Chinese Sentence Similarity Computing Based on Frame Semantic Parsing

Sentence similarity computing plays an important role in many tasks of natural language processing. Recent approaches to sentence similarity computing have focused on word-level information without considering the semantic structural information; these methods based on the sentence structure are not generally desirable as they are severely affected by the incomplete description of sentence semantic. Hence, similarity computing isn't able to get better results. To solve this problem, this paper proposes a novel similarity computing approach based on Chinese FrameNet. The approach implements to measure the sentences' semantics similarity by multi-frame semantic parsing, importance measure of frames, similar match of frames, similarity computing between frames and so on. From the frame perspective, the multi-frame semantic parsing comprehensively describes sentences' semantics by identifying all the target words, choosing corresponding frames and labeling the frame elements. On that basis, the similarity result can be more accurate by distinguishing the different frames' importance in accordance with the semantic coverage area of the frame. In addition, by means of extracting the semantic core words of the frame element, the approach improves the precision of similarity among the frames of chunk form. The sentences which contain multiple target words are chosen as the corpus of the experiments. In contrast with traditional approaches, the results show that the proposed approach could achieve better similarity results.