Development of Language Information System and Systematic Extraction of Latent Knowledge, Based on Graph Computation of Semantic Network

点击翻译按钮获取中文摘要

We developed a new method of solving the cluster-size imbalance problem observed when documents and corpora are processed with MCL. The Branching MCL (BMCL) or the latent adjacency matrix can resize overly inclusive Markov clusters (core clusters) into appropriate subsets. This method is applied to a semantic network built from the large-scale corpus of Gakken's Large Dictionary of Japanese (GLDJ), covering 100,000 words, definitions, examples, and grammatical explanations. The effectiveness of these techniques is currently being tested by creating a clustered semantic network for the GLDJ.As the applications of the graph clustering to the field of Humanities, we made the semantic networks from the lexical co-occurrence data of some historical documents or novels : the books of two contemporary thinkers, Cabanis and Mesmer, to measure the similarity of thinking between them ; the very famous novel of Saint-Exupery, "Le petit prince" to objectively propose a method of word sense disambiguation applicable to his enigmatic word usage. In this study we proposed as an alternative to the keyword-based clustering a new windowing method called Incrementally Advancing Window (TAW) that generates co-occurring word pairs that can be used as inputs to the Incremental Routing Algorithm. The results of the MCL applied to co-occurrence and/or adjacency data matrices were evaluated by using the indexes as weighted curvature, modularity Q and F measure.