Exploiting the Scope of Negations and Heterogeneous Features for Relation Extraction: A Case Study for Drug-Drug Interaction Extraction

Proceedings of NAACL-HLT 2013, pages 765–771, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Exploiting the Scope of Negations and Heterogeneous Features for Relation Extraction: A Case Study for Drug-Drug Interaction Extraction Md. Faisal Mahbub Chowdhury † ‡ and Alberto Lavelli ‡ ‡ Fondazione Bruno Kessler (FBK-irst), Italy † University of Trento, Italy fmchowdhury@gmail.com, lavelli@fbk.eu Abstract This paper presents an approach that exploits the scope of negation cues for relation extrac- tion (RE) without the need of using any specif- ically annotated dataset for building a separate negation scope detection classifier. New fea- tures are proposed which are used in two dif- ferent stages. These also include non-target entity specific features. The proposed RE ap- proach outperforms the previous state of the art for drug-drug interaction (DDI) extraction. 1 Introduction Negation is a linguistic phenomenon where a nega- tion cue (e.g. not) can alter the meaning of a partic- ular text segment or of a fact. This text segment (or fact) is said to be inside the scope of that negation (cue). In the context of RE, there is not much work that aims to exploit the scope of negations.1 The only work on RE that we are aware of is Sanchez- Graillet and Poesio (2007) where they used various heuristics to extract negative protein interaction. Despite the recent interest on automatically de- tecting the scope of negation2 till now there seems to be no empirical evidence supporting its exploita- tion for the purpose of RE. Even if we could man- age to obtain highly accurate automatically detected 1In the context of event extraction (a closely related task of RE), there have been efforts in BioNLP shared tasks of 2009 and 2011 for (non-mandatory sub-task of) event negation detection (3 participants in 2009; 2 in 2011) (Kim et al., 2009; Kim et al., 2011). The participants approached the sub-task using either pre-defined patterns or some heuristics. 2This task is popularized by various recently held shared tasks (Farkas et al., 2010; Morante and Blanco, 2012). negation scopes, it is not clear how to feed this infor- mation inside the RE approach. Simply considering whether a pair of candidate mentions falls under the scope of a negation cue might not be helpful. In this paper, we propose that the scope of nega- tions can be exploited at two different levels. Firstly, the system would check whether all the target en- tity3 mentions inside a sentence along with possible relation clues (or trigger words), if any, fall (directly or indirectly) under the scope of a negation cue. If such a sentence is found, then it should be discarded (i.e. candidate mention pairs4 inside that sentence would not be considered). Secondly, for each of the remaining pairs of candidate mentions, the system should exploit features related to the scope of nega- tion (rather than simply adding a feature for negation cue, approach adopted in various RE systems) that can provide indication (if any such evidence exists) that the corresponding relation of interest …

Content truncated to 500 words