Opinion Target Extraction Based On Deep Neural Network

Opinion target extraction is an important part of fine-grained sentiment analysis,and it is also an essential basis for accurate product review mining or public opinion monitoring.Template rules based methods,have become the preferred choice in the actual scene due to their simplicity and efficiency.However,the formulation of template rules is often targeted at a specific domain,and often encounters problems such as template matching order conflict.Conditional Random Fields based methods,have no problems as template rules based method,but they also rely on a large number of manual process of feature engineering.These features are decisive for the extraction performance.Accordingly,researchers began to study how to automatically learn these features from the review text.Recent studies have found that deep neural networks have the ability to automatically learn feature representations.Therefore,the deep neural networks based opinion target extraction has become an important research direction in this field.This paper aims at the problems existing in the deep neural network based opinion target extraction,and studies how to combine linguistic morphology,lexicology,syntactic structure and other prior knowledge with deep neural network to further boost the extraction performance of opinion targets.Based on careful analysis and research on the issues in opinion target extraction,the main research effort and innovation of this paper are as follows:(a)We analyze and deconstruct the currently leading opinion target extraction methods based on deep learning framework,and study the differences between traditional machine learning methods and deep learning methods in the semantic representation of words and the advantages of representing words through dense vectors in the context of short texts,and compare different construction manners of the deep network in the opinion target extraction task and the links between these networks in the hierarchy and topology structure.(b)Aiming at the shortcomings of current deep neural network in extracting opinion targets,an improved model incorporating syntactic structure information is proposed.We explore the linguistic characteristics of opinion targets such as morphology,Part-of-Speech,and dependency syntax,and affix these characteristic information to the recurrent network in which the character vector is for alleviating the problem of Out-of-Vocabulary words,the Part-ofSpeech vector for capturing properties of particular words and the transfer mechanism of dependency information between words for enhancing the transmission of syntactical information.And finally,the outputs of the recurrent network are fed to a CRF layer which is capable of capturing label dependencies information.(c)The comparison experiments prove that the proposed model is superior to other latest methods,and an ablation experiment is designed to verify the contribution of different components of the full model.On the public data set provided by SemEval 2014 Task 4,the experimental results show that the proposed model is not only better than traditional machine learning methods,but also goes beyond the current leading models.