Finding Contradictions in Text

In this paper, I seek to understand the ways contradictions occur across texts and I describe a system for automatically detecting such constructions. Finding conflicting statements is foundational for text understanding, a problem which recently received a surge of interest in the computational linguistics community. Condoravdi et al. (2003) first recognized the importance of handling both entailment and contradiction for text understanding: “relations of entailment and contradiction are the key data of semantics, as traditionally viewed as a branch of linguistics. The ability to recognize such semantic relations is clearly not a sufficient criterion for language understanding: there is more than just being able to tell that one sentence follows from another. But we would argue that it is a minimal, necessary criterion.” (p. 38). So far however, work in robust text understanding has focused on entailment: systems aimed at providing textual inference in arbitrary domains. The task of textual inference first appeared latent within the field of question answering (Pasca & Harabagiu 2001; Moldovan et al. 2003). As schematized in Figure 1, the question What company sells most greetings cards? can be viewed as a statement containing a variable (what company) which in this case is of the organization type. If the system finds a text passage that entails the statement and contains a possible assignment for the variable (mainly a concept of the same type), the variable assignment is taken as the answer to the question. In this example, the passage Hallmark remains the largest maker of greetings cards entails the question, Hallmark is of the type organization, and will be the answer to the question. The task of textual inference then received attention within the PASCAL Recognizing Textual Entailment (RTE) Challenges (Dagan et al. 2006; Bar-Haim et al. 2006; Giampiccolo et al. 2007), and related work within the U.S. Government AQUAINT program. In the RTE challenges, systems are given pairs of sentences, called text (T) and hypothesis (H), and the goal is to identify whether the hypothesis follows from the text and general background knowledge, according to the intuitions of an intelligent human reader. That is, the standard is not whether the hypothesis is logically entailed, but whether it can reasonably be inferred: “We say that T entails H if the meaning of H can be inferred from the meaning of T, as would typically be interpreted by people. This somewhat informal definition is based on (and assumes) common human understanding of language as well as common background knowledge.” (Dagan et al. 2005, p. 1).