Multi-word complex concept retrieval via lexical semantic similarity

This paper first presents a simple computational means of measuring universal object similarity that is based on classical feature-based similarity models. This computational model is implemented with the help of semantic network representations (e.g. WordNet taxonomy) and corpus statistics. It is then extended and applied to a higher level and practical information retrieval task-retrieving multi-word complex concepts. The extension is performed by pair-wise comparison of all decomposed sub-concepts or terms in a query and the texts, trying different schemes for combining averaging and maximization of the pair-wise similarities. Series of experiments are conducted to compare it with classic statistical methods and the results are supportive of our work.