Get text information:

[0009] Obtain sentiment classification results based on the acquired text information and the preset aspect-level sentiment classification model;

[0010] Among them, in the aspect-level sentiment classification model, the context representation obtained based on the text information is input into the multi-head attention layer to realize the interaction between the context sequence and the aspect words, and to increase the integration of dependency type information in the graph neural network. The processing of the edges between words is given different attention according to different dependency information, and a multi-layer dependency-aware graph convolutional network layer is used for comprehensive learning.

[0011] Furthermore, the training process of the aspect-level sentiment classification model is as follows:

[0012] Obtain the text to be analyzed, perform word embedding training on the text to be analyzed, and convert the text into a word embedding vector;

[0013] Convert the word embedding vector into a contextual representation;

[0014] Input the context representation into a multi-head attention layer to achieve interaction between context sequences and aspect words;

[0015] The context representation processed by the multi-head attention layer is input into the dependency-aware graph convolutional network for learning.

[0016] Furthermore, the learning of dependency-aware graph convolutional networks includes: inputting the context hidden state vector processed by the multi-head attention layer into the graph neural network, using the graph neural network to capture the grammatical relationship between aspect word nodes and their adjacent nodes, the node features are transmitted through the graph neural network layer, and the representation of each node is enriched by the grammatical information of the dependency tree; the integration of dependency type information is added to the graph neural network, different attention is given to the processing of edges between words according to different dependency information, and comprehensive learning is performed using multi-layer dependency-aware graph convolutional network layers.

[0017] Furthermore, the prediction of sentiment classification results includes:

[0018] The output of the dependency-aware graph convolutional network is fed into multiple conditional random fields fused with structured attention to capture opinion features and obtain a new context representation. The sentences of all conditional random fields are connected to obtain the output representation.

[0019] The output representation is sent to the sentiment classifier to obtain the final sentiment polarity prediction result.

[0020] Furthermore, after the output representation is transmitted to the sentiment classifier, it is first embedded in a fully connected layer and then a normalization layer to obtain the distribution of sentiment polarity and obtain the final sentiment polarity prediction result.

[0021] Furthermore, potential labels are added to indicate whether the context words affect the sentiment polarity of a specific aspect, and the marginal probability on the label is set to the influence of the current context words on the sentiment polarity of the aspect words; the marginal distribution of potential labels is calculated using the forward-backward algorithm to obtain the sentence representation.

[0022] Furthermore, when converting the word embedding vector into a contextual representation, a bidirectional long short-term memory network hidden state vector is used to obtain the contextual representation of the word embedding.

[0023] In a second aspect, the present disclosure further provides an aspect-level sentiment classification system based on a dependency-aware graph convolutional network, comprising a collection module and a classification module;

[0024] The acquisition module is configured to: obtain text information:

[0025] The classification module is configured to: obtain a sentiment classification result based on the acquired text information and a preset aspect-level sentiment classification model;

[0026] Among them, in the aspect-level sentiment classification model, the context representation obtained based on the text information is input into the multi-head attention layer to realize the interaction between the context sequence and the aspect words, and to increase the integration of dependency type information in the graph neural network. The processing of the edges between words is given different attention according to different dependency information, and a multi-layer dependency-aware graph convolutional network layer is used for comprehensive learning.

[0027] In a third aspect, the present disclosure further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the aspect-level sentiment classification method based on a dependency-aware graph convolutional network described in the first aspect.

[0028] In a fourth aspect, the present disclosure further provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of the aspect-level sentiment classification method based on the dependency-aware graph convolutional network described in the first aspect are implemented.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] 1. This disclosure solves the problem that the traditional GCN model omits information in dependency types and treats all word relationships in the graph equally, thus failing to distinguish unimportant relationships and misleading ASC accordingly;

[0031] 2 The present invention marks edges and constructs a graph according to the dependency types corresponding to the dependency relationships on the dependency tree, increases the integration of dependency type information, pays different attention to the processing of edges between words according to different dependency information, distinguishes important contextual information from these relationships to enhance ASC, and improves the accuracy and stability of sentiment polarity judgment on different data sets. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The drawings in the specification that constitute a part of this embodiment are used to provide a further understanding of this embodiment. The schematic embodiments of this embodiment and their descriptions are used to explain this embodiment and do not constitute improper limitations on this embodiment.

[0033] Figure 1 This is a flowchart of Embodiment 1 of the present disclosure;

[0034] Figure 2 This is a network model diagram of Example 1 of the present disclosure. DETAILED DESCRIPTION

[0035] The present disclosure is further described below in conjunction with the accompanying drawings and embodiments.

[0036] It should be noted that the following detailed descriptions are illustrative and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0037] Embodiment 1:

[0038] like Figure 1 As shown, this embodiment provides an aspect-level sentiment classification method based on a dependency-aware graph convolutional network, including:

[0039] Step 1: Obtain the text to be analyzed; perform word embedding training on the text to be analyzed, convert the text into a word embedding vector, embed the word to obtain the word vector expression of each word in the text. In this embodiment, two embedding methods are adopted: GloVe (Global Vectors for Word Representation) embedding and BERT (Bidirectional Encoder Representations from Transformers) embedding.

[0040] Step 1.1: GloVe embedding, given a context sequence consisting of m words and aspect sequence (1≤τ≤n≤m), where X a Yes X c Subsequence of . Through the pre-trained GloVe embedding matrix where d m represents the embedding dimension of the word vector, and V represents the size of the vocabulary.

[0041] Step 1.2: BERT embedding. In this embodiment, a pre-trained BERT model is used as an encoder to obtain the embedding of context words. First, in this embodiment, "[CLS]+context+[SEP]+term+[SEP]" is fed into BERT as input, thereby realizing the explicit interaction between the entire sentence and the aspect words. Then, the word vector is initialized after the input and fine-tuned during the training process.

[0042] The word embedding layer finally generates h l , and pass it to downstream tasks as a context representation.

[0043] Step 2: Convert the word vector embedding obtained in step 1 into a contextual representation. In this embodiment, a bidirectional long short-term memory (LSTM) network and multi-head attention are used to obtain a contextual representation in a specific direction.

[0044] Step 2.1: Input the context output of step 1 into the bidirectional LSTM layer; Applying the bidirectional LSTM to the context encoder can better capture the long-distance dependencies in the sentence; In this embodiment, the hidden representations in the forward direction and the backward direction are connected to obtain the context representation, which is defined as:

[0045]

[0046]

[0047]

[0048] in, and Represent the forward and backward hidden states respectively, t represents a certain moment, θ LSTM represents the parameters used in the bidirectional LSTM, Indicates connection, h L Represents the output context representation of the bidirectional LSTM.

[0049] Step 2.2: Input the context representation obtained in step 2.1 into the multi-head attention layer. The multi-head attention can perform multiple attention functions at the same time, which can realize the interaction between the context sequence and the aspect words, enhance the integration, and obtain the context information more accurately. It is defined as:

[0050]

[0051]

[0052] in represents the i-th head of the multi-head attention layer, Q, K, and W represent the query sequence, key sequence, and value sequence respectively, and W 0 , W Q , W K , W V is a learnable matrix, h MHA Represents the output of the multi-head attention layer, and obtains the final output of the context representation layer for a specific aspect.

[0053] Step 3: Input the context representation obtained by the multi-head attention layer in step 2.2 into the dependency-aware graph convolutional network module.

[0054] Step 3.1: Build a dependency-aware graph. First, get the dependency result from the obtained dependency tree. You can use the dependency tuple List to represent, express and The dependency type between them; secondly, using the adjacency matrix Represents the dependency relationship between context words in a tuple. If two context words and If there is a dependency relationship between them, then the corresponding position of the connection matrix is ​​1, otherwise it is 0; then use the dependency type matrix Denotes the dependency type corresponding to the dependency relationship; finally, in order to utilize the dependency type, a transfer matrix is ​​used to transform all Mapped to their embedded items

[0055] Step 3.2: Input the context representation obtained by the multi-head attention layer into the dependency-aware graph convolutional network module. In this embodiment, an L-layer dependency-based graph convolutional network is applied. In the L-layer, for any two context words and The edges between them use the hidden layer vectors in the L-1 layer and ( and Output from multi-head attention layer in aspect-specific context representation) and dependency-type embedding Connect it up.

[0056]

[0057]

[0058] Next, in this embodiment, the weights are calculated for the edges, and and Dimensions are aligned.

[0059]

[0060]

[0061] Finally, Applied to the edge and calculated at the lth layer after the L-layer dependency-based graph convolutional network process Output.

[0062]

[0063] Among them, W (l) and b (l) represents the learnable parameters in the L-th layer of the deep dependency-aware graph convolutional network, and σ represents the activation function. In this embodiment, the ReLU activation function is used to obtain the final output of the deep dependency-aware graph convolutional network layer. The above process is for each The whole L-layer deep dependency-aware graph convolutional network is used for this purpose, so the dependency type information embedded in the deep dependency-aware graph convolutional network layer can enhance the aspect-level sentiment analysis accordingly.

[0064] Step 4: Input the vector obtained by the dependency-aware graph convolutional network layer in step 3 into multiple conditional random fields (CRFs) attention fusion layers; in order to better capture the opinion features, in this embodiment, multiple linear chain CRFs are introduced to centrally add structural dependencies. In addition, in this embodiment, a potential label Y∈{Yes,No} is created to indicate whether each context word is an opinion word. Given a sentence representation X, the CRF is defined as:

[0065]

[0066]

[0067] Where score(y,x) represents a score from the sum of the emission score and the transfer score of the deep dependency-aware graph convolutional network layer. D represents the transfer matrix, Indicates z i to z i+1 The transfer score of the label, Represents label z i The emission score at the i-th position is the output of the deep dependency-aware graph convolutional network layer. get.

[0068] In this embodiment, the purpose of adding the potential label Y to the CRF layer is to indicate whether the context word affects the sentiment polarity of a specific aspect. The marginal probability on the Yes label is understood as the influence of the current context word on the sentiment polarity of the aspect word. Then the forward-backward algorithm is used to calculate the marginal distribution of the potential label, and finally the sentence representation S is obtained. c :

[0069]

[0070] The final representation is obtained by concatenating the sentence representations of all CRFs:

[0071]

[0072] Where m represents the number of CRFs.

[0073] Step 5: Feed the output obtained in step 4 into the output layer. The sentence representation C is passed to the sentiment classifier, which is first embedded in a fully connected layer and then a SoftMax normalization layer to obtain the distribution of sentiment polarity:

[0074] p(y|S c )=Softmax(W S +b s ) (15)

[0075] Where W S and b s are the learnable parameters of the sentiment classifier layer.

[0076] Step 6: The model in this example is trained by a standard gradient descent algorithm with cross entropy loss and L2 regularization:

[0077]

[0078] Where D represents the training data set, Y represents the true value label, Indicates the first elements. θ represents all parameters that need to be trained, and λ represents the coefficient of the L2 regularization term.

[0079] Embodiment 2:

[0080] This embodiment provides an aspect-level sentiment classification system based on a dependency-aware graph convolutional network, including a collection module and a classification module;

[0081] The acquisition module is configured to: obtain text information:

[0082] The classification module is configured to: obtain a sentiment classification result based on the acquired text information and a preset aspect-level sentiment classification model;

[0083] Among them, in the aspect-level sentiment classification model, the context representation obtained based on the text information is input into the multi-head attention layer to realize the interaction between the context sequence and the aspect words, and to increase the integration of dependency type information in the graph neural network. The processing of the edges between words is given different attention according to different dependency information, and a multi-layer dependency-aware graph convolutional network layer is used for comprehensive learning.

[0084] Embodiment 3:

[0085] This embodiment provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps of the aspect-level sentiment classification method based on the dependency-aware graph convolutional network described in Example 1 are implemented.

[0086] Embodiment 4:

[0087] This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps of the aspect-level sentiment classification method based on the dependency-aware graph convolutional network described in Example 1 are implemented.

[0088] The above description is only a preferred embodiment of the present embodiment and is not intended to limit the present embodiment. For those skilled in the art, the present embodiment may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present embodiment shall be included in the protection scope of the present embodiment.

Claims

1. An aspect-level sentiment classification method based on dependency-aware graph convolutional network, characterized in that: include: Get text information: Obtain sentiment classification results based on the acquired text information and the preset aspect-level sentiment classification model; In the aspect-level sentiment classification model, the context representation obtained based on the text information is input into the multi-head attention layer to realize the interaction between the context sequence and the aspect words, and the integration of dependency type information is added to the graph neural network. Different attention is paid to the processing of the edges between words according to different dependency information, and a multi-layer dependency-aware graph convolutional network layer is used for comprehensive learning. The specific steps involved are: Step 1: Get text information; Step 2: Convert the word vector embedding obtained in step 1 into a contextual representation, and use a bidirectional long short-term memory (LSTM) network and multi-head attention to obtain a contextual representation in a specific direction; Step 2.1: Feed the context output of step 1 into a bidirectional LSTM layer; apply a bidirectional LSTM to the context encoder to capture the long-range dependencies in the sentence; concatenate the hidden representations in the forward and backward directions to obtain the context representation, defined as: in, and Represent the forward and backward hidden states respectively, t represents a certain moment, θ LSTM represents the parameters used in the bidirectional LSTM, Indicates connection, h L Represents the output context representation of the bidirectional LSTM; Step 2.2: Input the context representation obtained in step 2.1 into the multi-head attention layer. The multi-head attention layer simultaneously performs the attention of multiple attention functions, realizes the interaction between the context sequence and the aspect words, enhances the integration, and obtains the context information; it is defined as: in represents the i-th head of the multi-head attention layer, Q, K, and W represent the query sequence, key sequence, and value sequence respectively, and W 0 , W Q , W K , W V is a learnable matrix, h MHA Represents the output of the multi-head attention layer, and obtains the final output of the context representation layer for a specific aspect; Step 3: Input the context representation obtained by the multi-head attention layer in step 2.2 into the dependency-aware graph convolutional network module; Step 3.1: Construct a dependency-aware graph; first, obtain the dependency results from the obtained dependency tree, using the dependency tuple List to represent, express and The dependency type between them; secondly, using the adjacency matrix Represents the dependency relationship between context words in a tuple. If two context words and If there is a dependency relationship between them, then the corresponding position of the connection matrix is ​​1, otherwise it is 0; then use the dependency type matrix Denotes the dependency type corresponding to the dependency relationship; finally, in order to utilize the dependency type, a transfer matrix is ​​used to transform all Mapped to their embedded items Step 3.2: Input the context representation obtained by the multi-head attention layer into the dependency-aware graph convolutional network module; apply L layers of dependency-based graph convolutional networks, and for any two context words in L layers and The edges between them use the hidden layer vectors in the L-1 layer and ( and Output from multi-head attention layer in aspect-specific context representation) and dependency-type embedding Connect Next, we calculate the weights for the edges and and Dimension alignment; Finally, Applied to the edge and calculated at the lth layer after the L-layer dependency-based graph convolutional network process Output: Among them, W (l) and b (l) represents the learnable parameters in the L-th layer of deep dependency-aware graph convolutional network, and σ represents the activation function. The ReLU activation function is used to obtain the final output of the deep dependency-aware graph convolutional network layer. The above process is for each The whole L-layer deep dependency-aware graph convolutional network is used for this purpose, so the dependency type information embedded in the deep dependency-aware graph convolutional network layer can enhance the aspect-level sentiment analysis accordingly.

2. The aspect-level sentiment classification method based on dependency-aware graph convolutional network according to claim 1, characterized in that: The training process of the aspect-level sentiment classification model is as follows: Obtain the text to be analyzed, perform word embedding training on the text to be analyzed, and convert the text into a word embedding vector; Convert the word embedding vector into a contextual representation; Input the context representation into a multi-head attention layer to achieve interaction between context sequences and aspect words; The context representation processed by the multi-head attention layer is input into the dependency-aware graph convolutional network for learning.

3. The aspect-level sentiment classification method based on dependency-aware graph convolutional network as claimed in claim 2, characterized in that: The learning of dependency-aware graph convolutional networks includes: inputting the context hidden state vector processed by the multi-head attention layer into the graph neural network, using the graph neural network to capture the grammatical relationship between aspect word nodes and their adjacent nodes, the node features are transmitted through the graph neural network layer, and the representation of each node is enriched by the grammatical information of the dependency tree; the integration of dependency type information is added to the graph neural network, different attention is given to the processing of edges between words according to different dependency information, and comprehensive learning is performed using multi-layer dependency-aware graph convolutional network layers.

4. The aspect-level sentiment classification method based on dependency-aware graph convolutional network as claimed in claim 2, characterized in that: The predictions of sentiment classification results include: The output of the dependency-aware graph convolutional network is fed into multiple conditional random fields fused with structured attention to capture opinion features and obtain a new context representation. The sentences of all conditional random fields are connected to obtain the output representation. The output representation is sent to the sentiment classifier to obtain the final sentiment polarity prediction result.

5. The aspect-level sentiment classification method based on dependency-aware graph convolutional network as claimed in claim 4, characterized in that: After the output representation is transmitted to the sentiment classifier, it is first embedded in a fully connected layer and then a normalization layer to obtain the distribution of sentiment polarity and obtain the final sentiment polarity prediction result.

6. The aspect-level sentiment classification method based on dependency-aware graph convolutional network as claimed in claim 4, characterized in that: Add potential tags to indicate whether the context words affect the sentiment polarity of a specific aspect, and set the marginal probability on the tag to the influence of the current context words on the sentiment polarity of the aspect words; use the forward-backward algorithm to calculate the marginal distribution of potential tags and obtain the sentence representation.

7. The aspect-level sentiment classification method based on dependency-aware graph convolutional network as claimed in claim 4, characterized in that: When converting the word embedding vector into a contextual representation, a bidirectional long short-term memory network hidden state vector is used to obtain the contextual representation of the word embedding.

8. Aspect-level sentiment classification system based on dependency-aware graph convolutional network, characterized by: Including acquisition module and classification module; The acquisition module is configured to: obtain text information: The classification module is configured to: obtain a sentiment classification result based on the acquired text information and a preset aspect-level sentiment classification model; In the aspect-level sentiment classification model, the context representation obtained based on the text information is input into the multi-head attention layer to realize the interaction between the context sequence and the aspect words, and the integration of dependency type information is added to the graph neural network. Different attention is paid to the processing of the edges between words according to different dependency information, and a multi-layer dependency-aware graph convolutional network layer is used for comprehensive learning. The specific steps involved are: Step 1: Get text information; Step 2: Convert the word vector embedding obtained in step 1 into a contextual representation, and use a bidirectional long short-term memory (LSTM) network and multi-head attention to obtain a contextual representation in a specific direction; Step 2.1: Feed the context output of step 1 into a bidirectional LSTM layer; apply a bidirectional LSTM to the context encoder to capture the long-range dependencies in the sentence; concatenate the hidden representations in the forward and backward directions to obtain the context representation, defined as: in, and Represent the forward and backward hidden states respectively, t represents a certain moment, θ LSTM represents the parameters used in the bidirectional LSTM, Indicates connection, h L Represents the output context representation of the bidirectional LSTM; Step 2.2: Input the context representation obtained in step 2.1 into the multi-head attention layer. The multi-head attention layer simultaneously performs the attention of multiple attention functions, realizes the interaction between the context sequence and the aspect words, enhances the integration, and obtains the context information; it is defined as: in represents the i-th head of the multi-head attention layer, Q, K, and W represent the query sequence, key sequence, and value sequence respectively, and W 0 , W Q , W K , W V is a learnable matrix, h MHA Represents the output of the multi-head attention layer, and obtains the final output of the context representation layer for a specific aspect; Step 3: Input the context representation obtained by the multi-head attention layer in step 2.2 into the dependency-aware graph convolutional network module; Step 3.1: Construct a dependency-aware graph; first, obtain the dependency results from the obtained dependency tree, using the dependency tuple List to represent, express and The dependency type between them; secondly, using the adjacency matrix Represents the dependency relationship between context words in a tuple. If two context words and If there is a dependency relationship between them, then the corresponding position of the connection matrix is ​​1, otherwise it is 0; then use the dependency type matrix Denotes the dependency type corresponding to the dependency relationship; finally, in order to utilize the dependency type, a transfer matrix is ​​used to transform all Mapped to their embedded items Step 3.2: Input the context representation obtained by the multi-head attention layer into the dependency-aware graph convolutional network module; apply L layers of dependency-based graph convolutional networks, and for any two context words in L layers and The edges between them use the hidden layer vectors in the L-1 layer and ( and Output from multi-head attention layer in aspect-specific context representation) and dependency-type embedding Connect Next, we calculate the weights for the edges and and Dimension alignment; Finally, Applied to the edge and calculated at the lth layer after the L-layer dependency-based graph convolutional network process Output: Among them, W (l) and b (l) represents the learnable parameters in the L-th layer of deep dependency-aware graph convolutional network, and σ represents the activation function. The ReLU activation function is used to obtain the final output of the deep dependency-aware graph convolutional network layer. The above process is for each The whole L-layer deep dependency-aware graph convolutional network is used for this purpose, so the dependency type information embedded in the deep dependency-aware graph convolutional network layer can enhance the aspect-level sentiment analysis accordingly.

9. A computer-readable storage medium having a computer program stored thereon for calculating fingerprint similarity, characterized in that: When the program is executed by a processor, the steps of the aspect-level sentiment classification method based on a dependency-aware graph convolutional network as described in any one of claims 1 to 7 are implemented.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the aspect-level sentiment classification method based on the dependency-aware graph convolutional network as described in any one of claims 1 to 7 are implemented.