Collocation Extraction Using Web Statistics

Proceedings of the Fourth International Conference on Language Resources and Evaluation (LREC 2004)

Abstract

This paper mines collocations from two different web usage corpora, NTU proxy log and TTS search log. The precisions for NTU and TTS test data are 61.76% and 57.50%, respectively, by human judgment for 2% sampling of extracted collocations. For automatic evaluation, we submit extracted collocation to Google search engine, and the resulting page counts are used to compute the mutual information of the collocation. Experimental results show that total 43.27% and 42.65% of collocations mined from NTU and TTS corpora passed the examination of MIs.