Towards a Framework for Comparing Automatic Term Recognition Methods

Automatic Term Recognition focuses on the extraction of words and multi-word expressions that are significant for a given domain. There is a considerable interest in using ATR for automatic metadata generation, creation of thesauri and terminological glossaries, keyword extraction, ontology building, etc. In this paper, we build upon the work done at the University of Sheffield, where a library with a few algorithms for ATR was recently developed. We enrich this library with new ATR algorithms and tools for evaluation. Our aim is to perform an experimental study comparing the base ATR methods as well as their combinations under various conditions. The results of the study indicate that better precision can be usually reached by combining ATR methods using foreground and ATR methods using background knowledge. The created platform is freely available and prepared for extensions by other researchers.