Machine Learning of Computational Construction Grammars

The natural languages that underlie human communication are not static and innate, but emerge and evolve through the communicative interactions of interlocutors. The emergence and evolution of linguistic systems is driven by evolutionary processes, in particular variation, selection, self-organisation and level formation. These evolutionary processes take place within the linguistic system itself (rather than in the genes of the interlocutors), on the levels of concepts, words, grammar and discourse. The main selective pressures are success in communication and a reduction of cognitive effort, while sources of variation mostly stem from the creativity and problem solving capabilities of the language users. It is one of the main goals of the research fields of emergent communication and evolutionary linguistics to precisely model the evolutionary processes that allow a population of communicative agents to establish such a conventionalised linguistic system.

The evolutionary pressures involved in the emergence and evolution of concepts and words are relatively well understood, while the processes driving the emergence and evolution of grammatical systems are still to a large extent uncharted territory. In my PhD project, I will contribute to this understanding by investigating how basic grammatical structures can emerge and evolve through task-based communicative interactions in a tutor-learner scenario. More concretely, I will design and implement learning operators that allow an autonomous agent to learn basic grammatical structures as operationalised by computational construction grammar, i.e. in the form of conventionalised form-meaning mappings, which can range from being entirely idiomatic to being completely abstract. The resulting grammars are transparent in the sense that they are explainable in human terms, and well-adapted to the communicative needs of its users, as they have been shaped by past successes and failures in communication.

The capacity for an autonomous agent to learn computational construction grammars has a potential impact in many application domains, ranging from human-robot interaction, through personal assistants, to intelligent language tutoring systems.