Adding Context to Semantic Data-Driven Paraphrasing

Proceedings of the Fifth Joint Conference on Lexical and Computational Semantics (*SEM 2016), pages 108–113, Berlin, Germany, August 11-12, 2016. Adding Context to Semantic Data-Driven Paraphrasing Vered Shwartz Ido Dagan Computer Science Department Bar-Ilan University Ramat-Gan, Israel vered1986@gmail.com dagan@cs.biu.ac.il Abstract Recognizing lexical inferences between pairs of terms is a common task in NLP applications, which should typically be performed within a given context. Such context-sensitive inferences have to con- sider both term meaning in context as well as the fine-grained relation hold- ing between the terms. Hence, to de- velop suitable lexical inference methods, we need datasets that are annotated with fine-grained semantic relations in-context. Since existing datasets either provide out- of-context annotations or refer to coarse- grained relations, we propose a method- ology for adding context-sensitive anno- tations. We demonstrate our methodol- ogy by applying it to phrase pairs from PPDB 2.0, creating a novel dataset of fine- grained lexical inferences in-context and showing its utility in developing context- sensitive methods. 1 Introduction Recognizing lexical inference is an essential com- ponent in semantic tasks. In question answering, for instance, identifying that broadcast and air are synonymous enables answering the question “When was ‘Friends’ first aired?” given the text “‘Friends’ was first broadcast in 1994”. Semantic relations such as synonymy (tall, high) and hyper- nymy (cat, pet) are used to infer the meaning of one term from another, in order to overcome lexi- cal variability. In semantic tasks, such terms appear within cor- responding contexts, thus making two aspects nec- essary in order to correctly apply inferences: First, the meaning of each term should be considered within its context (Szpektor et al., 2007; Pantel et al., 2007), e.g., play entails compete in certain contexts, but not in the context of playing the na- tional anthem at a sports competition. Second, the soundness of inferences within context is condi- tioned on the fine-grained semantic relation that holds between the terms, as studied within natural logic (MacCartney and Manning, 2007). For in- stance, in upward-monotone sentences a term en- tails its hypernym (“my iPhone’s battery is low” ⇒ “my phone’s battery is low”), while in down- ward monotone ones it entails its hyponym (“talk- ing on the phone is prohibited” ⇒ “talking on the iPhone is prohibited”). Accordingly, developing algorithms that prop- erly apply lexical inferences in context requires datasets in which inferences are annotated in- context by fine-grained semantic relations. Yet, such a dataset is not available (see 2.1). Most ex- isting datasets provide out-of-context annotations, while the few available in-context annotations re- fer to coarse-grained relations, such as relatedness or similarity. In recent years, the PPDB paraphrase database (Ganitkevitch et al., 2013) became a popular re- source among semantic tasks, such as monolin- gual alignment (Sultan et al., 2014) and recog- nizing textual entailment (Noh et al., 2015). Re- cently, Pavlick et al. (2015) classified each para- phrase pair to the fine-grained semantic relation that holds between the phrases, following natu- ral logic (MacCartney and Manning, 2007). To that end, …

Content truncated to 500 words