Lexical Substitution for the Medical Domain

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 610–614, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Lexical Substitution for the Medical Domain Martin Riedl1 Michael R. Glass2 Alfio Gliozzo2 (1) FG Language Technology, CS Dept., TU Darmstadt, 64289 Darmstadt, Germany (2) IBM T.J. Watson Research, Yorktown Heights, NY 10598, USA riedl@cs.tu-darmstadt.de, {mrglass,gliozzo}@us.ibm.com Abstract In this paper we examine the lexical substitu- tion task for the medical domain. We adapt the current best system from the open domain, which trains a single classifier for all instances using delexicalized features. We show sig- nificant improvements over a strong baseline coming from a distributional thesaurus (DT). Whereas in the open domain system, features derived from WordNet show only slight im- provements, we show that its counterpart for the medical domain (UMLS) shows a signif- icant additional benefit when used for feature generation. 1 Introduction The task of lexical substitution (McCarthy and Navigli, 2009) deals with the substitution of a target term within a sentence with words having the same meaning. Thus, the task divides into two subtasks: • Identification of substitution candidates, i.e. terms that are, for some contexts, substitutable for a given target term. • Ranking the substitution candidates according to their context Such a substitution system can help for semantic text similarity (B¨ar et al., 2012), textual entailment (Dagan et al., 2013) or plagiarism detection (Chong and Specia, 2011). Datasets provided by McCarthy and Navigli (2009) and Biemann (2012) offer manually annotated substi- tutes for a given set of target words within a context (sentence). Contrary to these two datasets in Kremer et al. (2014) a dataset is offered where all words have are annotated with substitutes. All the datasets are suited for the open domain. But a system performing lexical substitution is not only of interest for the open domain, but also for the medical domain. Such a system could then be applied to medical word sense disambiguation, entailment or question answering tasks. Here we introduce a new dataset and adapt the lexical substitution system, pro- vided by Szarvas et al. (2013), to the medical domain. Additionally, we do not make use of WordNet (Miller, 1995) to provide similar terms, but rather employ a Dis- tributional Thesaurus (DT), computed on medical texts. 2 Related Work For the general domain, the lexical substitution task was initiated by a Semeval-2007 Task (McCarthy and Navigli, 2009). This task was won by an unsupervised method (Giuliano et al., 2007), which uses WordNet for the substitution candidate generation and then relies on the Google Web1T n-grams (Brants and Franz, 2006)1 to rank the substitutes. The currently best system, to our knowledge, is pro- posed by Szarvas et al. (2013). This is a supervised ap- proach, where a single classifier is trained using delex- icalized features for all substitutes and can thus be ap- plied even to previously unseen substitutes. Although there have been many approaches for solving the task for the general domain, only slight effort …

Content truncated to 500 words