Aspect-level sentiment analysis method based on knowledge enhancement and machine reading understanding

Abstract

The invention discloses an aspect-level sentiment analysis method based on knowledge enhancement and machine reading understanding. The aspect-level sentiment analysis method comprises the following steps: S1, preprocessing an original comment text by adopting a word segmentation strategy; s2, based on the sentiment analysis external knowledge data set, post-training is performed on the BERT model through mask language modeling and a next sentence prediction task, exclusive vocabulary distribution and semantic association are learned, and a domain and task knowledge enhancement model is obtained; s3, inputting the preprocessed text sequence into a domain and task knowledge enhancement model, and establishing a bidirectional query mechanism; s4, after the coding features of the domain and task knowledge enhancement model and the query probability of the bidirectional query mechanism are spliced, bidirectional LSTM is input to capture long-range dependence, label transfer logic is constrained through a CRF layer, and an emotional polarity label is output. By introducing an external knowledge base, the defects of the model in the aspect of understanding a text background and implicit information are made up, and the reliability of the model in the aspect of understanding text backgrounds and implicit information is improved. And the accuracy and diversity of sentiment analysis are improved.

Description

Technical Field

[0001] The present application relates to the field of information processing and intelligent recognition technology, and in particular to an aspect-level sentiment analysis method based on knowledge enhancement and machine reading comprehension. Background Art

[0002] Aspect-level sentiment analysis systems focus on identifying sentiment related to specific aspect or opinion terms within text. Unlike traditional document- or sentence-level sentiment analysis, this system deeply explores aspect-level features of specific objects within a text and determines the sentiment of these features (positive, negative, or neutral). The system accurately captures complex semantic relationships and identifies implicit emotional expressions. It is widely used in scenarios such as product analysis and public opinion monitoring, efficiently enabling fine-grained analysis and insight into user sentiment.

[0003] Traditional sentiment lexicon approaches first construct a sentiment lexicon, then use it to calculate sentiment scores for each word in a text. Subsequently, an algorithm calculates the sentiment score and polarity of the entire sentence based on the sentiment scores of each word. Due to the rapid evolution of language and the constant addition of new words, as well as the relatively infrequent update frequency of constructed sentiment lexicons, it is difficult to fully account for the impact of new words on sentiment classification results.

[0004] Machine learning models combine aspect-level sentiment classification tasks into a single classification task. Machine learning approaches to sentiment analysis involve training a classification model on a dataset and then using the trained classification model to predict outcomes. Compared to lexicon-based sentiment analysis methods, machine learning methods achieve higher sentiment classification accuracy. Consequently, machine learning methods have been widely used in sentiment analysis in recent years and have contributed significantly to the development of the field. However, these methods rely on training machine learning algorithms on manually annotated datasets to create pretrained classifiers. Consequently, machine learning methods often struggle to effectively collect contextual data and have limitations when processing long texts.

[0005] The application of deep learning technology in sentiment analysis primarily relies on neural network architectures. These models can autonomously learn the underlying features of text data, establish information associations within sentences, and automatically identify sentiment categories based on the learned feature representations. However, while deep learning models can infer a certain degree of sentiment polarity using contextual information, they are unable to accurately capture the different aspects of sentiment when faced with text that contains multiple emotional aspects, ambiguous sentiment polarity, or mixed emotional expressions.

[0006] Furthermore, pre-trained models also suffer from a serious hallucination problem. First, pre-trained models typically reason based on linguistic patterns learned from large corpora, but these models lack a comprehensive understanding of the text's underlying meaning. Second, while these models excel at grammatical structures and common language patterns, they often over-rely on certain surface-level grammatical features, ignoring subtle emotional nuances and complex context within the text. Furthermore, when faced with ambiguous sentiment polarity or complex context, the model may make unreasonable assumptions, leading to incorrect sentiment classification or sentiment polarity judgments. Summary of the Invention

[0007] In response to the above-mentioned deficiencies in the prior art, the present invention provides an aspect-level sentiment analysis method based on knowledge enhancement and machine reading comprehension to solve the problems of inaccurate capture of emotional elements and inability to understand the deep semantics and background information behind the text due to lack of external knowledge in existing sentiment analysis methods.

[0008] In order to achieve the above-mentioned purpose, the technical solution adopted by the present invention is: an aspect-level sentiment analysis method based on knowledge enhancement and machine reading comprehension, comprising: S1. Preprocess the original review text using word segmentation strategy; S2. Based on the sentiment analysis external knowledge dataset, the BERT model is post-trained through masked language modeling and next sentence prediction tasks to learn the distribution and semantic association of exclusive vocabulary and obtain a domain and task knowledge enhanced model; S3. Input the preprocessed text sequence into the domain and task knowledge enhancement model to establish a bidirectional query mechanism; S4. After concatenating the encoding features of the domain and task knowledge enhancement model with the query probability of the bidirectional query mechanism, the model is fed into a bidirectional LSTM to capture long-range dependencies. The label transfer logic is constrained through the CRF layer to output the sentiment polarity label.

[0009] Further: S2 includes: S21. Based on the sentiment analysis external knowledge dataset, the BERT model is trained through masked language modeling to enable the BERT model to learn the distribution of exclusive vocabulary, the semantic relationship between vocabulary, and contextual dependencies. S22. Based on the sentiment analysis external knowledge dataset, the BERT model is trained through the next sentence prediction task. By determining whether two given sentences are adjacent in the original text, the BERT model can capture sentence-level context information. S23. The BERT model trained through masked language modeling and next sentence prediction tasks is used as a domain and task knowledge enhancement model.

[0010] Further: S21 includes: S211, for each comment in the sentiment analysis external knowledge dataset, randomly select words for masking; S212, segmenting the masked comments and inserting marks at the starting point and segmentation point of the comments to obtain marked comments; S213. Input the marked comments into the BERT model, and use the BERT model to predict the masked words, so that the BERT model can learn the exclusive word distribution, the semantic relationship between words, and the context dependency.

[0011] Further: S22 includes: S221, for each comment in the sentiment analysis external knowledge dataset, randomly select words for masking; S222, segmenting the masked comments and inserting marks at the starting point and segmentation point of the comments to obtain marked comments; S223, randomly select positive samples and negative samples based on the marked comments; S224. Based on the positive and negative samples, the BERT model is trained on the next sentence prediction task by perceiving the task text. The expression is: in, Indicates the probability that sentence B is the next sentence of sentence A, Indicates the probability that sentence B is not the next sentence of sentence A, represents the sentence A in the perception task text, Sentence B represents the perception task text; is the activation function, is the model's prediction score for sentence A and sentence B.

[0012] Further: In S224, the method for generating the perception task text is: after marking the segmentation point of the marked comment, select the mark with the highest probability as the starting position s , and in Among the candidate positions, the mark with the highest probability is determined as the end position , and the starting position s To the end position The text between is used as the perception task text.

[0013] Furthermore: In S3, the bidirectional query mechanism includes: Unrestricted query: used to query a set of aspect words or a set of opinion words; Restricted query: used to query the set of opinion words included in a certain aspect word or to query the set of aspect words included in a certain opinion word; Sentiment polarity query: The sentiment polarity category of the aspect word corresponding to a certain opinion word.

[0014] Furthermore: S4 includes: S41. Obtain the encoding features of the domain and task knowledge enhancement model and the query probability of the bidirectional query mechanism; S42: After concatenating the encoding features of the domain and task knowledge enhancement model with the query probability of the bidirectional query mechanism, the bidirectional LSTM is input to capture the long-range dependency. The label transfer logic is constrained by the CRF layer to obtain the legal transfer relationship between labels. S43. Calculate and input sentiment polarity labels based on the legal transfer relationship between labels.

[0015] Further: In S41, the encoding features of the domain and task knowledge enhancement model Expressed as: in, Represents domain and task knowledge enhancement model, represents a subword sequence, is the dimension symbol, N is the length of the sequence after the word segmentation strategy is applied, d Enhance the hidden layer dimensions of the model with domain and task knowledge, Indicates the word segmentation strategy, is the original comment text sequence, Indicates the first words, is the length of the original review text sequence; Query probability of bidirectional query mechanism Overall probability of passing the span , candidate pairing probabilities of aspect words and opinion words and the probability distribution of sentiment polarity The weighted adjustment is obtained, and the expressions are: in, and are the starting position probability and ending position probability of the candidate span respectively, is the candidate probability of aspect words, is the candidate probability of opinion words, is the activation function, is the bias term, is the weight matrix.

[0016] Further: In S42, the legal transfer relationship between tags E The expression is: in, is the weight matrix, is the bias term, represents the long-range dependency captured by the bidirectional LSTM, Represents a bidirectional LSTM model.

[0017] Further: Sentiment Polarity Label The expression is: in, represents the sentiment polarity label sequence, Indicates in i Positions correspond to sentiment polarity labels The score, Represents the label transfer matrix From the label Transfer to transfer score.

[0018] The beneficial effects of the present invention are: 1. The introduction of external knowledge datasets for sentiment analysis enriches the model's understanding of specific fields, enabling it to better adapt to sentiment analysis tasks in specific fields and improving the accuracy of sentiment analysis; 2. Post-training the BERT model through masked language modeling and next sentence prediction tasks, extracting sentiment elements based on encoded word vectors, optimizing the sentiment element extraction process and avoiding analysis errors caused by ignoring contextual information. 3. It can more efficiently utilize limited annotated data, reduce dependence on large-scale annotated data, and improve model performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 Flowchart of the aspect-level sentiment analysis method based on knowledge enhancement and machine reading comprehension. DETAILED DESCRIPTION

[0020] The specific embodiments of the present invention are described below to facilitate understanding of the present invention by those skilled in the art. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the concepts of the present invention are protected.

[0021] The present invention provides an aspect-level sentiment analysis method based on knowledge enhancement and machine reading comprehension. The basic idea is to post-train the pre-trained language model BERT through domain-specific and task-related datasets to enhance its adaptability and domain knowledge comprehension ability in sentiment analysis tasks. By using the sentiment element extraction model, through bidirectional machine reading comprehension, using bidirectional long short-term memory network and conditional random field method, the effective use of context information is achieved, thus solving the problem of lack of external knowledge in aspect-level sentiment analysis.

[0022] Sentiment analysis that relies solely on the text itself tends to overlook important information in specific scenarios, resulting in inaccurate analysis results. External knowledge is introduced to enhance the model's understanding of specific fields through unsupervised domain data, enabling the model to better process domain-specific texts. Supervised MRC task data is used to improve the model's understanding of tasks and the accuracy of sentiment element extraction. Bi-LSTM (Bi-directional Long Short-Term Memory) and CRF (Conditional Random Field) are used to accurately extract aspect words and opinion words from the text and perform aspect classification to improve the accuracy of sentiment analysis. Figure 1 As shown, the specific steps include: S1. Preprocess the original review text using word segmentation strategy; S2. Based on the sentiment analysis external knowledge dataset, the BERT model is post-trained through masked language modeling and next sentence prediction tasks to learn the distribution and semantic association of exclusive vocabulary and obtain a domain and task knowledge enhanced model. S3. Input the preprocessed text sequence into the domain and task knowledge enhancement model to establish a bidirectional query mechanism; S4. After concatenating the encoding features of the domain and task knowledge enhancement model with the query probability of the bidirectional query mechanism, the model is fed into a bidirectional LSTM to capture long-range dependencies. The label transfer logic is constrained through the CRF layer to output the sentiment polarity label.

[0023] Specifically, a word segmentation strategy is adopted in S1 to decompose low-frequency words into high-frequency sub-words (for example: "Continuity" → "Continuity##"), enhancing the model's ability to process unregistered words.

[0024] Specifically, S2 includes: S21. Based on the sentiment analysis external knowledge dataset, the BERT model is trained through masked language modeling to enable the BERT model to learn the distribution of exclusive vocabulary, the semantic relationship between vocabulary, and contextual dependencies. S22. Based on the sentiment analysis external knowledge dataset, the BERT model is trained through the next sentence prediction task. By determining whether two given sentences are adjacent in the original text, the BERT model can capture sentence-level context information. S23. The BERT model trained through masked language modeling and next sentence prediction tasks is used as a domain and task knowledge enhancement model.

[0025] Preferably, S21 includes: S211, for each comment in the sentiment analysis external knowledge dataset, randomly select words for masking; S212, segmenting the masked comments and inserting marks at the starting point and segmentation point of the comments to obtain marked comments; S213. Input the marked comments into the BERT model, and use the BERT model to predict the masked words, so that the BERT model can learn the exclusive word distribution, the semantic relationship between words, and the context dependency.

[0026] For example, in this embodiment, for each comment in the sentiment analysis external knowledge dataset, 15% of the words are randomly selected for masking. These words are replaced with [MASK] tags. Split into two parts ( and ),in is the complete review, Is the split point, add at the beginning Mark, inserted between two parts Marker used to separate two parts.

[0027] In the process of domain knowledge injection through masked language modeling, the word representation learned by BERT through the masked language modeling task may be affected by data bias. For example, BERT may predict [MASK] in "[MASK] is bright" as "sun" because "sun" frequently co-occurs with "bright" in the text. However, this tendency will cause the model to perform poorly in other fields (such as the computer field) because in the computer field, "screen" is a more appropriate prediction. To overcome this problem, the process of domain knowledge injection performs MLM training on unsupervised data in a specific field. By randomly masking the words in the input text and requiring the model to predict these words, the model can learn the word representation and semantic relationship of the specific field. The probability distribution obtained from the output of the BERT model after masked language modeling training is The expression is: in, For vocabulary The probability of occurrence, is the unnormalized logits vector output by the BERT model after training based on masked language modeling, representing the The prediction scores of all words at positions, is the activation function, For the given context.

[0028] To enhance the model's understanding of the task through the next sentence prediction task, domain knowledge is enhanced by performing NSP training on supervised machine reading task data. By determining whether two sentences are continuous in the original text, the model can more accurately understand the relationship between the question and the context in the machine reading task, thereby more precisely locating the answer. Specifically, S22 includes: S221, for each comment in the sentiment analysis external knowledge dataset, randomly select words for masking; S222, segmenting the masked comments and inserting marks at the starting point and segmentation point of the comments to obtain marked comments; S223, randomly select positive samples and negative samples based on the marked comments; S224. Based on the positive and negative samples, the BERT model is trained on the next sentence prediction task by perceiving the task text. The expression is: in, Indicates the probability that sentence B is the next sentence of sentence A, Indicates the probability that sentence B is not the next sentence of sentence A, represents the sentence A in the perception task text, Sentence B represents the perception task text; is the model's prediction score for sentence A and sentence B.

[0029] In terms of machine reading modeling, in S224, the method for generating the perception task text is as follows: after marking the segmentation point of the marked comment, select the mark with the highest probability as the starting position s , and in Among the candidate positions, the mark with the highest probability is determined as the end position , and the starting position s To the end position The text between them is used as the perception task text; Specifically, in the machine reading modeling process, first, the hidden representation of the input sequence is obtained through the BERT model. , whose dimensions are ; Next, hide the Passed to two independent fully connected layers to calculate the probability distribution of the starting position and the ending position respectively and , whose expression is: in, is the length of the input sequence, is the dimension of the hidden layer, and is the weight matrix with dimension , and is the bias term, is the dimension symbol; Then, through the probability distribution of the starting position and the ending position and , calculate the starting position s To the end position , whose expression is: in, yes The position of the token, so the pointer never points to the token in the question, which ensures that the final answer text fragment is a valid text fragment in the input sequence; Finally, the starting position s To the end position The text between is used as the perception task text.

[0030] Based on machine reading modeling, the model extracts accurate answers from a given question and context. Retraining the aforementioned masked language modeling and next sentence prediction tasks on a task knowledge dataset enables the model to learn task-related sentence-level semantic relationships and reading comprehension capabilities, thereby better understanding the semantic relationship between questions and paragraphs and extracting accurate answers from paragraphs.

[0031] Specifically, in S3, the bidirectional query mechanism includes: Unrestricted query: used to query a set of aspect words or a set of opinion words; Restricted query: used to query the set of opinion words included in a certain aspect word or to query the set of aspect words included in a certain opinion word; Sentiment polarity query: The sentiment polarity category of the aspect word corresponding to a certain opinion word.

[0032] Preferably, the question template of the two-way query mechanism includes: Non-restrictive query: Which are aspect words [aspect word n]? Which are opinion words [opinion word n]? Restricted query: What opinion words [opinion word n] are there for aspect word [aspect word 1]? What aspect words [aspect word n] are there for opinion word [opinion word 1]? Sentiment polarity query: What sentiment polarity does the aspect word [aspect word 1] corresponding to the opinion word [opinion word 1] belong to? Combined with machine reading comprehension tasks, the model's sensitivity to question-text associations is improved through answer span prediction training, thereby strengthening task-oriented semantic understanding capabilities.

[0033] Among them, the sequence start label of aspect words (A) is B-Aspect and the sequence end label is I-Aspect; the sequence start label of opinion words (O) is B-Opinion and the sequence end label is I-Opinion; the labels of sentiment polarity are Positive, Neutral, and Negative.

[0034] Specifically, S4 includes: S41. Obtain the encoding features of the domain and task knowledge enhancement model and the query probability of the bidirectional query mechanism; S42: After concatenating the encoding features of the domain and task knowledge enhancement model with the query probability of the bidirectional query mechanism, the bidirectional LSTM is input to capture the long-range dependency. The label transfer logic is constrained by the CRF layer to obtain the legal transfer relationship between labels. S43. Calculate and input sentiment polarity labels based on the legal transfer relationship between labels.

[0035] Furthermore, in S41, the encoding features of the domain and task knowledge enhancement model Expressed as: in, Represents domain and task knowledge enhancement model, represents a subword sequence, is the dimension symbol, N is the length of the sequence after the word segmentation strategy is applied, d Enhance the hidden layer dimensions of the model with domain and task knowledge, Indicates the word segmentation strategy, is the original comment text sequence, Indicates the first words, is the length of the original review text sequence; Query probability of bidirectional query mechanism Overall probability of passing the span , candidate pairing probabilities of aspect words and opinion words and the probability distribution of sentiment polarity The weighted adjustment is obtained, and the expressions are: in, and are the starting position probability and ending position probability of the candidate span respectively, is the candidate probability of aspect words, is the candidate probability of opinion words, is the activation function, is the bias term, is the weight matrix.

[0036] Furthermore, in S42, the legal transfer relationship between tags E The expression is: in, is the weight matrix, is the bias term, represents the long-range dependency captured by the bidirectional LSTM, Represents a bidirectional LSTM model.

[0037] Furthermore, the sentiment polarity label The expression is: in, Represents the emotional polarity label sequence, which is the emotional polarity category corresponding to each position that the model ultimately determines, such as positive, negative, neutral, etc. represents the emission fraction, which is the i Positions correspond to sentiment polarity labels The score, Represents the label transfer matrix From the label Transfer to transfer score.

[0038] Preferably, after performing the above steps, the performance of the entire aspect-level sentiment analysis model can be evaluated to measure the model's accuracy. The aspect words, opinion words, and sentiment polarity extracted by the model are compared with the source dataset, and the number of correctly matched elements is counted. By calculating metrics such as precision, recall, and F1 value for sentiment polarity judgment, the model's accuracy in determining sentiment tendency can be comprehensively evaluated.

[0039] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. Aspect-level sentiment analysis method based on knowledge enhancement and machine reading comprehension, characterized by: include: S1. Preprocess the original review text using word segmentation strategy; S2. Based on the sentiment analysis external knowledge dataset, the BERT model is post-trained through masked language modeling and next sentence prediction tasks to learn the distribution and semantic association of exclusive vocabulary and obtain a domain and task knowledge enhanced model; S3. Input the preprocessed text sequence into the domain and task knowledge enhancement model to establish a bidirectional query mechanism; S4. After concatenating the encoding features of the domain and task knowledge enhancement model with the query probability of the bidirectional query mechanism, the model is fed into a bidirectional LSTM to capture long-range dependencies. The label transfer logic is constrained through the CRF layer to output the sentiment polarity label.

2. The aspect-level sentiment analysis method based on knowledge enhancement and machine reading comprehension according to claim 1 is characterized in that S2 include: S21. Based on the sentiment analysis external knowledge dataset, the BERT model is trained through masked language modeling to enable the BERT model to learn the distribution of exclusive vocabulary, the semantic relationship between vocabulary, and contextual dependencies. S22. Based on the sentiment analysis external knowledge dataset, the BERT model is trained through the next sentence prediction task. By determining whether two given sentences are adjacent in the original text, the BERT model can capture sentence-level context information. S23. The BERT model trained through masked language modeling and next sentence prediction tasks is used as a domain and task knowledge enhancement model.

3. The aspect-level sentiment analysis method based on knowledge enhancement and machine reading comprehension according to claim 2 is characterized in that: S21 includes: S211, for each comment in the sentiment analysis external knowledge dataset, randomly select words for masking; S212, segmenting the masked comments and inserting marks at the starting point and segmentation point of the comments to obtain marked comments; S213. Input the marked comments into the BERT model, and use the BERT model to predict the masked words, so that the BERT model can learn the exclusive word distribution, the semantic relationship between words, and the context dependency.

4. The aspect-level sentiment analysis method based on knowledge enhancement and machine reading comprehension according to claim 3 is characterized in that: S22 includes: S221, for each comment in the sentiment analysis external knowledge dataset, randomly select words for masking; S222, segmenting the masked comments and inserting marks at the starting point and segmentation point of the comments to obtain marked comments; S223, randomly select positive samples and negative samples based on the marked comments; S224. Based on the positive and negative samples, the BERT model is trained on the next sentence prediction task by perceiving the task text. The expression is: in, Indicates the probability that sentence B is the next sentence of sentence A, Indicates the probability that sentence B is not the next sentence of sentence A, represents the sentence A in the perception task text, Sentence B represents the perception task text; ( ) is the activation function, is the model's prediction score for sentence A and sentence B.

5. The aspect-level sentiment analysis method based on knowledge enhancement and machine reading comprehension according to claim 4 is characterized in that: In S224, the method for generating the perception task text is as follows: after marking the segmentation point of the marked comment, select the mark with the highest probability as the starting position s , and in Among the candidate positions, the mark with the highest probability is determined as the end position , and the starting position s To the end position The text between is used as the perception task text.

6. The aspect-level sentiment analysis method based on knowledge enhancement and machine reading comprehension according to claim 1 is characterized in that: In S3, the bidirectional query mechanism includes: Unrestricted query: used to query a set of aspect words or a set of opinion words; Restricted query: used to query the set of opinion words included in a certain aspect word or to query the set of aspect words included in a certain opinion word; Sentiment polarity query: The sentiment polarity category of the aspect word corresponding to a certain opinion word.

7. The aspect-level sentiment analysis method based on knowledge enhancement and machine reading comprehension according to claim 1 is characterized in that S4 include: S41. Obtain the encoding features of the domain and task knowledge enhancement model and the query probability of the bidirectional query mechanism; S42: After concatenating the encoding features of the domain and task knowledge enhancement model with the query probability of the bidirectional query mechanism, the bidirectional LSTM is input to capture the long-range dependency. The label transfer logic is constrained by the CRF layer to obtain the legal transfer relationship between labels. S43. Calculate and input sentiment polarity labels based on the legal transfer relationship between labels.

8. The aspect-level sentiment analysis method based on knowledge enhancement and machine reading comprehension according to claim 7 is characterized in that: In S41, the encoding features of the domain and task knowledge enhancement model Expressed as: in, Represents domain and task knowledge enhancement model, represents a subword sequence, is the dimension symbol, N is the length of the sequence after the word segmentation strategy is applied, d Enhance the hidden layer dimensions of the model with domain and task knowledge, Indicates the word segmentation strategy, is the original comment text sequence, Indicates the first words, is the length of the original review text sequence; Query probability of bidirectional query mechanism Overall probability of passing the span , candidate pairing probabilities of aspect words and opinion words and the probability distribution of sentiment polarity The weighted adjustment is obtained, and the expressions are: in, and are the starting position probability and ending position probability of the candidate span respectively, is the candidate probability of aspect words, is the candidate probability of opinion words, is the activation function, is the bias term, is the weight matrix.

9. The aspect-level sentiment analysis method based on knowledge enhancement and machine reading comprehension according to claim 8 is characterized in that: In S42, the legal transfer relationship between tags E The expression is: in, is the weight matrix, is the bias term, represents the long-range dependency captured by the bidirectional LSTM, Represents a bidirectional LSTM model.

10. The aspect-level sentiment analysis method based on knowledge enhancement and machine reading comprehension according to claim 9 is characterized in that: Sentiment polarity label The expression is: in, represents the sentiment polarity label sequence, Indicates in i Positions correspond to sentiment polarity labels The score, Represents the label transfer matrix From the label Transfer to transfer score.

Images