Semantic mapping of entity relation mentions aims to link the relation mentions in natural language texts to the attribute relations in Knowledge Base. It is a key supporting technique for many applications, e.g., konwledge base population, semantic retrieval etc. Benefitting from the fundamental research of open information extraction and the population of konwledge base, relation mapping has been a hot issue in the natural languae processing research area. There are two main difficulties in relation mapping. One is the mismatch between relation mentions and attributes since the relation mentions are always various. The other is the ambiguity of relation mentions because a relation mention may refer to different attributes whlle there may be multiple attributes between an entity pair. This thesis makes an intensive study on the technique of mapping entity relation mentions to semantic items in Knowledge Base. The main contributions and innovative points are summarized as follows. 1. Relation Mapping Based on Instances and Attribute Names Semantics Expansion Because of the problems of relation mention variations and coincidental matches, relation mapping algorithms are required to capture the semantics behind the various expressions while immuning the noise. To solve this problem, existing methods rely on the information of entity pairs, i.e., they assume that if a relation mention share much more entity pairs with attribute, then the mention is more likely to express the semantics of that attribute. We believe that the relation mapping relies on not only the entity pairs information but also the relation mentions themselves. Therefore, we propose a relation mapping method which combines entity pairs and attribute name semantics expansion. We first expand attribute candidates with their synsets and then match the semantics similarity between the elements in the synsets with relation mentions. The matching results are combined with entity pairs by stacking technique to achieve the goal of relation mapping. Experimental results demonstrate that the average accuracy of our method can achieve 0.744 for relation mapping on PATTY dataset, which improves current methods rely on entity pairs by 0.245. 2. Generative Model for Relation Mapping Open Information Extraction (Open IE) could extract domain-independent relational triples from natural language texts which keep the variety and richness of the language. However, open IE didn’t recognize the exact semant...
修改评论