DS-2018-06: Signaling under Uncertainty

DS-2018-06: Brochhagen, Thomas (2018) Signaling under Uncertainty. Doctoral thesis, University of Amsterdam.

[thumbnail of Full Text] Text (Full Text)
DS-2018-06.text.pdf

Download (3MB)
[thumbnail of Samenvatting] Text (Samenvatting)
DS-2018-06.samenvatting.txt

Download (3kB)

Abstract

What is conveyed often goes beyond what is said. Rather than avoiding it, natural communication seems to thrive in the implicit; in the unsaid; in the contextually determined. This investigation centers around this issue by asking why and under which conditions language (use) may come to leverage or accommodate the unsaid when matters could be conveyed more explicitly. More precisely, at a fundamental level, we seek to better understand why there is a division of labor between semantics and pragmatics. We do so by looking at the conditions under which properties that draw from this division arise, which we analyze by combining, in novel ways, game-theoretic models of rational language use, reinforcement learning, (iterated) Bayesian learning, and population dynamics such as the replicator-mutator dynamic.

Our analysis traces linguistic change at the level of iterated interactions as well as at that of populations. Both levels come with their own perspective and thereby shed their own light on a given linguistic property. This allows us to explore different yet connected answers to questions such as why natural communication is rife with semantic ambiguity; under which conditions systematic pragmatic inferences may (fail to) lexicalize; and, more generally, what kinds of divisions of labor between semantics and pragmatics we can expect to arise from pressures and environmental factors that shape language.

At the level of iterated interactions, we analyze the deliberate use of ambiguous expressions in dialog. With previous explanations of ambiguity, we argue that context plays an important role in allowing for the safe exploitation of ambiguity. However, we inject some wrinkles into this explanation by calling into question and giving up the assumption that interlocutors have access to the same contextual information. This issue unravels into a larger one, where the interplay between context, interlocutors' subjective contextual expectations, and their beliefs about each other's expectations play an important role. We argue that the joint outcome of these factors determines the conditions under which a functional advantage for ambiguity crystallizes. We propose a model of rational language use and couple it with simple adaptive dynamics to capture these ideas, and show that it succeeds in predicting empirically attested patterns of ambiguous language use.

At the population level, properties that draw from interactions at the semantics-pragmatics interface face challenges not only in language use, but also in their faithful transmission across generations. Neither semantics nor pragmatics are directly observable. Learners instead only witness the behavior in which their combination results. This raises an issue because different divisions could result in (almost) indistinguishable overt linguistic behavior. More precisely, we ask why and when regular pragmatic inferences do (not) lexicalize, and when semantic underspecification is either maintained or gives way to more precise expressions. To address these questions, we formulate a model of the (co-)evolution of semantics and pragmatics. This model tracks the effects of functional pressure toward efficient information transfer and the effects of pressure for learnability on separate as well as on combined evolutionary trajectories. We combine this model with individual-level models of pragmatic language use, and couch it in different environments of language use and learning.

Item Type: Thesis (Doctoral)
Report Nr: DS-2018-06
Series Name: ILLC Dissertation (DS) Series
Year: 2018
Subjects: Language
Depositing User: Dr Marco Vervoort
Date Deposited: 14 Jun 2022 15:17
Last Modified: 14 Jun 2022 15:17
URI: https://eprints.illc.uva.nl/id/eprint/2155

Actions (login required)

View Item View Item