The Research Of Aspect-based Fine-grained Sentiment Analysis

The sentiment analysis task can help people to efficiently analyze the massive text data containing the emotional information presented by the internet,and to some extent help people make decisions more efficiently,so it is of great research significance.The traditional sentiment analysis task is to directly judge the emotional tendency of a given text.However,when the text contains multiple emotional targets and the emotions conflict,it can not make appropriate judgments.Therefore,it is very necessary to study the fine-grained sentiment analysis task based on the aspect terms.This task can not only recognize multiple aspect terms in the text,but also discriminate the sentiment polarities corresponding to the aspect terms.This thesis is based on two sub-tasks of the aspect-level fine-grained sentiment analysis task—aspect term extraction task and aspect-based sentiment analysis task.The main task of the aspect term extraction task is to extract the emotional target,that is,the aspect term from the comment text,and the main task of the aspect-based sentiment analysis task is to analyze the sentiment polarity of the specific aspect term in the comment.Therefore,the main work and research results of this thesis are as follows:Based on the transfer learning method,this thesis proposes a method based on the pretraining model to train the domain embedding and sentiment embedding to compensate for the defects that the universal embedding expression can only capture the general semantic information.This embedded expression can be closer to the needs of the two subtasks of the aspect-level based sentiment analysis.In addition,through transfer learning,the general semantic information covered by large-scale corpus and the specific knowledge embedded in the extended corpus can be transferred to the word embedding representation of learning.For the aspect extraction task,an aspect extraction model based on multi-dimensional embedding and self-attention mechanism is proposed.The introduction of the self-attention mechanism effectively solves the problem of insufficient capture of long-distance dependencies between aspect term and contexts.In order to solve the strong problems in experimental data with strong domain correlation,the model adds a word embedding containing domain information.In addition,the model embedding layer uses both general word embedding,domain domain embedding,and word features to improve model performance.Aiming at the aspect-based sentiment analysis task,an improved model based on pretraining method and an improved multi-task method-based model are proposed to solve the problem that the deep learning model has limited performance on small data sets.The introduction of the gated multi-attention mechanism which solves the problem of insufficient long-distance text feature capture by traditional attention mechanism,and learn the different contributions of context to the emotional judgment of specific aspect terms.Secondly,the use of sentiment embedding is used to solve the problem that the generic embedding does not contain emotional information.