Research Methods For Aspect-level Sentiment Analysis Based On Attention Mechanism

Aspect-level sentiment analysis is one of the main research orientations on sentiment analysis.In recent years,it has received widespread attention from many researchers.Compared with document-level sentiment analysis,it has a more fine-grained sentiment evaluation object.Aspect-level sentiment analysis includes two sub-tasks: Aspect-term Extraction and Aspect-term Sentiment Classification.The goal of aspect-term extraction is to identify all the evaluation objects(i.e.aspects),and the goal of aspect-term sentiment classification is to identify all sentiment polarities of different evaluation objects.This thesis focuses on two subtasks of the aspect-level sentiment analysis,conducts research on general attention-based methods of aspect-term extraction and aspect-term sentiment classification.In summary,the main research work of this thesis is summarized as follows:(1)Aspect-term extraction.Frist,the thesis analyzes the aspect-level sentiment analysis tasks,describes the problems and challenges in the aspect-level task.For aspect-term extraction,the thesis introduces an advanced unsupervised neural attention-based model and analyzes the experimental results of the model.(2)Aspect-term sentiment classification.The thesis proposes a new multi-layer aspect-context interactive attention representation model for the limitation of LSTM combined with the general structure of attention mechanism in aspect-term sentiment classification.The model only relies on the attention mechanism to generate interactive sequence-sequence representations of aspects and their contexts,replacing the previous LSTM method of separating model aspect and context.It can extract features related to specified aspects during the context sequence modeling process,while generating high-quality aspect representations.In the experiments,compared with the existing thirteen advanced methods on multiple datasets,the experimental results show that the model proposed in this thesis can be significantly improved on Restaurant dataset,and two very competitive results have also been obtained on Laptop and Twitter datasets.(3)Design and implementation for aspect-level sentiment analysis system.We combine the proposed aspect-term sentiment classification model with the current advanced aspect-term extraction model to develop a system for aspect-level sentiment analysis of user comment text.First,the system can extract all aspect terms from the roughest user review text.On this basis,it can predict the sentiment polarity corresponding to each aspect-term.By integrating aspect-term extraction and aspect-term sentiment classification functions,the system has good practicality.