Bayesian Inference Semantics: A Modelling System and A Test Suite

Proceedings of the Eighth Joint Conference on Lexical and Computational Semantics (*SEM), pages 263–272 Minneapolis, June 6–7, 2019. c⃝2019 Association for Computational Linguistics 263 Bayesian Inference Semantics: A Modelling System and A Test Suite Jean-Philippe Bernardy Rasmus Blanck Stergios Chatzikyriakidis Shalom Lappin Aleksandre Maskharashvili University of Gothenburg firstname.lastname@gu.se Abstract We present BIS, a Bayesian Inference Seman- tics, for probabilistic reasoning in natural lan- guage. The current system is based on the framework of Bernardy et al. (2018), but de- parts from it in important respects. BIS makes use of Bayesian learning for inferring a hy- pothesis from premises. This involves estimat- ing the probability of the hypothesis, given the data supplied by the premises of an argument. It uses a syntactic parser to generate typed syn- tactic structures that serve as input to a model generation system. Sentences are interpreted compositionally to probabilistic programs, and the corresponding truth values are estimated using sampling methods. BIS successfully deals with various probabilistic semantic phe- nomena, including frequency adverbs, gener- alised quantifiers, generics, and vague predi- cates. It performs well on a number of interest- ing probabilistic reasoning tasks. It also sus- tains most classically valid inferences (instan- tiation, de Morgan’s laws, etc.). To test BIS we have built an experimental test suite with examples of a range of probabilistic and clas- sical inference patterns. 1 Introduction On a traditional view of inference, the entailment relation between the premises of an argument and its conclusion holds iff the argument is logically valid in a proof or model theory. More recently, computational approaches to entailment in natu- ral text, such as Recognising Textual Entailment (RTE, Dagan et al. (2009)) have attempted to cap- ture inferences that depend on lexical meaning and real world knowledge, as well as logical structure. In the latter sorts of inference, the conclusions of- ten follow from the premises within a certain range of probability values. In this paper we present Bayesian Inference Se- mantics (BIS), a probabilistic semantics for natu- ral language that assigns probability values, rather than Boolean truth-values, to sentences. The prob- ability of a sentence is the likelihood that an idealised speaker, as represented by our model, would accept the assertion that it expresses. Our framework builds on the approach proposed by Bernardy et al. (2018). It is Bayesian in that it con- structs models in which asserted constraints pro- vide Bayesian evidence that models use to deter- mine whether objects satisfy particular properties. Objects are represented as vectors in a model space S, and properties are subspaces in S. Sat- isfaction of a property is expressed as mem- bership in the corresponding subspace of S. The probability density over the space of pos- sible situations corresponds to the a priori density of objects in these subspaces, and is specified through Bayesian priors. The sys- tem leverages the probabilistic functional pro- gramming language WebPPL (Goodman and Stuhlm¨uller, 2014) to evaluate Bayesian posteri- ors. English sentences are parsed using Ranta’s Grammatical Framework (GF, http://www. grammaticalframework.org/, 2004), and …

Content truncated to 500 words