Short-Text Conceptualization Based on a Co-ranking Framework via Lexical Knowledge Base

Abstract

The problem of short-text conceptualization is important, and has attracted increasing attention. Recent probabilistic algorithms have demonstrated remarkable successes. However most of them are limited to the assumption that all the observed terms in given short-text are conditionally independent, ignoring the interaction among terms (and concepts), as well as the beneficial reactions from concepts to terms. To overcome these problems, recently some co-rank paradigms are proposed, unfortunately neither they fails to integrate the co-occurrence feature nor they fails to utilize the semantic similarity implicit in the lexical knowledge base. Therefore, previous works could not release robust concept representation. Faced with this problem, this paper proposes a novel framework based on both statistic information (e.g., co-occurrence feature in large-scale corpus) and semantic information (e.g., semantic similarity in lexical knowledge base), for co-ranking terms and their corresponding concepts simultaneously, This co-ranking framework utilizes several graphs: the concept graph, the term graph and the subordination graph. The experimental results show that our method achieves higher accuracy and efficiency in short-text conceptualization than the state-of-the-art algorithms.