Text Mining for the Extraction of Domain Relevant Terms and Term collocations

The domain adaptation capability of information extraction (IE) systems relies on automatic acquisition of domain specific knowledge. The domain specific knowledge contains domain relevant terms, semantic relations for ontology building, or lexicosyntactic patterns for template filling [Riloff & Jones 1999 and Yangarber et al 2000]. Recently, an ever-growing interest in automatic term extraction methods in NLP [Church & Hanks 1989, Smadja 1994, Daille 1996 and Evert & Krenn 2001] has been observed. In this paper, we present an approach to automatic acquisition of singleword terms, multi-word terms and collocations by taking classified documents as input. A word in our approach corresponds to a token unit after the text tokenization. A single-word term is a term consisting of a single word, whereas a multi-word term normally consists of more than one word. By collocations, we consider combinations of words that are not only lexically determined but also semantically related words. Our method is based on the integration of term classification methods and statistical measures for word association. It exhibits that very good results may be achieved on training corpora of different sizes. In particular, we can handle free word-order languages like German using special term collocation techniques. Thus all combinations of elements in a collocation candidate are allowed instead of using a window of predefined size.