Exploring variation across biomedical subdomains

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 689–697, Beijing, August 2010 Exploring variation across biomedical subdomains Tom Lippincott and Diarmuid ´O S´eaghdha and Lin Sun and Anna Korhonen Computer Laboratory University of Cambridge {tl318,do242,ls418,alk23}@cam.ac.uk Abstract Previous research has demonstrated the importance of handling differences be- tween domains such as “newswire” and “biomedicine” when porting NLP systems from one domain to another. In this paper we identify the related issue of subdomain variation, i.e., differences between subsets of a domain that might be expected to be- have homogeneously. Using a large corpus of research articles, we explore how subdo- mains of biomedicine vary across a variety of linguistic dimensions and discover that there is rich variation. We conclude that an awareness of such variation is necessary when deploying NLP systems for use in single or multiple subdomains. 1 Introduction One of the most noticeable trends in the past decade of Natural Language Processing (NLP) re- search has been the deployment of language pro- cessing technology to meet the information re- trieval and extraction needs of scientists in other disciplines. This meeting of fields has proven mu- tually beneficial: scientists increasingly rely on automated tools to help them cope with the expo- nentially expanding body of publications in their field, while NLP researchers have been spurred to address new conceptual problems in theirs. Among the fundamental advances from the NLP perspec- tive has been the realisation that tools which per- form well on textual data from one source may fail to do so on another unless they are tailored to the new source in some way. This has led to signifi- cant interest in the idea of contrasting domains and the concomitant problem of domain adaptation, as well as the production of manually annotated domain-specific corpora.1 One definition of domain variation associates it with differences in the underlying probability distributions from which different sets of data are drawn (Daum´e III and Marcu, 2006). The concept also mirrors the notion of variation across thematic subjects and the corpus-linguistic notions of reg- ister and genre (Biber, 1988). In addition to the differences in vocabulary that one would expect to observe, domains can vary in many linguistic variables that affect NLP systems. The scientific domain which has received the most attention (and is the focus of this paper) is the biomedical domain. Notable examples of corpus construction projects for the biomedical domain are PennBioIE (Kulick et al., 2004) and GENIA (Kim et al., 2003). These corpora have been used to develop systems for a range of processing tasks, from entity recognition (Jin et al., 2006) to parsing (Hara et al., 2005) to coreference resolution (Nguyen and Kim, 2008). An implicit assumption in much previous work on biomedical NLP has been that particular subdo- mains of biomedical literature – typically molec- ular biology – can be used as a model of biomed- ical language in general. For example, GENIA consists of abstracts dealing with a specific set of subjects in molecular …

Content truncated to 500 words