Parsing Biomedical Literature

Parsing Biomedical Literature⋆ Matthew Lease and Eugene Charniak Brown Laboratory for Linguistic Information Processing (BLLIP), Brown University, Providence, RI USA {mlease, ec}@cs.brown.edu Abstract. We present a preliminary study of several parser adaptation techniques evaluated on the GENIA corpus of MEDLINE abstracts [1,2]. We begin by observing that the Penn Treebank (PTB) is lexically im- poverished when measured on various genres of scientific and techni- cal writing, and that this significantly impacts parse accuracy. To re- solve this without requiring in-domain treebank data, we show how ex- isting domain-specific lexical resources may be leveraged to augment PTB-training: part-of-speech tags, dictionary collocations, and named- entities. Using a state-of-the-art statistical parser [3] as our baseline, our lexically-adapted parser achieves a 14.2% reduction in error. With oracle- knowledge of named-entities, this error reduction improves to 21.2%. 1 Introduction Since the advent of the Penn Treebank (PTB) [4], statistical approaches to nat- ural language parsing have quickly matured [3,5]. By providing a very large corpus of manually labeled parsing examples, PTB has played an invaluable role in enabling the broad analysis, automatic training, and quantitative evalu- ation of parsing techniques. However, while PTB’s Wall Street Journal (WSJ) corpus has historically served as the canonical benchmark for evaluating statis- tical parsing, the need for broader evaluation has been increasingly recognized in recent years. Furthermore, since it is impractical to create a large treebank like PTB for every genre of interest, significant attention has been directed to- wards maximally reusing existing training data in order to mitigate the need for domain-specific training examples. These issues have been most notably ex- plored in parser adaptation studies conducted between PTB’s WSJ and Brown corpora [6,7,8,9]. As part of our own exploration of these issues, we have been investigating statistical parser adaptation to a novel domain: biomedical literature. This lit- erature presents a stark contrast to WSJ and Brown: it is suffused with domain- specific vocabulary, has markedly different stylistic constraints, and is often writ- ten by non-native speakers. Moreover, broader consideration of technical litera- ture shows this challenge and opportunity is not confined to biomedical literature ⋆ We would like to thank the National Science Foundation for their support of this work (IIS-0112432, LIS-9721276, and DMS-0074276), as well as thank Sharon Goldwater and our anonymous reviewers for their valuable feeback. R. Dale et al. (Eds.): IJCNLP 2005, LNAI 3651, pp. 58–69, 2005. c ⃝ Springer-Verlag Berlin Heidelberg 2005 Parsing Biomedical Literature 59 alone, but is also demonstrated by patent literature, engineering manuals, and field-specific scientific discourse. Through our work with biomedical literature, we hope to gain insights into effective techniques for adapting statistical parsing to technical literature in general. Our interest in biomedical literature is also motivated by a real need to im- prove information extraction in this domain. With over 15 million citations in PubMed today, biomedical literature is the largest and fastest growing knowl- edge domain of any science. As such, simply managing the sheer volume of its accumulated information has become a significant problem. In response …

Content truncated to 500 words