UMichigan: A Conditional Random Field Model for Resolving the Scope of Negation

26290

Authors

nan

Abstract

In this paper, we present a system for de- tecting negation in English text. We address three tasks: negation cue detection, negation scope resolution and negated event identifi- cation. We pose these tasks as sequence la- beling problems. For each task, we train a Conditional Random Field (CRF) model on lexical, structural, and syntactic features ex- tracted from labeled data. The models are trained and tested using the dataset distributed with the *sem Shared Task 2012 on resolving the scope and focus of negation. The system detects negation cues with 90.98% F1 mea- sure (94.3% and 87.88% recall). It identifies negation scope with 82.70% F1 on token-by- token level and 64.78% F1 on full scope level. Negated events are detected with 51.10% F1 measure. 1

Keywords

nan

Full Content

First Joint Conference on Lexical and Computational Semantics (*SEM), pages 328–334, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics UMichigan: A Conditional Random Field Model for Resolving the Scope of Negation Amjad Abu-Jbara EECS Department University of Michigan Ann Arbor, MI, USA amjbara@umich.edu Dragomir Radev EECS Department University of Michigan Ann Arbor, MI, USA radev@umich.edu Abstract In this paper, we present a system for de- tecting negation in English text. We address three tasks: negation cue detection, negation scope resolution and negated event identifi- cation. We pose these tasks as sequence la- beling problems. For each task, we train a Conditional Random Field (CRF) model on lexical, structural, and syntactic features ex- tracted from labeled data. The models are trained and tested using the dataset distributed with the *sem Shared Task 2012 on resolving the scope and focus of negation. The system detects negation cues with 90.98% F1 mea- sure (94.3% and 87.88% recall). It identifies negation scope with 82.70% F1 on token-by- token level and 64.78% F1 on full scope level. Negated events are detected with 51.10% F1 measure. 1 Introduction Negation is a linguistic phenomenon present in all languages (Tottie, 1993; Horn, 1989). The seman- tic function of negation is to transform an affirma- tive statement into its opposite meaning. The auto- matic detection of negation and its scope is a prob- lem encountered in a wide range of natural language processing applications including, but not limited to, data mining, relation extraction, question answering, and sentiment analysis. For example, failing to ac- count for negation may result in giving wrong an- swers in question answering systems or in the pre- diction of opposite sentiment in sentiment analysis systems. The occurrence of negation in a sentence is deter- mined by the presence of a negation cue. A nega- tion cue is a word, a phrase, a prefix, or a postfix that triggers negation. Scope of negation is the part of the meaning that is negated (Huddleston and Pul- lum, 2002). The negated event is the event or the en- tity that the negation indicates its absence or denies its occurrence. For example, in the sentence below never is the negation cue. The scope is enclosed in square brackets. The negated event is underlined. [Andrew had] never [liked smart phones], but he received one as a gift last week and started to use it. Negation cues and scopes may be discontinuous. For example, the negation cue neither ... nor is dis- continuous. In this chapter, we present a system for automat- ically detecting negation cues, negated events, and negation scopes in English text. The system uses conditional random field (CRF) models trained on labeled sentences extracted from two classical En- glish novels. The CRF models are trained using lex- ical, structural, and syntactic features. The experi- ments show promising results. This paper is organized as follows. Section 2 re- views previous work. Section 3 describes the data. Section 4 describes the CRFs models. Section 5 presents evaluation, results, …

Content truncated to 500 words

Labeled Text

In this paper, we present a system for detecting negation in English text. We address three tasks: negation cue detection, negation scope resolution and negated event identification. We pose these tasks as sequence labeling problems. For each task, we train a Conditional Random Field (CRF) model on lexical, structural, and syntactic features extracted from labeled data. The models are trained and tested using the dataset distributed with the *sem Shared Task 2012 on resolving the scope and focus of negation. The system detects negation cues with 90.98% F1 measure (94.3% and 87.88% recall). It identifies negation scope with 82.70% F1 on token-bytoken level and 64.78% F1 on full scope level. Negated events are detected with 51.10% F1 measure. 1

Publication Details

Paper ID:
26290
Year:
2012
Publication:
TIPSTER
Publication Type:
conf
DOI:
nan