Learning to Predict Denotational Probabilities For Modeling Entailment

Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, pages 721–730, Valencia, Spain, April 3-7, 2017. c⃝2017 Association for Computational Linguistics Learning to Predict Denotational Probabilities For Modeling Entailment Alice Lai and Julia Hockenmaier Department of Computer Science University of Illinois at Urbana-Champaign {aylai2, juliahmr}@illinois.edu Abstract We propose a framework that captures the denotational probabilities of words and phrases by embedding them in a vector space, and present a method to induce such an embedding from a dataset of denota- tional probabilities. We show that our model successfully predicts denotational probabilities for unseen phrases, and that its predictions are useful for textual entail- ment datasets such as SICK and SNLI. 1 Introduction In order to bridge the gap between vector-based distributional approaches to lexical semantics that are intended to capture which words occur in sim- ilar contexts, and logic-based approaches to com- positional semantics that are intended to capture the truth conditions under which statements hold, Young et al. (2014) introduced the concept of “denotational similarity.” Denotational similarity is intended to measure the similarity of simple, declarative statements in terms of the similarity of their truth conditions. From classical truth-conditional semantics, Young et al. borrowed the notion of the deno- tation of a declarative sentence s, �s�, as the set of possible worlds in which the sentence is true. Young et al. apply this concept to the domain of image descriptions by defining the visual denota- tion of a sentence s as the set of images that s de- scribes. The denotational probability of s, P��(s), is the number of images in the visual denotation of s over the size of the corpus. Two sentences are denotationally similar if the sets of images (possi- ble worlds) they describe have a large overlap. For example, “A woman is jogging on a beach” and “A woman is running on a sandy shore” can of- ten be used to describe the same scenario, so they will have a large image overlap that corresponds to high denotational similarity. Given the above definitions, Young et al. es- timate the denotational probabilities of phrases from FLICKR30K, a corpus of 30,000 images, each paired with five descriptive captions. Young et al. (2014) and Lai and Hockenmaier (2014) showed that these similarities are complementary to standard distributional similarities, and poten- tially more useful for semantic tasks that involve entailment. However, the systems presented in these papers were restricted to looking up the de- notational similarities of frequent phrases in the training data. In this paper, we go beyond this prior work and define a model that can predict the denotational probabilities of novel phrases and sentences. Our experimental results indicate that these predicted denotational probabilities are use- ful for several textual entailment datasets. 2 Textual entailment in SICK and SNLI The goal of textual entailment is to predict whether a hypothesis sentence is true, false, or neither based on the premise text (Dagan et al., 2013). Due in …

Content truncated to 500 words