Using Seme Based Graph to Estimate Chinese Lexical Semantic Relatedness

A robust numerical measures of lexical relatedness is significant for many applications, such as text summarization system and information retrieval researches. Standard seme-based measures of word pair relatedness are based on only the comparison of semes of the two words. This paper propose a new model called seme based graph using an extended random walk to measure explicit and implicit relatedness between two semes. Comparing to traditional random graph walk model, our model uses average encounter probability instead of average arrival probability and avoid an ambiguous divergence measure method. Then, this paper presents an dynamic programming algorithm to compute the relatedness scores. According to our experiments, the model has following advantages: 1) relatedness scores is determined highly congruous with the perspective of human cognition. 2)high correlated with human judgments at ρ=0.8. 3)very fast and low cost.