LLF – Bât. ODG – 5e étage – Salle du conseil (533)
Quentin Dénigot (LLF)
Formal Approaches to the Communication of Conflicting Identities in Discourse
This dissertation focuses on a proposal for the formal modelling of a very specific phenomenon of political communication usually called ``dogwhistles''. A secondary contribution is an attempt to use (and a reflexive approach to the use of) machine learning techniques for the study of said dogwhistles.
From a theoretical point of view, this work draws inspiration from game theoretic approaches of linguistics and pragmatics, game theoretic approaches of social meaning, and distributional semantics. Beyond the main contributions, the thesis presents briefly each of these areas of research, since it is likely that readers might not know much about at least one of these areas, and I therefore hope that these introductory texts can be also counted among the contributions of this work.
Formal linguistics has devised a number of sophisticated mathematical objects to try and describe linguistic phenomena with as much precision as possible for the last decades. There is however one area of linguistics that has historically evaded attempts at formalization: sociolinguistics. The few attempts at formalizing sociolinguistics results are fairly recent. Inspired by tools from formal pragmatics, they have expanded the concept of linguistic interpretation to the interpretation not only of message content, but of social meaning as well, leading to elaborate formal models that are at the basis of fascinating research programmes.
There has been however few attempts at linking the works from formal pragmatics and semantics -- focusing mostly on the interpretation of content -- and formal sociolinguistics -- focusing on the interpretation of social meaning. Notably, there have been no attempts at giving a picture in formal terms of situations whereby the social meaning contained within an utterance can influence the content interpretation of that utterance by listeners.
In this work, we focus on situations where speakers send messages where the social meaning and content interpretations vary according to who receives them. This phenomenon is sometimes referred to as dogwhistle politics, since it has been mostly studied in the context of political discourse. Using formal tools from pragmatics (Rational Speech Act models) and their adaptations in the realm of sociolinguistics (Social Meaning Games), we aim to construct a model that can accurately describes these situations, where the retrieval of the identity of the speaker can supposedly lead to a variety of interpretations, linking the notions of social meaning and content meaning in a single kind of linguistic game. Beyond this endeavor, a key characteristic of dogwhistles is that they are supposedly hidden messages to a point. This leads to a situation where it can be difficult to assess whether a given utterance does or does not contain a dogwhistle. Acknowledging this, we have tried to give a definition of dogwhistles that goes beyond their mere description and could be applied to the enterprise of finding them in text, using concepts from distributional semantics, vector space models of language and machine learning.