The Sentence Similarity Research Based On Frame Kernel Semantic Dependency Graph

The computation of sentence similarity is important and basic in the system of natural language processing. Improving the efficiency of sentence similarity plays an important role in question answering, information retrieval and machine translation. This thesis proposes the frame semantic dependency graph model based on Chinese FrameNet to present the frame semantic information for a sentence after analysis the domestic and aboard research. We compute the similarity of two sentences by computing the similarity of their frame semantic dependency graphs and their external component similarity. Furthermore, we identify the semantic key words of the frame elements by labeling multi-word chuck for it, thus, we convert frame semantic dependency graph into frame kernel semantic dependency graph. The main work of this paper is as follows:Chinese frame semantic dependency graph (FSDG) was proposed referencing the kernel dependency graph in English. It is a formalization representation for the core semantic structure of a sentence based on Chinese FrameNet. Extracting the frame semantic dependency graph of a sentence means obtaining the semantic framework of the sentence.The mathematical models of CFN and the Dependency graph are established, based on which calculates the dependency graph similarity and their external components similarity respectively, and we regard the two similarity convex combination as the two sentences similarity. Furthermore, this paper propose the concept differentiability which can measure the efficiency of algorithm for computing the sentence similarity, and the concept of support is also proposed to measure the importance of a FE for a special semantic frame. Experimental results show that the precision rate of this approach has increased 15% compared with the approach based on HowNet.This paper proposes a method to identify and extract core words of frame elements on the basis of multi-word chunk theory. By comparative analyzing, we establish the strategy that integrates the multi-word chunk and FE, then the rules system for extracting core words of frame elements based on multi-word chunk labeling is established. The experimental results on 6771 FEs show that the average precision and average coverage are 95.58% and 82.91%. Accordingly, the precision and recall of sentence similarity computation have increased 1.33% and 2.2% with FKDG than FSDG.