Generalisation and Specialisation Operators for Computational Construction Grammar and their Application in Evolutionary Linguistics Research

Abstract

The natural languages that underlie human communication are remarkably expressive, robust and well-adapted to the communicative needs of their users. However, the question of how these languages have emerged and through which mechanisms they continue to evolve remains heavily debated. A common methodology for studying this question is to simulate the emergence and evolution of language using agent-based models. In these models, a population of autonomous agents, which are either physical robots or software entities, participates in a series of communicative interactions, known as language games. Each game is played by two agents in the population, one being the speaker and the other being the hearer. The game involves a scripted, communicative task, which either succeeds or fails. At the end of the game, the speaker provides feedback to the hearer, so that learning can take place. The goal of the models is to determine the exact mechanisms that need to be present in the individual agents, so that a communication system with human language-like properties can emerge and evolve. While agent-based models have within the language game paradigm most extensively been used to study concept learning and vocabulary formation, they have more recently also been successfully applied to experiments on the emergence and evolution of grammar. In these models, the agents need to be equipped with a computational grammar formalism that supports robust and flexible language processing, including mechanisms for inventing and adopting grammatical structures. This dissertation presents three major contributions to the field of research that studies the modelling of the emergence and evolution of grammar. The first contribution consists in the implementation of a new, higher-level notation for Fluid Construction Grammar (FCG). FCG is an advanced computational grammar formalism that is often used in evolutionary linguistics experiments. The new notation represents grammatical structures in a more intuitive way and abstracts away from low-level implementation details. This facilitates the use of FCG in language evolution experiments and the new notation has indeed already become FCG's standard notation. The second contribution introduces powerful mechanisms for generalising and specialising grammatical constructions. The impasse that arises when agents are faced with utterances that they cannot process can often be overcome by adapting constraints that block the application of existing grammatical constructions. Previous experiments relied on ad hoc ways to detect and adapt these constraints. Here, I extend FCG with three general mechanisms: (i) an anti-unification based operator that finds the blocking constraints and their least general generalisations, (ii) a hierarchical type system that can capture these generalisations is a fine-grained way, and (iii) a pro-unification operator that imposes additional constraints on a construction, specialising it to specific cases. The third contribution consists in a case study that demonstrates how the representations and mechanisms introduced above can be incorporated in an actual agent-based experiment. The experiment that I present here studies how early syntactic structures can emerge and evolve in a population of agents. In particular, it models how shared word order patterns can come into place and reduce the referential ambiguity of the language. The experiment makes use of the type hierarchy system to capture the association strength between words and slots in the word order patterns, and relies on the anti-unification operator to expand the coverage of existing patterns to novel words. The experiment shows that a coherent and efficient word order system rapidly emerges in a population of agents that are equipped with these general, local mechanisms.

Cite this work

No public copy is stored for this record, so only the metadata is cited.

Stored only when both papers are in the corpus — 11 citations exist upstream.