UiO1: Constituent-Based Discriminative Ranking for Negation Resolution

First Joint Conference on Lexical and Computational Semantics (*SEM), pages 310–318, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics UiO1: Constituent-Based Discriminative Ranking for Negation Resolution Jonathon Read Erik Velldal Lilja Øvrelid Stephan Oepen University of Oslo, Department of Informatics {jread,erikve,liljao,oe}@ifi.uio.no Abstract This paper describes the first of two systems submitted from the University of Oslo (UiO) to the 2012 *SEM Shared Task on resolving negation. Our submission is an adaption of the negation system of Velldal et al. (2012), which combines SVM cue classification with SVM-based ranking of syntactic constituents for scope resolution. The approach further ex- tends our prior work in that we also identify factual negated events. While submitted for the closed track, the system was the top per- former in the shared task overall. 1 Introduction The First Joint Conference on Lexical and Compu- tational Semantics (*SEM 2012) hosts a shared task on resolving negation (Morante and Blanco, 2012). This involves the subtasks of (i) identifying nega- tion cues, (ii) identifying the in-sentence scope of these cues, and (iii) identifying negated (and factual) events. This paper describes a system submitted by the Language Technology Group at the University of Oslo (UiO). Our starting point is the negation system developed by Velldal et al. (2012) for the domain of biomedical texts, an SVM-based system for classi- fying cues and ranking syntactic constituents to re- solve cue scopes. However, we extend and adapt this system in several important respects, such as in terms of the underlying linguistic formalisms that are used, the textual domain, handling of morpho- logical cues and discontinuous scopes, and in that the current system also identifies negated events. The data sets used for the shared task include the following, all based on negation-annotated Co- nan Doyle (CD) stories (Morante and Daelemans, 2012): a training set of 3644 sentences (hereafter referred to as CDT), a development set of 787 sen- tences (CDD), and a held-out evaluation set of 1089 sentences (CDE). We will refer to the combination of CDT and CDD as CDTD. An example of an an- notated sentence is shown in (1) below, where the cue is marked in bold, the scope is underlined, and the event marked in italics. (1) There was no answer. We describe two different system configurations, both of which were submitted for the closed track (hence we can only make use of the data provided by the task organizers). The systems only differ with respect to how they were optimized. In the first configuration, (hereafter I), all components in the pipeline had their parameters tuned by 10-fold cross-validation across CDTD. The second config- uration (II) is tuned against CDD using CDT for training. The rationale for this strategy is to guard against possible overfitting effects that could result from either optimization scheme, given the limited size of the data sets. For the held-out testing all mod- els are estimated on the entire CDTD. Unless otherwise noted, all reported scores are generated using the evaluation script provided …

Content truncated to 500 words