Targeted Sentiment Analysis

Targeted sentiment analysis(TSA)is a crucial task for fine-grained public opinion mining,which aims to find the sentiment towards a specific target in a given text.The emergence and development of TSA are closely related to the popularization of the social media platforms and the Internet,and it plays an important role in the fields of e-commerce,public opinion mining,and product analysis.Recently,deep learning methods have been wildly applied to various sort of Natural Language Processing tasks,including targeted sentiment analysis.TSA tasks require the model to identify the position of the entity as well as the emotional expression in the sentence,and the alignment of the entity and its sentiment context is difficult.So far,the existing models still have a long way to go in order to achieving human level performance.To tackle the above problems,we conduct an investigation on TSA and aim to improve the performance of neural network model on TSA tasks through introducing extra information into the model.The main contributions of this thesis include:(1)To address the target and context alignment issue,a novel syntax aware model named SAM is designed to introduce syntactic information into TSA tasks.Beyond extracting the contextual feature of the review text,the model captures syntactic information using a graph convolutional network over the dependency graph built from the dependency tree of the review text.By integrating the dependency relationship,SAM can efficiently propagate sentiment features between syntactically relevant words.This thesis also combines the designed model with the existing work of introducing external corpus for domain adaptation,which further improves the model performance.Experimental results on multiple public datasets show the effectiveness of the proposed method.(2)To address the issue of locating entity and emotional expression,a multi-task learning method using auxiliary task is proposed to introduce more supervision information into the model.To this end,Two different auxiliary tasks and their corresponding multi-task learning models are designed:one with the auxiliary task of target position prediction and the other with the auxiliary task of part-of-speech prediction.By sharing information between the main task and the auxiliary task,the model achieves a better performance on TSA tasks.This thesis verifies the effectiveness of the introduced auxiliary tasks through experiments,and further explores the performance of the methods combining auxiliary tasks and domain adaptation.The two methods proposed in the thesis obtain new state of the art results on two of the public datasets.The proposed methods are more accurate on predicting user’s attitude towards a specific target mentioned in the opinion text,which is helpful for obtaining more detailed and accurate product feedback information and can be applied to improving user experience and product analysis.