Aspect-level sentiment analysis method based on syntactic guidance Transform

Abstract

The invention relates to an aspect-level sentiment analysis method based on syntactic guidance Transform, which comprises the following steps of: firstly, constructing a dependency relationship matrix B through dependency analysis, then, calculating a syntactic distance through a syntactic dependency tree, modeling node importance by utilizing an inverse distance weighting function, and then, fusing the information into a self-attention mechanism of the Transform; secondly, aspect word extraction is conducted through a conditional random field, a local-global representation fusion layer is designed, local aspect word representation and global sentence representation are combined, and the semantic modeling capacity is enhanced; and inputting the data into a Softmax classifier, and outputting an emotion polarity tag. According to the method, the advantages of syntactic information and a self-attention mechanism are combined, the dependency relationship between aspect words and sentiment words can be more effectively modeled, the sentiment polarity classification accuracy is improved, and the context modeling capability of the model is enhanced. A multi-task learning framework is adopted, aspect word extraction and sentiment polarity classification are carried out in a combined mode, the generalization ability of the model is improved, and aspect-level sentiment analysis can be carried out more accurately.

Description

Technical Field

[0001] The present invention relates to the fields of natural language processing and deep learning, and particularly relates to an aspect-level sentiment analysis method based on syntactic-guided Transformer. This method combines syntactic information with the self-attention mechanism to achieve the extraction of aspect words in the text and the prediction of sentiment polarity. Background Art

[0002] The statements in this part only provide background technical information related to the present disclosure, and these statements may constitute the prior art. In the process of implementing the present invention, the inventors found that there are at least the following problems in the prior art.

[0003] Currently, natural language processing (NLP) technology has made remarkable progress in tasks such as text classification, machine translation, and sentiment analysis. Especially driven by deep learning, text processing methods based on neural networks have become the mainstream. Aspect-based sentiment analysis (ABSA), as a fine-grained task of sentiment analysis, aims to identify aspect words in the text and judge their corresponding sentiment polarities, and has important value in application scenarios such as e-commerce review analysis and social media public opinion monitoring.

[0004] For example, in the sentence "The food in this restaurant is very delicious, but the service is too slow.", the ABSA task includes two subtasks:

[0005] 1. Aspect term extraction (ATE): Identify aspect words in the sentence, such as "food" and "service".

[0006] 2. Sentiment polarity classification (SPC): Judge the sentiment of each aspect word, such as "food" has a positive sentiment and "service" has a negative sentiment.

[0007] Compared with traditional sentiment analysis methods (such as overall sentiment classification), ABSA can provide a more fine-grained analysis, making sentiment judgment more accurate. However, existing methods still face many challenges when dealing with the ABSA task.

[0008] Traditional ABSA methods rely on rule-based natural language processing techniques or machine learning models, by constructing sentiment lexicons, manually setting features, or using support vector machines (SVM), naive Bayes Classification algorithms such as Bayes are used to determine sentiment polarity. Although these methods have strong interpretability, their ability to handle complex syntactic structures is limited, and they perform poorly in cross-domain applications. With the development of deep learning, models such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs) have been introduced into the ABSA task, improving the classification performance of the model by automatically learning text features. However, these methods mainly rely on word order information and have limited ability to model long-distance dependencies, making it difficult to accurately capture the association between aspect words and sentiment words.

[0009] In recent years, to more effectively model long-distance dependencies in sentences, ABSA methods based on graph neural networks (GNNs) have gradually emerged. These methods use dependency parsing to construct syntactic dependency trees and extract features from the syntactic graph through models such as graph convolutional networks (GCNs) and graph attention networks (GATs) to enhance the association between aspect words and sentiment words. However, GNN-based methods highly rely on the quality of syntactic parsing and are easily affected by parsing errors. In addition, GNNs are prone to the phenomenon of over-smoothing when stacked in multiple layers, resulting in a decline in information expression ability and limiting their adaptability to complex situations.

[0010] In recent years, the Transformer model has demonstrated excellent performance in natural language processing tasks with its self-attention mechanism. The Transformer can effectively capture long-distance dependencies and has achieved remarkable results in multiple tasks such as machine translation, text summarization, and sentiment analysis. However, after testing, in the ABSA task, existing Transformer models are still not satisfactory and there is room for improvement. This is the case with the patent application No. 202411248130.7, titled "A Long-Text Sentiment Analysis Method Based on Modern Chinese Grammar in the Economic Security Field". This invention attempts to use a proprietary dictionary in the economic security field to replace traditional single-character word segmentation and uses relative position encoding to replace the absolute position encoding of traditional Transformers, allowing the model to learn and adjust the parameters in the network by itself.

[0011] However, according to the applicant's in-depth research, the reason why the above similar inventions are not satisfactory in actual use is mainly that they ignore the utilization of syntactic information and only rely on the serialized self-attention mechanism, which makes the performance of the model in the aspect-level sentiment analysis task still have room for improvement. In addition, existing research usually extracts aspect words or classifies sentiment polarity separately and fails to fully utilize the complementary information between the two, resulting in insufficient generalization ability of the model. Summary of the Invention

[0012] In view of the above problems, the present invention aims to solve some of the problems in the prior art, or at least alleviate these problems.

[0013] A method for aspect-level sentiment analysis based on syntax-guided Transformer, including the following steps:

[0014] Dependency parsing: select the input sentence S, use the dependency parsing tool to parse the syntactic structure, and construct the dependency matrix B∈R n×n ; Where n is the number of words in the sentence, and the matrix element B ij Indicates whether there is a dependency relationship between word i and word j, if yes, it is 1, otherwise it is 0;

[0015] Text encoding: Encode the input sentence D using a pre-trained language model to extract the embedding matrix H0∈R n ×d According to the dependency matrix B, the inverse distance weighting function is used and normalized to obtain the normalized weighted matrix M′∈Rn ×n ; Finally, we get the word vector enhanced representation matrix G embed =M′·H0; where d is the hidden layer dimension;

[0016] Syntax guides Transformer module processing: The dependency matrix B is used to calculate the Transformer's attention mechanism. The complete calculation formula of Transformer is as follows:

[0017]

[0018] Where Q, K, V are query, key, and value matrices, and their values are all represented by the word vector enhancement matrix H embed Linear transformation, β is the scaling factor, H out is the output of Transformer, T is the matrix transpose;

[0019] Aspect word extraction: Enhance the representation matrix H of word vectors through conditional random fields (CRF) embed Perform label space projection to generate predicted aspect label sequence;

[0020] Representation fusion and prediction of sentiment polarity: H out Perform masking and pooling to generate local aspect word representation H Pooling ∈R d , R represents the output belongs to the real vector, and according to the word vector enhancement matrix H embed and H Pooling Perform interactive attention to calculate the global sentence representation H Attn ∈R d ; Represent local aspect words HPooling and the global sentence representation H Attn Fuse and pass it to the Softmax classifier, and cooperate with the aspect label sequence to output the sentiment polarity label.

[0021] According to the dependency matrix B, cooperate with the inverse distance weighting function and normalize it to obtain the normalized weighted matrix M' ∈ R n×n , including the following steps:

[0022] Calculate the shortest path matrix D between each pair of words in the sentence according to the dependency matrix B;

[0023] Apply the inverse distance weighting function to the shortest path matrix D for calculation to obtain the inverse distance weighted matrix M;

[0024] Perform Softmax normalization on the inverse distance weighted matrix M by row to obtain the normalized weighted matrix M' ∈ R n ×n ; The calculation formula is:

[0025]

[0026] M' = Softmax(M)

[0027] where D ij represents the distance between word i and word j on the syntactic tree.

[0028] Use the dependency matrix B for the calculation of the attention mechanism of the Transformer, including the following steps:

[0029] Perform linear projections on the word vector enhanced representation matrix H embed respectively to obtain 3 matrices: the query matrix the key matrix and the value matrix where d k is the dimension of the attention head;

[0030] Introduce the dependency matrix B as additional information into the attention calculation, and the complete calculation formula is as follows:

[0031]

[0032] Perform masking operation and pooling on H out to generate the local aspect word representation H Pooling , including the following steps:

[0033] Perform a masking operation on the output H of the Transformer out to mask the non-aspect words, so that the representation of the aspect words is retained, and the representation of the non-aspect words is set to zero;

[0034] Pool the non-zero parts of the aspect words to calculate the representation of local aspect words:

[0035]

[0036] Among them, A represents the index set of aspect words, and |A| is the number of aspect words.

[0037] Fuse the local aspect word representation H Pooling and the global sentence representation H Attn and pass them to the Softmax classifier, and cooperate with the aspect label sequence to output sentiment polarity labels, including the following steps:

[0038] Perform interactive attention calculation on H Pooling and H embed to obtain the global sentence representation H Attn ;

[0039] Fuse the local and global representations through a concatenation operation to form the final classification input feature:

[0040] H Fusion = [H Pooling ; H Attn

[0041] Pass the final classification input feature H Fusion to the Softmax classifier, and cooperate with the aspect label sequence to output sentiment polarity labels; the sentiment polarity labels include positive, negative or neutral.

[0042] The aspect-level sentiment analysis method based on syntactic-guided Transformer further includes training and optimization; during the training process, the standard cross-entropy loss function is used to optimize the sentiment polarity, and at the same time, the loss function of conditional random field is used to optimize the aspect word extraction.

[0043] Furthermore, the optimization adopts the Adam (Adaptive Moment Estimation) optimizer, and the learning rate is initially set to 10 -5 and is dynamically adjusted according to the performance of the validation set during the training process; the training method uses an end-to-end training method to jointly optimize the aspect word extraction and sentiment polarity classification tasks through a multi-task learning framework.

[0044] A syntactic-guided Transformer framework for aspect-oriented sentiment analysis includes a dependency parsing module, a text encoding module, a syntactic-guided Transformer module, a fusion module, and an aspect word extraction module, which are used to implement the steps of the above-mentioned aspect-level sentiment analysis method based on syntactic-guided Transformer.

[0045] ​The present invention has the following beneficial effects:

[0046] 1. By syntactic-guided Transformer to enhance long-distance dependency modeling, the model can more accurately capture the correlation between aspect words and sentiment words, solving the problem that traditional methods are difficult to model long-distance dependencies.

[0047] 2. Through the fusion of local and global information, combining local aspect information and global context information at the same time, the accuracy of aspect-level sentiment analysis is improved, enabling the model to still maintain a high sentiment polarity prediction ability under complex syntactic structures.

[0048] 3. Through a multi-task learning framework, aspect word extraction and sentiment polarity classification are simultaneously performed under a unified framework, enhancing the generalization ability of the model on different tasks and making it applicable to multiple application scenarios, such as e-commerce review analysis, social media sentiment monitoring, brand public opinion analysis, etc.

[0049] The present invention proposes an aspect-level sentiment analysis method based on syntactic-guided Transformer, which combines the advantages of syntactic information and self-attention mechanism. By constructing an inverse distance weighting function, this method can more effectively model the dependencies between aspect words and sentiment words, improving the accuracy of sentiment polarity classification. At the same time, a local-global representation fusion layer is designed to combine the aspect-based local representation with the global sentence representation, enhancing the context modeling ability of the model. In addition, this method also adopts a multi-task learning framework to jointly perform aspect word extraction and sentiment polarity classification, improving the generalization ability of the model and enabling it to perform aspect-level sentiment analysis more accurately. Brief Description of the Drawings

[0050] Figure 1 is a general flowchart of the present invention;

[0051] Figure 2 is an example diagram of the model of the present invention. Detailed Embodiments

[0052] The following further describes the present invention in conjunction with the accompanying drawings. The embodiments of the present invention are only used to illustrate the present invention and do not limit the present invention. Without departing from the technical idea of the present invention, various substitutions and changes made according to ordinary technical knowledge and common means in the art shall be included within the scope of the present invention.

[0053] To address the limitations of existing technologies in long-distance dependence and complex syntactic structure modeling, the present invention provides an aspect-level sentiment analysis method based on syntactic-guided Transformer, which combines syntactic information and the Transformer self-attention mechanism, enhances the guiding effect of syntactic information through an inverse distance weighting function, and designs a local-global representation fusion layer to improve the accuracy of sentiment analysis. Specifically, during implementation, the system first requires a trained model file and sentence input, such as Figure 1 As shown in or 2, the sentence will be processed through the following main steps:

[0054] Step a: Dependency parsing and syntactic matrix construction

[0055] The input sentence is first parsed by a dependency parsing tool (such as SpaCy, a high-performance natural language processing toolkit that supports tasks such as text tokenization, part-of-speech tagging, named entity recognition, and dependency syntactic analysis for multiple languages and is widely used for extracting syntactic structure information) to generate a syntactic tree, and the dependency relationship between each pair of words is calculated based on the tree structure. Then, a dependency relationship matrix B∈R n×n is constructed using the dependency tree. Each element B ij in the matrix represents the existing dependency relationship between word i and word j, which is 1 if it exists and 0 otherwise. This matrix will serve as the basis for subsequent calculations to guide the attention mechanism of the Transformer model.

[0056] Step b: Text encoding and generation of embedding representations

[0057] The input sentence enters the text encoding module, and a pre-trained language model (such as BERT) is used to encode the sentence to extract the context-related embedding representation of each word. Assuming the input sentence is S = {s1, s2, …, s n}, after being processed by BERT, the text encoding module outputs an embedding matrix H0∈R n×d , where d is the hidden layer dimension (i.e., the hidden dimension of the BERT model), representing the features of each word in the context. Subsequently, the shortest dependency path between each pair of words in the syntactic tree is calculated according to the dependency relationship matrix B in step a, and a normalized inverse distance weighting matrix M′∈R n×n is generated to enhance the word vector representation as H embed . Finally, the syntactic-enhanced embedding representation (i.e., the enhanced word vector representation) is calculated:

[0058] H embed = M′·H0

[0059] where, to calculate the shortest dependency path, the shortest path matrix D between each node in the syntactic tree is calculated through a short path algorithm, where D ijRepresents the distance between word i and word j on the syntactic tree. Subsequently, an inverse distance weighting function is used to generate an inverse distance weighted matrix M, which is then Softmax-normalized row by row to obtain a normalized inverse distance weighted matrix M′. The calculation formula is as follows:

[0060]

[0061] M′ = Softmax(M)

[0062] Step c: Syntactic-guided Transformer module

[0063] In the self-attention calculation of Transformer, syntactic information is incorporated into the calculation process through the dependency matrix B, and thus the dependency matrix B calculated in step a is used as the structural guidance information in the attention mechanism.

[0064] Specifically, in the calculation process of Transformer, first, the input word vector enhanced representation matrix H embed is linearly projected to obtain a query matrix Q, a key matrix K, and a value matrix V. Subsequently, in the self-attention calculation process, B is introduced as additional information into the attention calculation, and the attention weights are calculated and weighted summed using the following formula:

[0065]

[0066] where β is the scaling factor, and H out is the output of Transformer guided by syntactic information. The finally generated H out combines the feature representations enhanced by syntactic information and can effectively enhance the modeling ability of syntactic information on the relationship between words.

[0067] Step d: Aspect word extraction

[0068] The aspect words are extracted through a conditional random field (CRF), which ensures the coherence of aspect word labels. Using the CRF, the word vector enhanced representation matrix H embed generated in step b is projected onto the label space to generate a predicted aspect label sequence for predicting the positions of aspect words. In the final output result, the aspect word part is marked as 1, and the non-aspect word part is marked as 0 to ensure the coherence of the labels.

[0069] Step e: Local-global representation fusion

[0070] To improve the accuracy of sentiment analysis, the present invention uses a local-global representation fusion mechanism. First, through a masking operation, the non-aspect word part in the H out output in step c is masked, and only the features of aspect words are retained, making H outSet the non - aspect - related part of the words to 0. Subsequently, perform a pooling operation on the non - zero part of the aspect - related words to generate a local aspect - related word representation H Pooling .

[0071] The generation of the local aspect - related word representation is as follows:

[0072]

[0073] where A is the index set of the aspect - related words, and |A| is the number of aspect - related words.

[0074] Then, perform an interactive attention calculation on H embed and H embed from step b to obtain the global sentence representation H Attn .

[0075] Finally, concatenate and fuse the local aspect - related word representation H Pooling and the global sentence representation H Attn to form the final input representation for sentiment classification:

[0076] H Fusion = [H Pooling ; H Attn

[0077] This fused representation is passed to the subsequent Softmax classifier, along with the aspect label sequence, to output sentiment polarity labels (such as positive, negative, or neutral), achieving accurate classification of sentiment polarity. Through the local - global representation fusion mechanism, the ability to model aspect - level sentiment features is enhanced.

[0078] Step f: Training and optimization

[0079] The method of the present invention uses an end - to - end training method to jointly optimize the aspect - related word extraction and sentiment polarity classification tasks through a multi - task learning framework. During the training process, the sentiment polarity is optimized using the standard cross - entropy loss function, and the aspect - related word extraction is optimized using the loss function of the conditional random field. The optimization process uses the Adam optimizer, and the initial learning rate is set to 10 -5 , and it is dynamically adjusted according to the performance of the validation set during the training process. The Adam (Adaptive Moment Estimation) optimizer combines the ideas of momentum gradient and adaptive learning rate, and can dynamically adjust the learning step size of each parameter according to the estimates of the first - order and second - order moments during the training process, thereby improving the convergence speed and stability.

[0080] Specific example:

[0081] Taking the sentence "The food in the restaurant is delicious, but the service is slow" as an example, the specific implementation steps of the present invention are as follows:

[0082] ​1. Dependency parsing: The dependency parsing tool parses the sentence structure, identifies the syntactic relationships between words, and constructs a dependency relationship matrix B.

[0083] 2. Text encoding: Generate word embedding representations through the BERT model, and combine with the syntactic matrix, and apply the inverse distance weighted function to calculate the enhanced representation matrix H of word vectors embed 。

[0084] 3. Syntax-guided Transformer module: Adjust the attention calculation of the Transformer based on the dependency relationship matrix B to make the dependency relationship between the aspect word and the sentiment word clearer.

[0085] 4. Local-global information fusion: Extract the local information of the aspect word, and combine with the global sentence representation to form the final classification input feature H Fusion 。

[0086] 5. Aspect word extraction: Use CRF to extract the aspect words "food" and "service".

[0087] 6. Sentiment classification: Based on the fused representation, the Softmax classifier outputs that the sentiment polarity of the aspect word "food" is "positive", and the sentiment polarity of "service" is "negative".

[0088] In view of the deficiencies of the existing aspect-level sentiment analysis methods in long-distance dependency modeling, syntactic information utilization, local and global representation fusion, etc., the present invention proposes an aspect-level sentiment analysis method based on syntax-guided Transformer. This method combines syntactic information with the self-attention mechanism, optimizes the attention calculation of the Transformer through the inverse distance weighted function, and designs a local and global information fusion layer to improve the modeling ability of the aspect-level sentiment analysis task. The present invention can perform aspect word extraction and sentiment polarity classification simultaneously, improve the model's ability to model the association between aspect words and sentiment polarities, and enhance the generalization ability across tasks.

[0089] A syntax-guided Transformer framework for aspect-oriented sentiment analysis, including a dependency parsing module, a text encoding module, a syntax-guided Transformer module, a fusion module, and an aspect word extraction module, is used to implement the steps of the above-mentioned aspect-level sentiment analysis method based on syntax-guided Transformer. Among them, the dependency parsing module (i.e., the dependency parser) generates syntactic dependency relationships, the text encoding module (i.e., the text encoder) generates context representations, the syntax-guided Transformer module uses syntactic information to optimize the attention mechanism, the aspect word extraction module is used to extract aspect words, and the fusion module uses the fusion mechanism to integrate local and global information.

[0090] The present invention is applied to aspect-level sentiment analysis, and can extract aspect terms and predict the sentiment polarity of aspect terms. It uses syntactic information to guide the attention calculation of the Transformer model, combines local and global information fusion, supports both aspect term extraction and sentiment classification, can achieve high-precision sentiment analysis in multiple real scenarios, and has broad application value.

Claims

1. An aspect-level sentiment analysis method based on syntactic-guided Transformer, characterized in that, Including the following steps: Dependency parsing: Select the input sentence S, use a dependency parsing tool to parse the syntactic structure, and construct a dependency relationship matrix B ∈ R n ×n ; where n is the number of words in the sentence, and the matrix element B ij indicates whether there is a dependency relationship between word i and word j. If there is, it is 1; otherwise, it is 0; Text encoding: Use a pre-trained language model to encode the input sentence S to extract the embedding matrix H0 ∈ R n×d ; According to the dependency matrix B, cooperate with the inverse distance weighting function and normalize it to obtain the normalized weighted matrix M' ∈ R n ×n ; Finally, obtain the word vector enhanced representation matrix H embed = M' · H0; where d is the hidden layer dimension; Syntactic-guided Transformer module processing: Using the dependency matrix B for the calculation of the attention mechanism of the Transformer. The complete calculation formula of the Transformer is as follows: where Q, K, and V are query, key, and value matrices, and their values are all enhanced and represented by the word vector matrix H embed through linear transformation, β is a scaling factor, and H out is the output of the Transformer, and T is the matrix transpose; Aspect term extraction: Enhance the representation matrix H of word vectors through Conditional Random Fields (CRF). embed Perform label space projection to generate a predicted aspect label sequence. Indicate the fused and predicted sentiment polarity: For H out perform a masking operation and pooling to generate the local aspect word representation H Pooling ∈ R d , where R represents that the output belongs to a real number vector, and enhance the representation matrix H embed and H Pooling perform interactive attention calculation to obtain the global sentence representation H Attn ∈ R d ; fuse the local aspect word representation H Pooling and the global sentence representation H Attn and pass them to the Softmax classifier, which cooperates with the aspect label sequence to output sentiment polarity labels.

2. The aspect-level sentiment analysis method based on syntactic-guided Transformer according to claim 1, characterized in that According to the dependency matrix B, in cooperation with the inverse distance weighting function and normalization, the normalized weighted matrix is M′∈R n×n , including the following steps: Calculating the shortest path matrix D between each pair of words in the sentence according to the dependency matrix B; Applying the inverse distance weighting function to the shortest path matrix D for calculation to obtain the inverse distance weighting matrix M; The inverse distance weighted matrix M is normalized row by row using Softmax to obtain the normalized weighted matrix M′ ∈ R n×n ; The calculation formula is as follows: Among them, D ij represents the distance between word i and word j on the syntactic tree.

3. The aspect-level sentiment analysis method based on syntax-guided Transformer according to claim 1, characterized in that, Using the dependency matrix B for the calculation of the attention mechanism of the Transformer, including the following steps: Enhance the word vector representation matrix H embed Perform linear projections respectively to obtain three matrices: a query matrix a key matrix and a value matrix where d k is the dimension of the attention head; Introducing the dependency matrix B as additional information into the attention calculation. The complete calculation formula is as follows:

4. The aspect-level sentiment analysis method based on syntax-guided Transformer according to claim 1, characterized in that Mask and pool H out to generate the local aspect word representation H Pooling , including the following steps: Mask the output H of the Transformer out to mask out non-aspect words, so that the representation of aspect words is retained while the representation of non-aspect words is set to zero; Pooling the non-zero part of the aspect words to calculate the representation of the local aspect words: Where A represents the index set of the aspect words, and |A| is the number of aspect words.

5. The aspect-level sentiment analysis method based on syntax-guided Transformer according to claim 1, characterized in that Represent the local aspect word as H Pooling and the global sentence representation H Attn Fuse and pass them to the Softmax classifier, and cooperate with the aspect label sequence to output sentiment polarity labels, including the following steps: Perform interactive attention calculation on H Pooling and H embed to obtain the global sentence representation H Attn ; Fusing the local and global representations through a concatenation operation to form the final classification input features: H Fusion = [H Pooling ; H Attn ​ Input the final classified feature H Fusion Pass it to the Softmax classifier, and cooperate with the aspect label sequence to output sentiment polarity labels; the sentiment polarity labels include positive, negative or neutral.

6. The aspect-level sentiment analysis method based on syntax-guided Transformer according to claim 1, characterized in that, It also includes training and optimization; during the training process, using the standard cross-entropy loss function to optimize the sentiment polarity, and at the same time using the loss function of the conditional random field to optimize the aspect word extraction.

7. The aspect-level sentiment analysis method based on syntax-guided Transformer according to claim 6, characterized in that, The optimizer uses the Adam (Adaptive Moment Estimation) optimizer, and the initial learning rate is set to 10 -5 , and it is dynamically adjusted according to the performance of the validation set during the training process; the training method uses an end-to-end training method, and jointly optimizes the aspect term extraction and sentiment polarity classification tasks through a multi-task learning framework.

8. A syntactic-guided Transformer framework for aspect-oriented sentiment analysis, characterized in that, Including a dependency parsing module, a text encoding module, a syntactic-guided Transformer module, a fusion module, and an aspect word extraction module, used to implement the steps of the aspect-level sentiment analysis method based on the syntactic-guided Transformer according to any one of claims 1 to 7.

Images