UWashington: Negation Resolution using Machine Learning Methods

First Joint Conference on Lexical and Computational Semantics (*SEM), pages 335–339, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics UWashington: Negation Resolution using Machine Learning Methods James Paul White University of Washington Department of Linguistics, Box 354340 Seattle, WA 98195, USA jimwhite@uw.edu Abstract This paper reports on a simple system for resolving the scope of negation in the closed track of the *SEM 2012 Shared Task. Cue detection is performed using regular expression rules extracted from the training data. Both scope tokens and negated event tokens are resolved using a Conditional Random Field (CRF) sequence tagger – namely the SimpleTagger library in the MALLET machine learning toolkit. The full negation F1 score obtained for the task evaluation is 48.09% (P=74.02%, R=35.61%) which ranks this system fourth among the six submitted for the closed track. 1 Introduction Resolving the scope of negation is an interesting area of research for Natural Language Processing (NLP) systems because many such systems have used methods that are insensitive to polarity. As a result it is fairly common to have a system that treats “X does Y” and “X does not Y” as having the same, or very nearly the same, meaning1. A few application areas that have been addressing this issue recently are in sentiment analysis, bio­ medical NLP, and recognition of textual entail­ ment. Sentiment analysis systems are frequently used in corporate and product marketing, call cen­ ter quality control, and within “recommender” sys­ tems which are all contexts where it is important to recognize that “X does like Y” is contrary to “X does not like Y”. Similarly in biomedical text such 1A one token difference between the strings surely indicating at least an inexact match. as research papers and abstracts, diagnostic proce­ dure reports, and medical records it is important to differentiate between statements about what is the case and what is not the case. The *SEM 2012 Shared Task is actually two re­ lated tasks run in parallel. The one this system was developed for is the identification of three features of negation: the cue, the scope, and the factual negated event (if any). The other task is concerned with the focus of negation. Detailed description of both subtasks, including definition of the relevant concepts and terminology (negation, cue, scope, event, and focus) appears in this volume (Morante and Blanco, 2012). Roser Morante and Eduardo Blanco describe the corpora provided to partici­ pants with numbers and examples, methods used used to process the data, and briefly describes each participant and analyzes the overall results. Annotation of the corpus was undertaken at the University of Antwerp and was performed on sev­ eral Sherlock Holmes works of fiction written by Sir Arthur Conan Doyle. The corpus includes all sentences from the original text, not just those em­ ploying negation. Roser Morante and Walter Daelemans provide a thorough explanation of those gold annotations of negation cue, scope, and negated event (if any) (Morante and Daelemans, 2012). Their paper explains the motivations for …

Content truncated to 500 words