Research And Application On Entity Relation Classification Using Relatedness Computing

With the rapid development of the Internet technology, data resources on the Webare showing exponential growth trend, information extraction is to obtain valuableinformation from these massive data resources. The ultimate aim of informationextraction is to obtain factual information from text resources, which are called entities.From a more abstract point of view, anything has a specific set of attributes can beregarded as an entity. Consequently, determining the relationship between entities is animportant task in entity related research.This thesis focuses on calculating the relatedness and classifying the relationshipbetween entities, we compute entity relatedness using their attributes, and treat thesemantic relatedness as features of entity relationship classification. Specifically, thispaper mainly consists of the following three parts.This thesis firstly analyzes the traditional relatedness calculation methods, aftercomparing the effect of different models, we take text entity as an example, this thesisproposes the word-text mutual guidance (WTMG) model. This model aims to exploitthe relationship between text and its component words, text relatedness can becalculated using words relatedness and vice versa. The WTMG model is then used tocalculate the relatedness between mobile app entities, review similarity and apprelatedness are iteratively calculated based on the model.After determining the relatedness between entities, this thesis attempts to classifythe relationship between entities. We firstly introduce the surface linguistic featuresused by traditional relationship classification methods, and then present the semanticrelatedness features, these two kinds of features are fused to the classification tasks.After analyzing the shortage of traditional kNN and SVM, this thesis proposes thesemi-supervised bootstrapping method. The bootstrapping method can avoid thedependence on massive labeled data and the effect is close to a supervised model.According to the characteristics of the mobile app entity relationship, this thesispresents an approach to select the initial labeled set based on stratified sampling strategyin the bootstrapping procedure, the method can achieve better results by controlling theiteration stops and other parameters.The relatedness and relationship between entities have a wide range of applicationsin recommendation systems. This thesis combines the results of the two partsmentioned above and constructs an entity relation network like the knowledge graph. Inthis network, the node represents entity and the edge between nodes is the relationshipbetween entities, while the edge weight is the relatedness between nodes, a mobile app recommendation system is achieved based on this network. Different from the similarrecommendations given by the existed systems, our recommendation system not onlypresents the similar recommendations but also gives related ones, which are more inline with user habits.