Opinion Retrieval Based On Knowledge Graph

With the rapid development of the Internet,more and more user interactive media have emerged,people like to send their opinions on these platforms.It has large commercial and academic value to study these opinion documents.With the increase of data volume,opinion retrieval has gradually become a hotspot in natural language processing,it studies the way to find related documents with opinion from social media and other document collections.Opinion retrieval requires the retrieved documents are not only related to the given topic,but also have comments or views on the subject.Currently,a lot of researches on opinion retrieval have been carried out by scholars at home and abroad,which have got some achievements.However,the user's input is usually too short to describe the information need accurately,and the current approaches often overlook this point.To solve the problem,this paper uses knowledge graph to understand the information need of queries,and the details of the methods are proposed as followed:(1)In view of the fact that user queries are usually too short to represent the information need of the queries accurately,an opinion retrieval model based on knowledge graph entity text is proposed,and we use the knowledge graph to understand the information need of the users.Firstly,we get the candidate of query expansion terms by knowledge graph,and calculate four kinds of features of each candidate named term distributions,co-occurrence frequency,proximity and collection frequency.Then,we choose the expansion terms by SVM classifier with the features.Finally,we expand the generative opinion retrieval model using the expansion terms to get the opinion retrieval result.Experimental results on Sina Microblog and Twitter datasets show that our proposed method obtains significant improvements in terms of MAP and NDCG over the baseline approaches.(2)Only the feature of entity text is used in the opinion retrieval model based on entity text,other information in the knowledge graph is underutilized.Thus,we propose an opinion retrieval model which integrates the categories of knowledge graph entities.First,the entities of the query and documents are linked to DBpedia to obtain the category attributes.Then,the entity category score is calculated according to the formula like BM25,and the category score is merged with the original topic relevance score to get the new retrieval model.Experimental results both on the Chinese and English datasets show that the retrieval efficiency of this model is better than that proposed in(1).(3)The opinion retrieval model which integrates the entity category only takes into account the description text and category attributes of a single entity,but the relationship between entities is not used.So,an opinion retrieval model based on entity relationship and category is proposed.When acquiring candidate expansion terms,we consider not only the entity text,but also the other entities associated with the entity.At the same time,we add a weight factor of the expansion terms in the model to measure the differences between the expansion words.The experimental results show that the proposed method can retrieve the required opinion documents more efficiently than(1)and(2).