Adaptive Strategies in the Emergence of Lexical Systems

Student thesis: Doctoral Thesis

Abstract

This thesis can be situated in a new scientific paradigm that investigates the emergence and evolution of linguistic conventions through evolutionary language games. My contributions are mainly related to lexical conventions and the gradual development of cognitive structures.
Two approaches to lexical acquisition exist in the literature. The first posits a strong division between concept formation and lexical learning. Concepts arise independently of language and the lexical acquisition problem can be reduced to a problem of mapping words to concepts. In the third chapter I present an overview of the cross-situational learning (XSL) approaches that have been successfully applied to this problem in the past.
The second view of lexical learning sees lexical acquisition as a process of creating and gradually shaping the meanings of words. In the fourth chapter I show that traditional XSL is not well suited for this more difficult task and that this second view of word learning needs more powerful processing and learning strategies.
I argue that, surprisingly, we should look at theories of grammar, such as cognitive grammar and grammaticalization theory, for inspiration. In chapter five I propose a strategy that is in line with the basic tenets of these theories and which addresses three shortcomings of the traditional XSL approach. First it no longer relies on an explicit enumeration of competing hypotheses, only a single hypothesis per word is maintained at all time. Second, the manner in which word meanings are used in processing stresses flexible re-use. Third, word meanings are internally adapted based on the feedback from flexible processing. I show that strategies embodying these tenets cope much more naturally with the problem of lexical acquisition and conventionalization when meanings have not been established beforehand.
The final part of the thesis ventures beyond the lexical domain toward grammatical categorization. The main research question adressed is whether the adaptive lexical learning algorithms proposed in the earlier chapters can be extended to also deal with the problem of semantic grammatical categorization and the new conventionalization problems this task brings along.
An overarching contribution, which motivated primarily the earlier chapters of the thesis, is to offer an overview of non-grounded lexical language games and show the progress that has been made over the past fifteen years. A substantial selection of strategies has been reimplemented so that, for ?the first time, the strategies are explained and compared within the same multi-agent setting, using the same data and the same measures.
Date of Award24 Feb 2012
Original languageEnglish
SupervisorLuc Steels (Promotor), Ann Nowe (Jury), Bart De Boer (Jury), Tom Lenaerts (Jury), Paul Vogt (Jury) & Andrew Smith (Jury)

Keywords

  • artificial intelligence
  • language
  • concept
  • lexicon
  • adaptive

Cite this

'