A Topic Description Model Based on Two-Layer Kl Distance

Abstract

The main challenge of Topic Detection and Tracking (TDT) for Blog is the insufficient information in a topic description and the lack of key words input by users. We propose a Two-layer KL Distance approach which combines the KL distance model with a lexical semantic association matrix model. First, the KL Distance model captured the weights of Initial feature words. Second, the KL Distance model was used again to estimate weights of words linked with initial feature words in the lexical Semantic Association Matrix. Extensive experiments show the advantages of our method over the baselines as well as the effectiveness of the two-layer of KL Distance.