Event Frame Extraction Based on a Gene Regulation Corpus

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 761–768 Manchester, August 2008 Event Frame Extraction Based on a Gene Regulation Corpus Yutaka Sasaki 1 Paul Thompson 1 Philip Cotter 1 John McNaught 1, 2 Sophia Ananiadou1, 2 1 School of Computer Science, University of Manchester 2 National Centre for Text Mining MIB, 131 Princess Street, Manchester, M1 7DN, United Kingdom Yutaka.Sasaki@manchester.ac.uk Abstract This paper describes the supervised ac- quisition of semantic event frames based on a corpus of biomedical abstracts, in which the biological process of E. coli gene regulation has been linguistically annotated by a group of biologists in the EC research project "BOOTStrep". Gene regulation is one of the rapidly advancing areas for which information extraction could boost research. Event frames are an essential linguistic resource for extraction of information from biological literature. This paper presents a specification for linguistic-level annotation of gene regu- lation events, followed by novel methods of automatic event frame extraction from text. The event frame extraction per- formance has been evaluated with 10- fold cross validation. The experimental results show that a precision of nearly 50% and a recall of around 20% are achieved. Since the goal of this paper is event frame extraction, rather than event instance extraction, the issue of low re- call could be solved by applying the methods to a larger-scale corpus. 1 Introduction This paper describes the automatic extraction of linguistic event frames based on a corpus of MEDLINE abstracts that has been annotated with gene regulation events by a group of do- © 2008. Licensed under the Creative Commons Attri- bution-Noncommercial-Share Alike 3.0 Unported license (http://creativecommons.org/licenses/by-nc- sa/3.0/). Some rights reserved. main experts. Annotation is centred on both verbs and nominalised verbs that describe rele- vant events. For each event, semantic arguments that occur within the same sentence are marked and labelled with semantic roles and named en- tity (NE) types. The focus of the paper is the extraction of event frames on the basis of the annotated corpus using machine learning techniques. Event frames are linguistic specifications concerning the be- haviour of verbs and nominalised verbs, in terms of the number and types of semantic arguments with which they typically co-occur in texts. Our eventual goal is to exploit such information to improve information extraction. Event frame ex- traction is different to event instance extraction (or template filling). Our event frames are des- tined for incorporation in the BOOTStrep BioLexicon to support identification of relevant event instances and discovery of event instance participants by NLP systems. 2 Background There are several well-established, large-scale repositories of semantic frames for general lan- guage, e.g., VerbNet (Kipper-Schuler, 2005), PropBank (Palmer et al., 2005) and FrameNet (Rupenhoffer et al, 2006). These all aim to char- acterise verb behaviour in terms of the semantic arguments with which verbs occur but differ in how they represent semantic arguments and groupings of verbs. In VerbNet, the semantic roles of arguments come from frame-independent roles, e.g. Agent, Patient, Location and Instrument. …

Content truncated to 500 words