Relation extraction using semantic information

Research works on relation extraction have put a lot of attention on finding features of surface text and syntactic patterns between entities. Much less work is done using semantically relevant features between entities because semantic information is difficult to identify without manual annotation. In this paper, we present a work for relation extraction using semantic information as we believe that semantic information is the most relevant and the least noisy for relation extraction. More specifically, we consider entity type matching as one of the additional feature because two entities of a relation must be confined to certain entity types. We further explore the use of trigger words which are semantically relevant to each relation type. Entity type matching controls the selective preference of arguments that participate in a relation. Trigger words add more positive evidences that are closely related to the target relations, which in turn help to reduce noisy data. To avoid manual annotation, we develop an automatic trigger word identification algorithm based on topic modeling techniques. Relation extraction is then carried out by incorporating these two types of semantic information in a graphical model along with other commonly used features. Performance evaluation shows that our relation extraction method is very effective, outperforming the state-of-the-art system on the CoNLL-2004 dataset by over 13 % in F-score and the baseline system without using these semantic information on Wikipedia data by over 12 %.