Fast Recursive Multi-class Classification of Pairs of Text Entities for Biomedical Event Extraction

Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 692–701, Gothenburg, Sweden, April 26-30 2014. c⃝2014 Association for Computational Linguistics Fast Recursive Multi-class Classification of Pairs of Text Entities for Biomedical Event Extraction Xiao Liu, Antoine Bordes, Yves Grandvalet Université de Technologie de Compiègne & CNRS Heudiasyc UMR 7253 60200 Compiègne Cedex, France firstname.lastname@hds.utc.fr Abstract Biomedical event extraction from arti- cles has become a popular research topic driven by important applications, such as the automatic update of dedicated knowl- edge base. Most existing approaches are either pipeline models of specific classi- fiers, usually subject to cascading errors, or joint structured models, more efficient but also more costly and more involved to train. We propose here a system based on a pairwise model that transforms event ex- traction into a simple multi-class problem of classifying pairs of text entities. Such pairs are recursively provided to the classi- fier, so as to extract events involving other events as arguments. Our model is more direct than the usual pipeline approaches, and speeds up inference compared to joint models. We report here the best results reported so far on the BioNLP 2011 and 2013 Genia tasks. 1 Introduction Huge amounts of biomedical documents, such as molecular biology reports or genomic papers are generated daily. Automatically organizing their content in dedicated databases enables ad- vanced search and eases information retrieval for researchers in biology, medicine or other related fields. Nowadays, these data sources are mostly in the form of unstructured free text, which is complex to incorporate into databases. Hence, many text-mining research initiatives are orga- nized around the issue of automatically extract- ing information from biomedical text. Efforts specifically dedicated to biomedical text are nec- essary because standard Natural Language Pro- cessing tools cannot be readily applied to extract biomedical events since such texts, articles or re- ports involve highly domain-specific jargon, syn- tax and dependencies (Kim et al., 2011a). This paper tackles the problem of event extrac- tion from biomedical documents. Building on pre- vious advances in named entity recognition (for detecting gene or protein mentions for instance), this task consists in associating to these entities the related events expressed in the text. Such events are of multiple types and involve at least one text entity as argument and another one as trigger; they can be quite complex since some events have sev- eral arguments, and recursive in the sense that ar- guments can themselves be events. An example of event is given in Figure 1. Biomedical event extraction is attracting more and more attention, especially thanks to the or- ganization of recurrent dedicated BioNLP chal- lenges (Kim et al., 2009; Kim et al., 2011b; Kim et al., 2013). We propose here a new approach which relies on a single multi-class classifier for recursively detecting events from (trigger, argu- ment) pairs. Compared to standard pipeline ap- proaches based on sequences of classifiers (Björne and Salakoski, 2013; Hakala et al., 2013), we avoid the intermediate problem of …

Content truncated to 500 words