Extracting Opinion Relation From Web Text Based On Word Alignment Model

Extracting opinion relation is the basic task of opinion mining,which focuses on extracting opinion targets and opinion words in review texts from web.Currently,it has caused a lot of attention of domestic and foreign scholars,and some achievements have been made.However existing methods are hard to deal with,for one thing the weight measurement of relation between opinion targets and opinion words is not accurate,for another do not consider the relations among opinion targets(opinion words)themselves,which affect the accuracy of final extraction.Thus,in the paper we focus on the problem and aim to improve the accuracy,the details of the research are proposed as followed:(1)Previous works in mining relation between opinion targets and opinion words use artificial templates and rules.But the weight measurement of relation between opinion targets and opinion words is not accurate.This paper proposes an approach to extract opinion targets and opinion words based on word alignment model.This method extracts the associations between opinion targets and opinion words by using word alignment model,whose strength is estimated with word distance.To model these associations,the approach constructs a bipartite graph.Then a domain relevance measure is used,and random-walk algorithm is applied to calculate the confidence of each opinion target candidates and opinion word candidates.The method is evaluated on the labeled corpus of task 3 in COAE2011.The experiment results show that our approach achieves a better average F1 than baseline method in three areas of the electronic products,the video entertainment and financial securities.(2)The approach based on word alignment model only applies word distance between opinion targets and opinion words to compute the weights of relations but do not consider the multi-layer relations among opinion targets and opinion words.Thus,we propose a novel model by using multiply relations for extracting opinion targets and opinion words.Similarly it extracts the associations between opinion targets and opinion words by using word alignment model,and then simultaneously considers syntactical dependency relations between opinion targets and opinion words,and co-occurrence relations among opinion targets(opinion words)themselves.Finally the approach constructs a bipartite graph,and random-walk algorithm is applied to calculate the confidence of each opinion target candidates and opinion word candidates.Experimental results indicate that average F1 can be improved by 3%in our method for extracting opinion targets compared with the method based on word alignment model(WAM_I).(3)The dependency syntactic parsing often only suitable for the text of complete sentence structure,but for some colloquial sentences or review texts that contain syntax error,it may cause a problem.So we proposes a review sentence compression model combined with conditional random fields(CRF)for extracting opinion targets and opinion words.We firstly compress the sentence by using CRF model,retain the main opinion elements of the sentence.Then the results are finally incorporated into the model by using multiply relations for extraction task.Experimental results demonstrate that the framework has a certain degree of improvement in the precision,recall and F1 than existing methods.