Computing Natural Language Inference

Workshop Description

Central within a theory of formal semantics for Natural Language (NL) is the study of Natural Language Inference (NLI). Roughly put, NLI is the task of determining whether an NL hypothesis can be inferred from an NL premise. Human beings do not only have the ability to understand infinitely many NL sentences but can also reason about these. In effect, understanding a NL sentence amounts (among other things) to knowing what can and cannot be inferred from such a sentence. It is thus no surprise that NLI is a central field of computational semantics. As Cooper et al. (1996) aptly put it, `inferential ability is not only a central manifestation of semantic competence but is in fact centrally constitutive of it'. Inferential ability, according to Cooper et al., is the best way to test the semantic adequacy of NLP systems. However, the three main datasets that have been designed to be used as test suites for such NLP systems are inadequate. The FraCas test suite (Cooper et al. 1996) offers good coverage in terms of inference types at the cost of unnatural data (data are constructed) and small size. The RTE platforms (Dagan et al. 2005) offer naturally occurring data at the cost of semantic coverage, while the latest SNLI platform (Bowman et al 2015), offers a huge crowdsourced platform but does not capture generic inferences, given that it is always dependent on the specific situation described by the caption. In general, it seems that existing platforms are limited to certain aspects of inference, and fail to represent the full complexity of NLI. Furthermore, dialogue -- arguably one of the most central aspects of human communication -- is left outside in all three cases. Finally, with the creation of large image and language corpora, textual inference may be related to visual representations that the text is referring to: both in terms of identifying a complement to textual underspecification and in identification of perceptual features as types of situations licensing certain kinds of inferences.

This workshop aims to create a forum where the issue of constructing NLI platforms will be discussed from both a theoretical and implementational point of view. We are further aiming at bringing together people from different computational semantics backgrounds, but who, at the same time, share the goal of creating a platform that is useful regardless the approach one takes w.r.t NLI. We invite 2-4 page abstract submissions (references included) on theoretical and implementational aspects of NLI. We specifically encourage submissions on the construction and development of state of the art platforms for NLI.

The workshop will take place one day before the main conference (19th of September 2017) in Montpellier. It is free for the main conference attendees. For more information see the conference main site