An Example-Based Approach to Difficult Pronoun Resolution

PACLIC 28 !358 An Example-Based Approach to Difficult Pronoun Resolution Canasai Kruengkrai Naoya Inoue Jun Sugiura Kentaro Inui Graduate School of Information Sciences, Tohoku University 6-6 Aramaki Aza Aoba, Aobaku, Sendai, Miyagi 980-8579, Japan {canasai,naoya-i,jun-s,inui}@ecei.tohoku.ac.jp Abstract A Winograd schema is a pair of twin sentences containing a referential ambiguity that is easy for a human to resolve but difficult for a com- puter. This paper explores the characteristics of world knowledge necessary for resolving such a schema. We observe that people tend to avoid ambiguous antecedents when using pro- nouns in writing. We present a method for au- tomatically acquiring examples that are simi- lar to Winograd schemas but have less ambi- guity. We generate a concise search query that captures the essential parts of a given source sentence and then find the alignments of the source sentence and its retrieved examples. Our experimental results show that the exist- ing sentences on the Web indeed contain in- stances of world knowledge useful for difficult pronoun resolution. 1 Introduction Consider the following pair of sentences:1 (1) a. The outlaw shot the sheriff, but he did not shoot the deputy. b. The outlaw shot the sheriff, but he shot back. Suppose that the target pronoun is he, and its two candidate antecedents are the outlaw and the sher- iff. The question is which of the two candidates is the correct antecedent for the target pronoun in each sentence? Most people resolve he to the outlaw in (1a) but to the sheriff in (1b) without noticing any 1The sentences are taken from the dataset created by Rah- man and Ng (2012). ambiguity. However, for a computer program, this pronoun resolution becomes extremely difficult, re- quiring the use of world knowledge and the ability to reason. We refer to the pair of sentences like (1) as a Winograd schema (Levesque, 2011; Levesque et al., 2012). Note that the two sentences differ only in a few words and have a referential ambiguity that is resolved in opposite ways. A previous work by Rahman and Ng (2012) showed that two sources of world knowledge, in- cluding narrative chains (Chambers and Jurafsky, 2008) and page counts returned by a search engine, are useful for resolving Winograd schemas. How- ever, these two knowledge sources have their own weaknesses and need some heuristics to bridge the gap. Narrative chains suffer from the lack of dis- course relations. For example, both sentences in (1) have a contrast relation indicated by but. How- ever, narrative chains rely only on temporal relations between two events (e.g., before and after). Page counts used for estimating n-gram statistics are un- stable and vary considerably over time (Lapata and Keller, 2005; Levesque et al., 2012). Therefore the answer to the question “what kind of world knowl- edge does a computer program need to have to re- solve Winograd schemas?” (Levesque, 2013) is still unclear. Rather than looking for new knowledge bases, we first examine whether existing sentences on the Web have sufficient evidence that could …

Content truncated to 500 words