A Connectionist Simulation of the Empirical Acquisition of Grammatical Relations

This paper proposes an account of the acquisition of grammatical relations using the basic concepts of connectionism and a construction-based theory of grammar. Many previous accounts of first-language acquisition assume that grammatical relations (e.g., the grammatical subject and object of a sentence) and linking rules are universal and innate; this is necessary to provide a first set of assumptions in the target language to allow deductive processes to test hypotheses and/or set parameters. In contrast to this approach, we propose that grammatical relations emerge rather late in the language-learning process. Our theoretical proposal is based on two observations. First, early production of childhood speech is formulaic and becomes systematic in a progressive fashion. Second, grammatical relations themselves are family-resemblance categories that cannot be described by a single parameter. This leads to the notion that grammatical relations are learned in a bottom up fashion. Combining this theoretical position with the notion that the main purpose of language is communication, we demonstrate the emergence of the notion of “subject” in a simple recurrent network that learns to map from sentences to semantic roles. We analyze the hidden layer representations of the emergent subject, and demonstrate that these representations correspond to a radially-structured category. We also claim that the pattern of generalization and undergeneralization demonstrated by the network conforms to what we expect from the data on children’s generalizations.