Presentations

Abstract

Nearly all human languages have grammars with complex recursive structures. These structures pose notable learning challenges. Two distributional properties of the input may facilitate learning: the presence of semantic biases (e.g. p(barks|dog) > p(talks|dog)) and the Zipf-distribution, with short sentences being extremely more frequent than longer ones. This project tested the effect of these sources of information on statistical learning of a hierarchical center-embedding grammar, using an artificial grammar learning paradigm. Semantic biases were represented by variations in transitional probabilities between words, with a biased input (p(barks|dog) > p(talks|dog)) compared to a non-biased input (p(barks|dog) = p(talks|dog)). The Zipf distribution was compared to a flat distribution, with sentences of different lengths occurring equally often. In a 2×2 factorial design, we tested for effects of biased transitional probabilities (biased/non-biased) and the distribution of sequences with varying length (Zipf distribution/flat distribution) on implicit learning and explicit ratings of grammaticality. Preliminary results show that a Zipf-shaped and semantically biased input facilitates grammar learnability. Thus, this project contributes to understanding how we learn complex structures with long-distance dependencies: learning may be sensitive to the specific distributional properties of the linguistic input, mirroring meaningful aspects of the world and favoring short utterances.