The semantics-syntax interface: Learning grammatical categories and hierarchical syntactic structure through semantics.

Abstract

Language is infinitely productive because syntax defines dependencies between grammatical categories of words and constituents, so there is interchangeability of these words and constituents within syntactic structures. Previous laboratory-based studies of language learning have shown that complex language structures like hierarchical center embeddings (HCE) are very hard to learn, but these studies tend to simplify the language learning task, omitting semantics and focusing either on learning dependencies between individual words or on acquiring the category membership of those words. We tested whether categories of words and dependencies between these categories and between constituents, could be learned simultaneously in an artificial language with HCE's, when accompanied by scenes illustrating the sentence's intended meaning. Across four experiments, we showed that participants were able to learn the HCE language varying words across categories and category-dependencies, and constituents across constituents-dependencies. They also were able to generalize the learned structure to novel sentences and novel scenes that they had not previously experienced. This simultaneous learning resulting in a productive complex language system, may be a consequence of grounding complex syntax acquisition in semantics.

Syntax through rapid synaptic changes

“…Presently, it does not reflect the hierarchical nature of sentence structures. This means that, to encode sentences like “the dog that chases the cat runs”, the current model needs to duplicate all its syntactic rules for the subordinate clause ( 8 ), and the model’s linear order will not explain rules of agreement ( 9 ). Nor is the model able to simultaneously assign the same word to two different roles, in the example “the little star is beside a big star”.…”

“…We hypothesize that working memory links each word to its role in the sentence, and those roles follow probabilistic orderings. Although this cannot account for true grammar ( 8 ), since it does not support hierarchical structure ( 9 ), it provides a biologically plausible way to understand some aspects of syntax. In this paper we propose a neural architecture that stores and generates syntactic sequences.…”