Learning Structures of Negations from Flat Annotations

Proceedings of the Fourth Joint Conference on Lexical and Computational Semantics (*SEM 2015), pages 71–81, Denver, Colorado, June 4–5, 2015. Learning Structures of Negations from Flat Annotations Vinodkumar Prabhakaran Department of Computer Science Columbia University New York, NY, USA vinod@cs.columbia.edu Branimir Boguraev IBM Watson Thomas J. Watson Research Center Yorktown Heights, NY, USA bran@us.ibm.com Abstract We propose a novel method to learn negation expressions in a specialized (medical) domain. In our corpus, negations are annotated as ‘flat’ text spans. This allows for some infelicities in the mark-up of the ground truth, making it less than perfectly aligned with the underly- ing syntactic structure. Nonetheless, the nega- tions thus captured are correct in intent, and thus potentially valuable. We succeed in train- ing a model for detecting the negated pred- icates corresponding to the annotated nega- tions, by re-mapping the corpus to anchor its ‘flat’ annotation spans into the predicate argu- ment structure. Our key idea—re-mapping the negation instance spans to more uniform syn- tactic nodes—makes it possible to re-frame the learning task as a simpler one, and to lever- age an imperfect resource in a way which en- ables us to learn a high performance model. We achieve high accuracy for negation detec- tion overall, 87%. Our re-mapping scheme can be constructively applied to existing flatly annotated resources for other tasks where syn- tactic context is vital. 1 Introduction Accounting for extra-propositional aspects of mean- ing in text is a very active NLP research area in recent years, exploring different aspects of mean- ing such as factivity (Saur´ı and Pustejovsky, 2009), uncertainty/hedging (Farkas et al., 2010), commit- ted belief (Prabhakaran et al., 2010), and modalities (Prabhakaran et al., 2012a). Among these, negation detection has generated special interest because of demonstrated needs for negation detection capabil- ity in practical applications such as information re- trieval (Averbuch et al., 2004), information extrac- tion (Meystre et al., 2008), sentiment analysis (Wie- gand et al., 2010; Councill et al., 2010), and relation detection (Chowdhury and Lavelli, 2013). Accurately detecting negations is especially im- portant in systems processing medical/clinical text. Consider the segment “Mild hyperinflation without focal pneumonia”, taken from a patient’s clinical record. It indicates the absence of focal pneumonia in the patient. Not capturing this extra-propositional aspect of negation concerning focal pneumonia will lead to wrong—and harmful—inferences in down- stream processing, e.g. by a clinical decision sup- port system. The need for sophisticated negation de- tection capabilities in clinical text is even more ur- gent given the broadening spectrum of applications in this domain: clinical question answering (Lee et al., 2006), clinical decision support (Demner- Fushman et al., 2009), medical information extrac- tion (Uzuner et al., 2010), medical entity relation mining (Tymoshenko et al., 2012), patient history tracking (Raghavan et al., 2012), etc. Our moti- vation for detecting negations in medical texts also stems from practical concerns of an operational medical question answering (QA) system (Ferrucci et al., 2013). Most recent approaches to negation detection adopt supervised machine learning techniques to learn the …

Content truncated to 500 words