Constructional language processing and learning starting from unsegmented linguistic input

Abstract

Constructionist theories of language acquisition claim that all linguistic knowledge can be captured in form-meaning mappings, which can vary in size and degree of abstraction and can contain information from all levels of linguistic analysis. Evidence from child language acquisition suggests that this linguistic knowledge is learned from situated, communicative interactions, where children observe unsegmented, continuous speech acts. However, current computational operationalisations of construction grammar do not fully corroborate these theoretical claims. Specifically, it is difficult to combine constructions of varying levels of granularity in the same constructional analysis and it is not feasible to learn computational construction grammars from unsegmented input. Both issues result from treating segmentation and language processing as separate steps, while in children these processes are intertwined. In response, we introduce two novel algorithms in this paper: one for the processing of constructions in which the linguistic forms are represented as unsegmented character sequences, and one for the learning of constructions starting from unsegmented linguistic forms. The novelty lies in operating directly on character sequences, thereby removing the need for a pre-segmentation step prior to linguistic processing. We operationalise these two algorithms within the framework of Fluid Construction Grammar and illustrate their application through several examples. Through these novel algorithms, we aim to offer greater flexibility to construction grammarians for implementing constructional analyses as well as to pave the way for experiments where the segmentation of continuous input is learned jointly with the constructions. In this way, we bring the computational implementation of construction grammar closer to its theoretical foundations.