Finding Contradictions in Text

Proceedings of ACL-08: HLT, pages 1039–1047, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics Finding Contradictions in Text Marie-Catherine de Marneffe, Linguistics Department Stanford University Stanford, CA 94305 mcdm@stanford.edu Anna N. Rafferty and Christopher D. Manning Computer Science Department Stanford University Stanford, CA 94305 {rafferty,manning}@stanford.edu Abstract Detecting conflicting statements is a foun- dational text understanding task with appli- cations in information analysis. We pro- pose an appropriate definition of contradiction for NLP tasks and develop available corpora, from which we construct a typology of con- tradictions. We demonstrate that a system for contradiction needs to make more fine-grained distinctions than the common systems for en- tailment. In particular, we argue for the cen- trality of event coreference and therefore in- corporate such a component based on topical- ity. We present the first detailed breakdown of performance on this task. Detecting some types of contradiction requires deeper inferen- tial paths than our system is capable of, but we achieve good performance on types arising from negation and antonymy. 1 Introduction In this paper, we seek to understand the ways con- tradictions occur across texts and describe a system for automatically detecting such constructions. As a foundational task in text understanding (Condoravdi et al., 2003), contradiction detection has many possi- ble applications. Consider applying a contradiction detection system to political candidate debates: by drawing attention to topics in which candidates have conflicting positions, the system could enable voters to make more informed choices between candidates and sift through the amount of available informa- tion. Contradiction detection could also be applied to intelligence reports, demonstrating which infor- mation may need further verification. In bioinfor- matics where protein-protein interaction is widely studied, automatically finding conflicting facts about such interactions would be beneficial. Here, we shed light on the complex picture of con- tradiction in text. We provide a definition of contra- diction suitable for NLP tasks, as well as a collec- tion of contradiction corpora. Analyzing these data, we find contradiction is a rare phenomenon that may be created in different ways; we propose a typol- ogy of contradiction classes and tabulate their fre- quencies. Contradictions arise from relatively obvi- ous features such as antonymy, negation, or numeric mismatches. They also arise from complex differ- ences in the structure of assertions, discrepancies based on world-knowledge, and lexical contrasts. (1) Police specializing in explosives defused the rock- ets. Some 100 people were working inside the plant. (2) 100 people were injured. This pair is contradictory: defused rockets cannot go off, and thus cannot injure anyone. Detecting con- tradictions appears to be a harder task than detecting entailments. Here, it is relatively easy to identify the lack of entailment: the first sentence involves no in- juries, so the second is unlikely to be entailed. Most entailment systems function as weak proof theory (Hickl et al., 2006; MacCartney et al., 2006; Zan- zotto et al., 2007), but contradictions require deeper inferences and model building. While mismatch- ing information between sentences is often a good cue …

Content truncated to 500 words