Goal-directed Fine-grained Sentiment Analysis Combined With Pretrained Models

Sentiment analysis is one of the important fields of natural language processing,and its purpose is to collect subjective opinions or feelings based on a specific topic from various data sources.But sentiment analysis may not be able to accurately analyze different aspects of sentiment in sentences with compound sentiment.Therefore,a more fine-grained sentiment analysis task,Aspect-based sentiment analysis,is proposed,which aims to identify the emotional tendencies of multiple aspects involved in the subject being analyzed.Aspect-level sentiment analysis consists of multiple sub-tasks,including aspect word extraction,opinion word extraction,aspect sentiment classification,and sentiment polarity analysis.Aspect word extraction task and opinion word extraction task are two subtasks of aspect-level sentiment analysis.Many works have been made in these two subtasks,but these two subtasks are often considered independently,ignoring aspect words and opinion words.The semantics are closely related.Aiming at this problem,based on the research of existing algorithms,this thesis proposes to treat the two subtasks as a whole,and implement a multi-task framework for them,so that the two subtasks can better complete the information extraction.The main research contents include: The following three points:1)In view of the limitation that the traditional word vector is a static word vector and cannot be adjusted according to specific sentences,this thesis proposes to use the BERT pre-training model as the word embedding layer of the model.BERT provides dynamic word vectors,which can dynamically adjust word vectors according to different data sets.The adjusted word vectors can better express semantic information,thereby improving the accuracy of extraction tasks.2)The aspect word information obtained by the aspect word extraction task is used as the input of the viewpoint word extraction task to improve the performance of the viewpoint word extraction task.And the feature extraction layer of the existing target-oriented opinion word extraction task model is improved.3)It is proposed to use the pipeline model and the joint model to combine the aspect word extraction task and the viewpoint word extraction task to build a multi-task framework.In the joint model,multiple tasks share the word embedding layer,so that multiple tasks can exchange information through the word embedding layer.This structure can effectively reduce the error propagation caused by the prediction error of the upstream model,thereby improving the extraction ability of the downstream model.In the two sub-tasks of aspect word extraction and viewpoint word extraction,this thesis links two related modules and proposes an optimization scheme for the existing model,and through experiments prove that the use of a multi-task framework to combine related modules can improve the model effect.