Aspect-level sentiment analysis method based on information bottleneck
Through the information bottleneck theory, word embedding is optimized, key features and interactions are identified, and the problem of insufficient explanatory nature of the aspect-level sentiment analysis model is solved, the performance and transparency of the model are improved, and the interpretability and trust of sentiment analysis are enhanced.
Patent Information
- Application Number
- CN202410944239.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-15
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2044-07-15
AI Technical Summary
Existing aspect-level sentiment analysis models have shortcomings in interpretability and transparency in decision-making processes, especially in complex model architectures, which are difficult to identify and select key dimensions, resulting in opacity in model predictions and difficulty in understanding.
Using an information bottleneck theory method, by maximizing the amount of information transfer between input and compressed representation, and minimizing information loss between compressed representation and output, identifying key features and their interactions, optimizing word embedding to reduce data complexity and retain important information.
Improve the explanatory power and performance of the model, enhance the ability to identify emotional tendencies, enable the model to understand and explain user emotions more transparently, reduce the dependence of computing resources, and improve the credibility of the model in complex emotions analysis tasks and A sense of trust.
Smart Images

Figure CN118917422A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of natural language processing, and in particular to an aspect-level sentiment analysis method based on information bottleneck, which is used to enhance the performance and interpretability of an aspect-level sentiment analysis model. Background Art
[0002] In today's rapidly developing information society, natural language processing technology, as a key branch of artificial intelligence, has made breakthrough progress in parsing and understanding human language. In particular, in the subfield of aspect-level sentiment analysis, researchers have invested a lot of effort in developing advanced models that can accurately capture the sentiment tendencies of specific entities or attributes in text. These models play a vital role in many fields, including but not limited to product review analysis, public opinion monitoring, customer feedback processing, etc. They provide data support and insights for business intelligence, public safety assessment and policy making.
[0003] However, as the model architecture becomes increasingly complex, the decision-making process of deep learning models often becomes elusive, forming the so-called "black box" problem. This problem has attracted widespread attention in the field of NLP, especially in those fields that have strict requirements on model interpretability, such as medical diagnosis, legal analysis, financial services, etc. The decision-making process of the model requires not only high accuracy, but also clear logic and transparency so that professionals can reproduce, verify and trust it.
[0004] The 2022 IEEE Transactions on Knowledge and Data Engineering paper titled "Asurvey on aspect-based sentiment analysis: Tasks, methods, and challenges" provides a new taxonomy of aspect-level sentiment analysis, organizing existing research along the axis of sentimental elements of interest, with a particular emphasis on the latest advances in composite aspect-level sentiment analysis tasks. From a solution perspective, the paper summarizes the use of pre-trained language models in aspect-level sentiment analysis, which significantly improve the performance of aspect-level sentiment analysis.
[0005] In the 2024 issue of the Chinese journal Computer Science and Exploration, the title is "A Dual-Feature Fusion Attention Aspect Sentiment Analysis Model Based on BERT", the author is Li Jin et al. This paper proposes a dual-feature fusion attention aspect sentiment analysis model (DFLGA-BERT), which solves the shortcomings of existing models in terms of contextual information and aspect word importance. By designing local and global feature extraction modules, as well as an improved "quasi" attention mechanism, the model shows significant performance improvements on the SentiHood and SemEval 2014Task4 datasets.
[0006] Based on gradient-based interpretation methods and information bottleneck theory, scholars have conducted extensive research on model interpretability in natural language processing, promoting the development of transparency and credibility of neural network models. In these studies, the inventors found during implementation that although existing methods have made progress in determining the importance of words, they did not take into account the differences in importance of different dimensions in aspect-oriented sentiment analysis. Specifically, the authors found in preliminary analysis that not all embedding dimensions are equally important in completing aspect-level sentiment analysis tasks, and that key dimensions vary in different datasets. This shows that although existing methods are effective in many applications, there is still room for further research in the field of aspect-level sentiment analysis:
[0007] The inventors found that there is still room for further research on the method of evolving the user emotional game:
[0008] (1) New methods need to be developed to automatically identify and select the dimensions that are most critical to sentiment analysis tasks. These methods should go beyond traditional assumptions and be able to dynamically adjust according to different data sets and contexts to more accurately reflect and explain the decision-making process of the model.
[0009] (2) Research should focus on how to optimize the model’s intrinsic representation through the information bottleneck theory. This will not only make the model’s predictions more transparent and intuitive to understand, but also enhance the model’s ability to recognize sentiment tendencies by streamlining and strengthening features. Summary of the invention
[0010] In view of the problems and shortcomings of the prior art mentioned above, the present invention provides an aspect-level sentiment analysis method based on information bottleneck, which enhances the explanatory power and performance of the model by identifying and refining key intrinsic dimensions. The IBG method uses the information bottleneck principle to compress and optimize word embeddings, ensuring that while reducing data complexity, information that is critical to the sentiment analysis task is retained. The framework extracts key sentiment features by refining the intrinsic dimensions of word embeddings to improve the model's ability to recognize the sentiment tendency of text.
[0011] In order to solve the above problems, the present invention adopts the following technical solutions:
[0012] The present invention is a decision-making method for the emotional evolution of social network users, which is characterized by: the specific steps are as follows:
[0013] A. Using the information bottleneck theory, by maximizing the amount of information transferred between the input and the compressed representation, minimizing the information loss between the compressed representation and the output;
[0014] B. Predict the gradient between output and input features through computational models, identify key features and their interactions;
[0015] C. In sentiment analysis tasks, the information bottleneck principle is used to compress and optimize word embeddings to ensure that important information is retained while reducing data complexity.
[0016] The specific steps of constructing the method of minimizing the information loss between the compressed representation and the output described in step (A) of the above technical solution are as follows:
[0017] A1. Using the information bottleneck theory, we maximize the mutual information I(X; Z) between the input X and the compressed representation Z to ensure that as much input information as possible is retained, as shown in formula (1):
[0018] L max =I(X; Z) (15)
[0019] where I(X;Z) is the mutual information between the input X and the compressed representation Z.
[0020] A2. While maximizing the information transfer between the input and the compressed representation, minimize the information loss between the compressed representation Z and the output Y, as shown in formula (2):
[0021] L min = -βI(Z;Y) (16)
[0022] Where β is a balance coefficient used to control the trade-off between information compression and preservation, and I(Z;Y) is the mutual information between the compressed representation Z and the output Y.
[0023] A3. Combining the above two objectives, define the total loss function L:
[0024] L=I(X;Z)-βI(Z;Y) (17)
[0025] Among them, L is the objective function to be optimized, and information compression and optimization are achieved by maximizing I(X; Z) and minimizing -βI(Z; Y). Its specific steps are as follows:
[0026] Processing of input feature X:
[0027] p(x) = probability distribution of input data (18)
[0028] According to the probability distribution p(x) of the input data, X is encoded by the encoder q(z|x) to obtain the compressed representation Z:
[0029] Z=q(z|x) (19)
[0030] Optimization of compressed representation Z:
[0031] Using the information bottleneck principle, maximize I(X;Z):
[0032]
[0033] Among them, p(x,z) is the joint probability distribution, p(z|x) is the conditional probability distribution, and p(z) is the marginal probability distribution.
[0034] Generation of output Y:
[0035] Decode Z through the decoder p(y|z) to generate output Y:
[0036] Y=p(y|z) (21)
[0037] Minimization of information loss:
[0038] Minimize I(Z;Y):
[0039]
[0040] Among them, p(z,y) is the joint probability distribution, p(y|z) is the conditional probability distribution, and p(y) is the marginal probability distribution.
[0041] Optimization of the overall goal:
[0042] Combining the above two parts, optimize the total loss function L:
[0043]
[0044] The calculation model described in step (B) of the above technical solution predicts the gradient between the output and input features, and the specific steps are as follows:
[0045] B1. By calculating the gradient of the model's output Y relative to the input feature X, we can determine the impact of each input feature on the output:
[0046]
[0047] Where f(X) represents the model output and Grad(f) represents the gradient.
[0048] B2: Based on the calculated gradient value, identify the input features that have the greatest impact on the model output. The specific method can be to select the first several features with the largest absolute value as key features, or to filter out features that have a significant impact on the output based on a certain threshold:
[0049] X k ={x i ||Grad(f) i| |>0} (25)
[0050] Where Grad(f) i| represents the gradient of the i-th input feature, and Grad(f) is the preset threshold.
[0051] B3. Identify the interactions between features by analyzing the gradient relationship between key features. The second-order gradient (Hessian matrix) can be used to measure the mutual influence between features:
[0052]
[0053] Among them, Hess(f) represents the Hessian matrix, whose elements Hess ij (f) represents the second-order derivatives between input features. The interactions between features can be judged by the off-diagonal elements of the Hessian matrix.
[0054] In step (C) of the above technical solution, the information bottleneck principle is used to compress and optimize word embedding to ensure that important information is retained while reducing data complexity. The specific steps are as follows:
[0055] C1: By maximizing the amount of information transferred between the input X and the compressed representation Z and minimizing the information loss between the compressed representation Z and the output Y, the model is able to reduce the complexity of the data while retaining the information that is critical to the sentiment analysis task, as shown in formula (3).
[0056] C2: In sentiment analysis tasks, word embeddings are compressed and optimized through the information bottleneck principle to ensure that important information is retained while reducing data complexity.
[0057] First, convert the input text data X into word embedding representation E:
[0058] E=Embed(X) (27)
[0059] Among them, Embed(X) represents the conversion of input text X into word embedding representation E:
[0060] Then, the information bottleneck principle is applied to compress and optimize the word embedding representation E to obtain the compressed representation Z:
[0061] Z=Compress(E) (28)
[0062] C3, through the information bottleneck principle, ensure that the information that is critical to the sentiment analysis task is retained during the compression process. During the training process, the balance coefficient β is adjusted to find the best balance between compression and information retention.
[0063] Compared with the prior art, the present invention has the following beneficial effects:
[0064] The present invention is based on the information bottleneck theory and aims to enhance the explanatory power and performance of the model by optimizing and compressing word embeddings in sentiment analysis tasks. In modern social networks, the expression and propagation of user emotions is a complex and dynamic process. The present invention proposes a method to ensure that information critical to sentiment analysis tasks is retained while reducing data complexity by identifying and refining key intrinsic dimensions. The present invention first uses the information bottleneck theory to optimize and compress information by maximizing the amount of information transferred between the input and the compressed representation and minimizing the information loss between the compressed representation and the output. By defining and optimizing the objective function, the present invention can effectively compress and optimize word embeddings to ensure that information critical to sentiment analysis tasks is retained while eliminating irrelevant or redundant information. This method not only improves the performance of the model, but also enhances its explanatory power, enabling the model to more accurately identify and understand users' emotional expressions.
[0065] The present invention predicts the gradient between the output and input features of the calculation model. The present invention can identify the key features that have the greatest impact on the prediction results and reveal the interactions between these features. In this way, the model can better understand and explain the important features and influencing factors in the sentiment analysis task. In the sentiment analysis task, the present invention further uses the information bottleneck principle to compress and optimize word embeddings. By defining and optimizing the objective function, the model can reduce the complexity of the data while retaining information that is critical to the sentiment analysis task. The optimized word embedding representation not only improves the performance of the model, but also enhances its ability to understand and interpret user emotions.
[0066] The aspect-level sentiment analysis interpretation method based on information bottleneck proposed in the present invention achieves a significant improvement in the model prediction accuracy by accurately identifying specific dimensions that are closely related to the sentiment analysis task. This method enhances the transparency of the model decision-making process through the learning of intrinsic dimensions and the introduction of information bottleneck mechanisms, thereby facilitating researchers and practitioners to deeply understand the internal working principles and prediction logic of the model. In addition, by reducing the complexity of the model, the present invention effectively reduces the dependence on computing resources, allowing the model to be efficiently deployed under resource-constrained conditions. More importantly, by improving the interpretability of the model, the present invention helps to build users' trust in the model's prediction results, especially in key application scenarios with extremely high requirements for prediction reliability. The establishment of such trust is of decisive significance for the adoption and application of the model.
[0067] In summary, the proposed method combines the information bottleneck theory and the gradient interpretation framework to provide an efficient and explanatory solution suitable for simulating and analyzing the emotional evolution process of social network users. This method not only improves the performance of sentiment analysis tasks, but also enhances the explanatory power of the model, enabling the model to better understand and cope with the complex social network sentiment expression and propagation process. BRIEF DESCRIPTION OF THE DRAWINGS
[0068] Figure 1 is a flow chart of the aspect-level sentiment analysis method based on information bottleneck theory of the present invention;
[0069] Figure 2 yes Figure 1 The specific steps of preparing input data in the sentiment analysis method based on information bottleneck theory described in step 101, including the detailed process of data collection, data preprocessing and feature extraction;
[0070] Figure 3 yes Figure 1 A flowchart of the detailed implementation steps and specific operation methods of the information bottleneck theory optimization process described in step 102;
[0071] Figure 4 yes Figure 1 Flow chart of aspect-level sentiment analysis in the sentiment analysis method based on information bottleneck theory described in step 103. DETAILED DESCRIPTION
[0072] The implementation process of the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.
[0073] Reference Figure 1 The aspect-level sentiment analysis method based on information bottleneck theory of the present invention comprises the following steps:
[0074] Step 101: Before conducting sentiment analysis, the raw data needs to be fully prepared and processed. This includes data collection, data preprocessing, and feature extraction. Figure 2 , the specific steps are as follows:
[0075] Step 201: collect text data posted by users from social networking platforms (such as Weibo, etc.), including comments, posts, and private messages. These data can be obtained through the platform's API, or by using web crawler tools to capture public text information. During the collection process, laws, regulations, and platform policies must be followed to ensure the legitimacy and privacy of the data.
[0076] Step 202: Clean the collected text data to remove noise and irrelevant information, including stop words, punctuation marks, special characters, and handle spelling errors and abbreviations. Then, perform word segmentation and part-of-speech tagging to ensure that the data is structured for subsequent analysis and modeling.
[0077] Step 203: Extract meaningful text features from the preprocessed data. Common methods include TF-IDF (term frequency-inverse document frequency) and word embedding technology (such as Word2Vec, GloVe). These features will be used to build model input data to support the implementation of tasks such as sentiment analysis.
[0078] Step 204: Store the preprocessed and feature-extracted data in a suitable database or file system for subsequent model training and analysis. You can choose to use a database system (such as MySQL, MongoDB) or a file format (such as CSV, JSON) for data storage and management.
[0079] Step 102: Apply information bottleneck theory to optimize the model. Improve data processing efficiency by defining compression representation and optimization goals to ensure effective information transmission. Figure 3 , the specific steps are as follows:
[0080] Step 301: In the model definition phase, the structure and objective function of the model are established according to the information bottleneck theory. In the process of defining the model, it is necessary to design a network architecture including an input layer, a compression layer, and an output layer. The input layer is responsible for receiving the original data, the compression layer compresses the data into a simplified representation, and the output layer generates the final prediction result. The goal is to maximize the amount of information transferred between the input data and the compressed representation by defining the calculation formula of the mutual information, while minimizing the information loss between the compressed representation and the target output. Establish an appropriate loss function to quantify the target, and design the network hierarchy and parameter configuration of the model to lay the foundation for model training and optimization.
[0081] Step 302: At this stage, an optimization objective function is set to balance information retention and information loss. When optimizing the objective function, techniques such as variational inference are usually used to calculate the mutual information, and algorithms such as gradient descent are applied to minimize the value of the objective function. Selecting an appropriate balance coefficient β\betaβ is a key experimental step and needs to be adjusted through experiments to achieve the best effect.
[0082] Step 303, in the training phase, the parameters of the model are updated through the optimization algorithm. Training includes forward propagation to calculate the prediction results, back propagation to calculate the gradient, and updating the model parameters through the optimization algorithm (such as gradient descent, Adam optimizer, etc.). During the training process, it is necessary to select appropriate initial parameters, set the learning rate, determine the number of training rounds and other hyperparameters, and monitor the progress of training to prevent overfitting. During the training process, the model needs to be tuned and hyperparameters such as learning rate and batch size need to be adjusted to ensure that the model can effectively compress data and achieve good performance on the task objectives.
[0083] Step 304: In the validation phase, the performance of the model in actual applications is evaluated. The generalization ability of the model is tested through an independent validation data set, and the performance of the model is measured using evaluation indicators such as accuracy, recall, F1-score, etc. Further, cross-validation is performed to obtain more reliable evaluation results, and misclassification is analyzed to improve the model. The validation process also includes a detailed analysis of the model results to find out the weaknesses of the model in specific scenarios, and further optimize and adjust the model based on the analysis results.
[0084] Step 103: In the sentiment analysis phase, sentiment classification is performed on the text data posted by the user, referring to Figure 4 , the specific steps are as follows:
[0085] Step 401: In this stage, a classification model suitable for aspect-level sentiment analysis is constructed. First, a suitable model architecture is selected to extract text features from the optimized word embedding. The input of the model is the text vector obtained by pre-training word embedding technology, and the output is the result of classifying the sentiment of each aspect in the text. The calculation formula is as follows:
[0086] v ω =Word2Vec(ω) (29)
[0087] v ω =Gole(ω) (30)
[0088] h i =ReLU(w·x i +b) (31)
[0089] h t =tanh(Wh ·x t +U h ·h t-1 +b h (32)
[0090] p=softmax(W o ·h T +b o ) (33)
[0091] Step 402: training the sentiment classification model. The main task is to improve the classification accuracy of the model by optimizing the loss function. The training process includes batch processing of data, forward propagation calculation of prediction results, calculation of loss function, back propagation update of model parameters, etc. The calculation formula is as follows (34-36):
[0092]
[0093] Step 403: In the sentiment analysis and interpretation stage, the newly input data is classified and the decision basis of the model is analyzed. Gradient calculation is used to analyze the impact of input features on the model prediction results, and visualization technology is used to display the results of sentiment analysis and the interpretation information of the model. These methods help understand the classification decision process of the model, improve the transparency of the model and the user's trust in the results. Sentiment analysis is performed on the newly input data, the sentiment classification results are output, and the transparency of the model is enhanced by explaining the classification decision of the model. The calculation formula is as follows (37):
[0094] Class = argmax(p) (37).
[0095] Example 2: Experiment on aspect-level sentiment analysis method based on information bottleneck theory:
[0096] In order to verify the performance and interpretability of the Information Bottleneck Based Gradient (IBG) framework, we conducted systematic experiments. These experiments used four common datasets: Res14, Lap14, Res15 and Res16, which are detailed with the opinion words of each aspect. In order to evaluate the performance of the present invention, the experiments adopted the most widely used evaluation indicators in the aspect-level sentiment analysis (ABSA) task: Accuracy and Macro F1-score. In order to verify the effectiveness of the present invention in improving the interpretability of the model, the experiment also used the area under the perturbation curve (AOPC) and the post hoc accuracy (Ph-Acc). AOPC evaluates the average decrease in the accuracy of the model when making predictions after removing the first k important words for explanation; while Ph-Acc retains the first k important words and masks the remaining words to evaluate whether the model can still make accurate predictions.
[0097] The present invention uses the bert-base-uncased version and uses the Adam optimizer for the additional components introduced in the original BERT and IBG frameworks. The learning rate of the original BERT is set to 1e-5 and the learning rate of the additional components is set to 1e-4. The compressed intrinsic dimension sizes are 5, 10, 20, 50, 100, and 300 dimensions. The hyperparameter α is set to 0.5.
[0098] In order to explore the performance and interpretability of the proposed framework, the experiment compares the IBG framework with several state-of-the-art baseline models. These baseline models include BERT-SPC, which concatenates the original sentence with aspect words, and other models such as AEN-BERT, LCF-BERT, and ASCM4ABSA. The experimental results are shown in Table 1:
[0099] Table 1 Main results of performance and interpretability
[0100]
[0101] Experimental results show that the IBG framework of the present invention excels in improving the performance and interpretability of existing models. Specifically, the Macro-F1 score of the BERT-SPC model after integrating IBG on the Lap14, Res15, and Res16 datasets has increased by more than 10 percentage points, while the accuracy of the Dual-GCNIBG model on all datasets has increased by about 1%, and the F1 score has increased by 5%. In addition, IBG also performs well in interpretability, significantly improving AOPC and Ph-Acc, verifying its ability to better capture opinion words and sentiment-related features by learning intrinsic dimensions. Ablation experiments, as shown in Table 2, further demonstrate the importance of iBiL and information bottleneck structures for aspect-level sentiment analysis tasks, and removing these components will lead to a significant decrease in model performance.
[0102] Table 2 Ablation experiment
[0103]
Claims
1. An aspect-level sentiment analysis method based on information bottleneck, characterized by: The specific steps are as follows: A. Using the information bottleneck theory, by maximizing the amount of information transferred between the input and the compressed representation, minimizing the information loss between the compressed representation and the output; B. Predict the gradient between output and input features through computational models, identify key features and their interactions; C. In sentiment analysis tasks, the information bottleneck principle is used to compress and optimize word embeddings, reducing data complexity while retaining important information.
2. The aspect-level sentiment analysis method based on information bottleneck according to claim 1, characterized in that: The specific steps of the above step (A) are as follows: A1. Using the information bottleneck theory, we maximize the mutual information I(X; Z) between the input X and the compressed representation Z to ensure that as much input information as possible is retained, as shown in formula (1): L max =I(X;Z) (1) where I(X;Z) is the mutual information between the input X and the compressed representation Z; A2. While maximizing the information transfer between the input and the compressed representation, minimize the information loss between the compressed representation Z and the output Y, as shown in formula (2): L min =-βI(Z;Y) (2) Where β is a balance coefficient used to control the trade-off between information compression and preservation, and I(Z;Y) is the mutual information between the compressed representation Z and the output Y; A3. Combining the above two objectives, define the total loss function L: L=I(X;Z)-βI(Z;Y) (3) Among them, L is the objective function to be optimized, and information compression and optimization are achieved by maximizing I(X; Z) and minimizing -βI(Z; Y). The specific steps are as follows: Processing of input feature X: p(x) = probability distribution of input data (4) According to the probability distribution p(x) of the input data, X is encoded by the encoder q(z|x) to obtain the compressed representation Z: Z=q(z|x) (5) The optimization of the compressed representation Z is: Using the information bottleneck principle, maximize I(X;Z): Among them, p(x,z) is the joint probability distribution, p(z|x) is the conditional probability distribution, and p(z) is the marginal probability distribution; Generation of output Y: Decode Z through the decoder p(y|z) to generate the output Y: Y=p(y|z) (7) Minimization of information loss: Minimize I(Z;Y): Among them, p(z,y) is the joint probability distribution, p(y|z) is the conditional probability distribution, and p(y) is the marginal probability distribution; Optimization of the overall goal: Combining the above two parts, optimize the total loss function L:
3. The aspect-level sentiment analysis method based on information bottleneck according to claim 1 is characterized in that: The specific steps of the above step (B) are as follows: B1. Determine the impact of each input feature on the output by calculating the gradient of the model's output Y relative to the input feature X: Where f(X) represents the model output and Grad(f) represents the gradient; B2: Based on the calculated gradient value, identify the input features that have the greatest impact on the model output; The specific method is to select the first several features with the largest absolute value as key features, or to filter out features that have a significant impact on the output based on a certain threshold: X k ={x i ||Grad(f) i| |>0} (11) Where Grad(f) i| represents the gradient of the i-th input feature, and Grad(f) is the preset threshold; B3. By analyzing the gradient relationship between key features, we can identify the interaction between features. We use the second-order gradient, namely the Hessian matrix, to measure the mutual influence between features: Among them, Hess(f) represents the Hessian matrix, whose elements Hess ij (f) represents the second-order derivatives between input features; the interaction between features is determined by the off-diagonal elements of the Hessian matrix.
4. The aspect-level sentiment analysis method based on information bottleneck according to claim 1 is characterized in that: The specific steps of the above step (C) are as follows: C1: By maximizing the amount of information transferred between the input X and the compressed representation Z and minimizing the information loss between the compressed representation Z and the output Y, the model is able to reduce the complexity of the data while retaining the information that is critical to the sentiment analysis task, as shown in formula (3); C2: In sentiment analysis tasks, word embeddings are compressed and optimized through the information bottleneck principle to ensure that important information is retained while reducing data complexity; First, convert the input text data X into word embedding representation E: E=Embed(X) (13) Among them, Embed(X) represents the conversion of input text X into word embedding representation E: Then, the information bottleneck principle is applied to compress and optimize the word embedding representation E to obtain the compressed representation Z: Z=Compress(E) (14) C3. Through the information bottleneck principle, ensure that the information that is critical to the sentiment analysis task is retained during the compression process; during the training process, adjust the balance coefficient β to find the best balance between compression and information retention.
Citation Information
Patent Citations
Text sentiment classification method through maximal independent multivariate logistic regression model based on Spark
CN108536838A
Aspect-level sentiment analysis method based on LSTM (Long Short Term Memory) and grammar distance
CN114662503A
Multi-level fusion aspect category sentiment analysis method based on self-attention
CN116384372A
Method for analyzing fine-grained text sentiment based on users' harshness
US20230100687A1
