Research On Aspect-Based Sentiment Analysis For Commodity Review Text

Aspect-based sentiment analysis for product review text refers to infer the user’s sentiment orientation for the aspect words(or attribute words)of a given product by combining the context of the review text.Currently,the known aspect-based sentiment analysis methods are mainly based on the neural network to perceive the emotional information contained by the target object in the sentence,and summarize and integrate the emotional information,so as to infer the emotional tendency expressed by the user for a given aspect word.However,due to the complexity of annotation of aspect-based data,there is a small scale of data available for model training,which results in the inability of neural networks to fully learn the semantic patterns and affective features of text,resulting in the bias of the model in predicting the affective tendency of aspectlevel words.To solve this problem,this paper proposes an aspect-based sentiment analysis model based on transfer learning and filtering mechanism.The main contents of this paper include the following two points:(1)This dissertation proposed an aspect-based sentiment analysis model based on transfer learning to solve the problem that the neural network could not be fully trained due to the lack of aspect-level training corpus.Firstly,on a large-scale document-level annotated corpus,the ABSA-TL model is prelearned for document-level sentiment analysis tasks,so that the model can prelearn rich text language patterns and sentiment features.Then,on the aspect-level annotated corpus,the pre-learned ABSA-TL model was trained twice.The purpose of this training was to make the model pay more attention to and perceive the emotional expressions corresponding to the given aspectlevel words.Experiments are carried out on the public aspect-based sentiment analysis dataset.The experimental results verify the effectiveness of the ABSA-TL model.(2)Furthermore,considering that the emotional noise unrelated to aspect words will be introduced in the process of transfer learning,an aspect-based sentiment analysis model based on transfer learning and filtering mechanism is proposed.Firstly,the document-level pre-training module is used to learn the general language pattern and sentiment characteristics of the text,and then the knowledge is transferred to the aspectlevel sentiment analysis module by sharing parameters.Then,an attention filter module is designed,which is mainly used to filter the emotion information unrelated to aspect words in document-level knowledge.Finally,with the assistance of the document-level pre-training module,the aspes-level sentiment analysis module and the filtering module are jointly trained to make use of document-level knowledge and reduce the impact of redundant sentiment noise.The experimental results show that the migration study level and filtering mechanism of the binding energy and effective improvement in affective forecasting accuracy.Figure [20] Table [8] Reference [75]...