Bruno Mery

Towards a Type-Theoretical Account of Lexical Semantics

After a quick overview of the field of study known as “Lexical Semantics”, where we advocate the need of accessing additional information besides syntax and Montague-style semantics at the lexical level in order to complete the full analysis of an utterance, we summarize the current formulations of a well-known theory of that field. We then propose and justify our own model of the Generative Lexicon Theory, based upon a variation of classical compositional semantics, and outline its formalization. Additionally, we discuss the theoretical place of informational, knowledge-related data supposed to exist within the lexicon as well as within discourse and other linguistic constructs. The formalization of the structure of natural language utterances around a surface form (phenogrammatics), a deep structure (tectogrammatics) and the meaning thereof as a logical form (semantics) has developed from the original theories of Curry and Montague to form coherent, type-driven models. Most of these new theories rely upon variations of the compositional analysis of the sentence: from pheno to tectogrammatics, and then to semantics. Our contribution to this work aims at giving such a model a means to overcome the problems posed by polysemous lexical units during the semantical analysis of the tectogrammatical form. Building upon an assumed “deep structure”, we formalize parts of Pustejovsky’s Generative Lexicon Theory, linguistically motivated in Pustejovsky (The generative lexicon, MIT Press, Cambridge, MA, 1995), in a pre-processing of the semantics of the sentence. The mechanisms of Lexical Semantics we propose are an additional layer of classical Montague compositional semantics, and, as such, integrate smoothly within such an analysis; we proceed by converting the lexical data to modifiers of the logical form. This treatment of Lexical Semantics furthermore induces us to think that some sort of non-evident background knowledge of the common use of words is necessary to perform a correct semantic analysis of an utterance. This “commonsense metaphysics” would therefore not be strictly confined to pragmatics, as is often assumed.

Challenges in the Computational Implementation of Montagovian Lexical Semantics

We present and discuss a general-purpose implementation of the process of lexical semantics analysis theorised in the Montagovian Generative Lexicon \(\varLambda TY_n\) (hereafter MGL). The prototype software itself constitutes a proof of concept of the MGL theory. The implementation process, as well as the data structures and algorithms, also provide valuable results as to the expressive power required by MGL. While the implementation of terms and types for the purpose of meaning assembly assumed by MGL is in itself straightforward, some lexical phenomena require additional mechanisms in order to process the logical representation in order to take into account implicit common-sense world knowledge. We therefore also present a minimal architecture for knowledge representation, and how it can be applied to different phenomena. The implementation illustrates the validity of the theory, but MGL requires a stronger corpus of types and terms in order to be thoroughly tested.

Collecting Weighted Coercions from Crowd-Sourced Lexical Data for Compositional Semantic Analysis

Type-theoretic frameworks for compositional semantics are aimed at producing structured meaning representations of natural language utterances.