UiO 2: Sequence-labeling Negation Using Dependency Features

First Joint Conference on Lexical and Computational Semantics (*SEM), pages 319–327, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics UiO2: Sequence-Labeling Negation Using Dependency Features Emanuele Lapponi Erik Velldal Lilja Øvrelid University of Oslo, Department of InformaticsJonathon Read {emanuel,erikve,liljao,jread}@ifi.uio.no Abstract This paper describes the second of two sys- tems submitted from the University of Oslo (UiO) to the 2012 *SEM Shared Task on re- solving negation. The system combines SVM cue classification with CRF sequence labeling of events and scopes. Models for scopes and events are created using lexical and syntactic features, together with a fine-grained set of la- bels that capture the scopal behavior of certain tokens. Following labeling, negated tokens are assigned to their respective cues using simple post-processing heuristics. The system was ranked first in the open track and third in the closed track, and was one of the top perform- ers in the scope resolution sub-task overall. 1 Introduction Negation Resolution (NR) is the task of determin- ing, for a given sentence, which tokens are affected by a negation cue. The data set most prominently used for the development of systems for automatic NR is the BioScope Corpus (Vincze et al., 2008), a collection of clinical reports and papers in the bio- medical domain annotated with negation and specu- lation cues and their scopes. The data sets released in conjunction with the 2012 shared task on NR hosted by The First Joint Conference on Lexical and Computational Semantics (*SEM 2012) are com- prised of the following negation annotated stories of Conan Doyle (CD): a training set of 3644 sentences drawn from The Hound of the Baskervilles (CDT), a development set of 787 sentences taken from Wis- teria Lodge (CDD; we will refer to the combina- tion of CDT and CDD as CDTD), and a held-out test set of 1089 sentences from The Cardboard Box and The Red Circle (CDE). In these sets, the con- cept of negation scope extends on the one adopted in the BioScope corpus in several aspects: Nega- tion cues are not part of the scope, morphological (affixal) cues are annotated and scopes can be dis- continuous. Moreover, in-scope states or events are marked as negated if they are factual and presented as events that did not happen (Morante and Daele- mans, 2012). Examples (1) and (2) below are exam- ples of affixal negation and discontinuous scope re- spectively: The cues are bold, the tokens contained within their scopes are underlined and the negated event is italicized. (1) Since we have been so unfortunate as to miss him [. . . ] (2) If he was in the hospital and yet not on the staff he could only have been a house-surgeon or a house-physician: lit- tle more than a senior student. Example (2) has no negated events because the sen- tence is non-factual. The *SEM shared task thus comprises three sub- tasks: cue identification, scope resolution and event detection. It is furthermore divided into two separate tracks: one closed track, …

Content truncated to 500 words