Relation Extraction From Complex Text In Open Domain

Relation extraction is to discover relations between entities mentionedin the plain text. It can be used to generate semantic data in form of RDFtriples representing facts. In this paper, we focus on relation extraction fromChinese text, which is less studied compared with that for English. Chinesewords and phrases have great ambiguities on syntax and semantic. Thus,Chinese NLP tools can be insufficient when the sentence is too long or thesentence structure is too complex. Unfortunately, this is the case in the realworld data. In order to tackle the limitation of the current Chinese NLP tools,we propose a method called sentence rolling to generate several enhancedinputs from the original input to help generate the correct relation candidates.In order to rank these candidates in an appropriate way, a voting approach isapplied based on several statistic-based ranking function. Further, a RelationKB is used to help determine the subject part and the object part for theselected relation candidate. We carried out comprehensive experiments onboth real world news corpus and benchmark data combining ChineseTreebank and Chinese Dependency Treebank. The experimental resultsshow that the method can improve the performance of relation extractionsignificantly compared with the existing ones and cost a reasonable time.