Opinion Targets And Opinion Words Collaborative Extraction Technology For User Reviews

The rapid development and popularization of network technology have exerted a great influence on the living pattern of human beings.At present,online shopping has become one of the main shopping modes for people,and users' reviews have important value for commodity/service providers and other potential users.However,the sheer volume of user reviews makes it difficult for users to quickly retrieve the desired information.Therefore,automated and intelligent analysis of user reviews is extremely important.Opinion target and opinion word are two core components in users' reviews.The former describes what users' review is about,and the latter expresses what users' attitude towards the object.These two items carry most of the users' opinion information.Therefore,the automatic extraction of opinion words and opinion targets from the reviews is a basic work for reviews analysis and intelligent application.This paper focuses on the cooperative extraction of opinion targets and opinion words.The main work includes:(1)In the existing work,the supervised opinion mining method can achieve better extraction effect.However,such methods rely on high-quality training samples,and to label these samples are time-consuming,laborious and error-prone process.We propose a method to iteratively obtain high-quality opinion word-pairs from users' reviews by crowdsourcing computing.Firstly,the reliability of workers is evaluated by EM algorithm.Secondly,task distribution is based on the reliability of workers.Finally,the final labeling results are produced by combining the reliability of the worker with the opinion dependency information of the opinion word-pairs in the returned results.Using the generated results in each iteration of the process to reevaluate the reliability of the workers can guarantee the quality of the results increasing the cost.Experimental results show our method is able to obtain more opinion word-pairs under a fixed budget.(2)A method of opinion word-pairs extraction,based on the analysis of opinion dependence with attention mechanism.In the same category of commodities,Opinion word-pairs usually have strong opinion dependence relation to the opinion targets and the opinion words contained in them.Therefore,in the extraction process of opinion word-pairs,they can be extracted by analyzing the opinion dependence relations among the words in the review sentences.Firstly,a dependency relation analysis model is constructed to obtain the dependency relation information of each word in a review sentence,which the basic model is defined as LSTM neural network in the paper.Secondly,it is assumed that one of the item that opinion word-pairs contained in review sentence is known,and the known item is used as the model's attention information,so that the model can focus on extracting the words of phrases associated with the known item with strong opinion dependence from the review sentence as another unknown item in the opinion word-pairs.Finally,the word-pairs with the highest score of the opinion dependence relation are output as the opinion word-pairs.Then a compound model is designed to realize the mining of opinion word pairs without the need to the known items in advance by combining the two models which contain the information of different known items in the opinion word-pairs.