Details

Autor(en) / Beteiligte

• Schrementi, Giancarlo

Titel

Language in the balance: Factors in the emergence of compositional communication

Ort / Verlag

ProQuest Dissertations Publishing

Erscheinungsjahr

2011

Link zum Volltext

• ProQuest

Quelle

ProQuest Dissertations & Theses A&I

Beschreibungen/Notizen

• One of the hallmarks of human language is that it is compositional. By creating novel combinations of morphemes we are able to generalize our communications to an infinite range of circumstances. Many animals communicate but only human language has this property. The question this thesis investigates is what conditions are sufficient for compositional communication to emerge. The thesis evaluates models of proto-linguistic communication for their insights on the factors that lead to the emergence of compositional communication. Two factors in language development are found to play a crucial role: a drive towards simpler grammars and utterances, and the necessity of bi-directional mappings between meaning and signal. Building on these models, a new model is proposed that strictly focuses on these two factors while eliminating any other mechanisms for the emergence of generalization. The agent in this model engages in a naming task in which it utters signals to represent meanings in its world and attempts to comprehend those signals back into the correct meaning. The agent's grammar is represented using a form of finite-state transducer. Using Rissanen's Minimum Description Length Principle (MDL) the two factors are balanced and unified into a fitness metric that rewards compact grammars and successful comprehension. A genetic algorithm is used to search over possible grammars to maximize this metric. An alternative fitness metric based on a Bayesian interpretation of MDL that has potential to weight the factors is also proposed. Generalization is measured using a test set of which only a portion is available for training in the search process. Monte Carlo simulations of the genetic algorithm find that both the MDL-based metric and the Bayesian metric are too susceptible to local optima for generalization to emerge. However, a weighted Bayesian metric that prioritizes bi-directionality shows a robust ability to drive the emergence of generalization across a variety of scenarios. The simulation results establish a new minimum foundation for the emergence of compositional communication. No cognitive mechanisms are necessary beyond balancing the development of bi-directional mappings with a desire for simplicity. However, this balance is not even but tilted in favor of bi-directional mappings.