Measuring Semantic Relatedness with Knowledge Association Network

Measuring semantic relatedness between two words is a fundamental task for many applications in both databases and natural language processing domains. Conventional methods mainly utilize the latent semantic information hidden in lexical databases (WordNet) or text corpus (Wikipedia). They have made great achievements based on the distance computation in lexical tree or co-occurrence principle in Wikipedia. However these methods suffer from low coverage and low precision because (1) lexical database contains abundant lexical information but lacks semantic information; (2) in Wikipedia, two related words (e.g. synonyms) may not appear in a window size or a sentence, and unrelated ones may be mentioned together by chance. To compute semantic relatedness more accurately, some other approaches have made great efforts based on free association network and achieved a significant improvement on relatedness measurement. Nevertheless, they need complex preprocessing in Wikipedia. Besides, the fixed score functions they adopt cause the lack of flexibility and expressiveness of model. In this paper, we leverage DBPedia and Wikipedia to construct a Knowledge Association Network (KAN) which avoids the information extraction of Wikipedia. We propose a flexible and expressive model to represent entities behind the words, in which attribute and topological structure information of entities are embedded in vector space simultaneously. The experiment results based on standard datasets show the better effectiveness of our model compared to previous models.