Comparable Study of Event Extraction in Newswire and Biomedical Domains

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 2270–2279, Dublin, Ireland, August 23-29 2014. Comparable Study of Event Extraction in Newswire and Biomedical Domains Makoto Miwa†,‡ Paul Thompson† Ioannis Korkontzelos† Sophia Ananiadou† †National Centre for Text Mining and School of Computer Science, University of Manchester, United Kingdom ‡Graduate School of Engineering, Toyota Technological Institute, Japan {makoto.miwa, paul.thompson, ioannis.korkontzelos, sophia.ananiadou}@manchester.ac.uk Abstract Event extraction is a popular research topic in natural language processing. Several event extrac- tion tasks have been defined for both the newswire and biomedical domains. In general, different systems have been developed for the two domains, despite the fact that the tasks in both domains share a number of characteristics. In this paper, we analyse the commonalities and differences between the tasks in the two domains. Based on this analysis, we demonstrate how an event extraction method originally designed for the biomedical domain can be adapted for application to the newswire domain. The performance is state-of-the-art for both domains, with F-scores of 52.7% for the biomedical domain and 52.1% for the newswire domain in terms of their primary evaluation metrics. 1 Introduction Research into event extraction was initially focussed on the general language domain, largely driven by the Message Understanding Conferences (MUC) series (e.g., Chinchor (1998)) and the Automated Con- tent Extraction (ACE) evaluations1. More recently, the focus of research has been widened to the biomed- ical domain, motivated by the ongoing series of biomedical natural language processing (BioNLP) shared tasks (STs) (e.g., Kim et al. (2013)). Although the textual characteristics and the types of relevant events to be extracted can vary consid- erably between domains, the same general features of events normally hold across domains. An event usually consists of a trigger and arguments (see Figures 1 and 2.) A trigger is typically a verb or a nom- inalised verb that denotes the presence of the event in the text, while the arguments are usually entities. In general, arguments are assigned semantic roles that characterise their contribution towards the event description. Until now, however, there has been little, if any, effort by researchers working on event extraction in different domains to share ideas and techniques, unlike syntactic tasks (e.g., (Miyao and Tsujii, 2008)) and other information extraction tasks, such as named entity recognition (e.g., (Giuliano et al., 2006)) and relation extraction (e.g., (Qian and Zhou, 2012)). This means that the potential to exploit cross- domain features of events to develop more adaptable event extraction systems is an under-studied area. Consequently, although there is a large number of published studies on event extraction, proposing many different methods, no work has previously been reported that aims to adapt an event extraction method developed for one domain to a new domain. In response to the above, we have investigated the feasibility of adapting an event extraction method developed for the biomedical domain to the newswire domain. To facilitate this, we firstly carry out a detailed static analysis of the differences that hold between event extraction …

Content truncated to 500 words