Bi-directional LSTM with character and dependency embedding based approach for bio-molecular event trigger extraction.

Extraction of bio-molecular event from bio-medical literature is very complex task. Bio-molecular event is nothing but a change in state of bio-molecules like proteins, genes etc. As bio-medical event is expressed with trigger word and its arguments, it needs to identify trigger words from data and then find out arguments of trigger words. Bio-medical data contains many ambiguous trigger words and can communicate multiple meanings in different settings. In order to clarify the meanings of such confusing trigger phrases, we employ a supervised technique. In this paper we propose a Bi-directional LSTM (Bi-LSTM) for extracting trigger words from biomedical text. These trigger words are most useful information in bio-molecular event expression. To extract the bio-molecular trigger words, we use a combination of three embedding techniques which are word embedding, character embedding and dependency relation embedding. For dependency embedding we identify dependency relation between current word and nearest protein within a sentence and a dictionary of dependency relations is formed for dependency embedding purpose. Our experiments on of BioNLP-2011 GENIA datasets for event extraction produced 72.54% recall, 75.14% precision and 73.74% F-score in detection of event triggers. [ABSTRACT FROM AUTHOR]