Abstract：In this paper, we propose an unsupervised approach to inferring implicit discourse relation (i.e. relation such as contingency or comparison that is not marked with a connective) based on information retrieval. With Google search engine, we extract candidate explicit relations which are similar to implicit relation on syntactic and semantic levels. These explicit relations which have achieved high accuracy are used to infer implicit relation. The proposed approach contains three modulesfirstly, we construct high-quality queries and extract candidate explicit relations; then three inference models (Similarity, Confidence, Relevance) are presented to evaluate the quality of queries and candidate relations; and finally, base on learning to rank candidate relations, we acquire the statistics of discourse senses distribution to realize the prediction of implicit discourse relation. Experimental results on the PDTB 2.0 show the accuracy of 54.3%, which is a significant improvement of 14.3% over the supervised system. Key wordsimplicit discourse relation; unsupervised; information retrieval; PDTB 2.0
2013, Vol. 27 
