Research Of Aspect-Based Sentiment Analysis Based On BERT

As the Internet penetrates into all walks of life,TB-level text messages are published and spread on the Internet every day.Massive text data provides rich training resources for machine learning and promotes the vigorous development of the field of natural language processing.Attribute-level sentiment analysis is a technology that automatically extracts the emotional information of entities in different attributes from the original text.Attribute-level sentiment analysis technology has broad application prospects.For example,it automatically extracts consumer evaluations of different aspects of the product from product reviews,automatically sorts out the public's views on all aspects of hot events from Weibo tweets,and automatically extracts user-to-company,organization-wide Functional satisfaction,etc.Pre-trained models + fine-tuned transfer learning methods have developed rapidly since 2018,and have made breakthrough progress in the field of natural language processing.Open AI GPT,ELMo,BERT and other deep models have greatly improved the degree of semantic understanding,and also brought new ideas to sentiment analysis tasks.This paper studies the attribute-level sentiment analysis technology based on BERT language model.The research work of this paper includes:(1)This paper proposes an attribute word expansion method based on the LDA topic model.By expanding the attribute words into a series of related words,it not only solves the difficulty that the attribute words in the attribute-level sentiment analysis task have context dependence,but also resolves the implicit attributes in the attribute-level sentiment analysis task and their corresponding sentiment words Difficult to extract.(2)This paper proposes an attention mechanism that combines attribute-related words and emotional tendencies.By integrating attribute-related words and sentiment tendencies into attribute-emotional text as supplementary information of attributes,the BERT language model increases the attention of attribute-related emotional information.Solve the problem that attribute-level sentiment analysis is easy to be disturbed by other sentiment information when performing sentiment analysis on a certain attribute.(3)This paper proposes a method for inputting sentence pairs based on question and answer patterns that is close to natural language.This input method can reduce the confusion of the BERT language model on the input sentence pairs,and reduce the training difficulty of the BERT language model when fine-tuning the attribute-level sentiment analysis task.(4)This paper verifies the effectiveness of the proposed algorithm through comparative experiments.Comparing the latest research results of attribute-level sentiment analysis,the performance of the proposed algorithm is 6% higher than the BERT-Condition-CNN algorithm and 1% higher than the BERT-pair-QA-M algorithm.