Invited paper presented at a symposium on New Directions in Statistical Learning at the 2002 International Conference on Infant Studies, Toronto, Canada.
When acquiring the syntactic structure of their native language children face a difficult "bootstrapping" problem. Discovering syntactic constraints requires knowledge about grammatical categories; but, conversely, grammatical categories are only useful for acquisition insofar as they support syntactic constraints. Recent work in developmental psycholinguistics suggests that children may solve the bootstrapping problem by exploiting distributional, phonological, and prosodic cues. However, these cues are probabilistic and individually unreliable. The simulations presented in this talk demonstrate how a statistical learning device attuned to phonological and prosodic cues can overcome some of the limitations of a simple distributional learning, resulting in significant improvements in syntactic acquisition.
A phrase-structure grammar was constructed based on previous research and independent corpus analyses. This fairly complex grammar was intended to capture the general syntactic trends prevalent in child-directed speech, including declarative, imperative, and interrogative sentences. Five groups of simple recurrent networks were trained on sentences generated from this grammar. Each group of networks integrated a different constellation of three probabilistic cues: word length (in syllables), stress, and pitch. A baseline network relied entirely on distributional information, with cues added to the remaining groups, including a three-cue network that took advantage of all three cues.
The multiple-cue integration networks all performed better than corresponding trigram models — a standard computational linguistics benchmark for distributional syntax learning. Multiple-cue integration also resulted in faster and more uniform learning when the performance of baseline networks was compared with the cue-augmented networks. Additional simulations demonstrated that the presence of distractor cues uncorrelated with syntactic structure does not prevent the networks from successfully integrating the partially reliable cues. These simulations (and others) underscore the computational feasibility of multiple-cue approach to the acquisition of syntax, and, more generally, the advantage of going beyond simple distributional information to statistical learning through the integration of multiple probabilistic information sources.
Please email suggestions/errors to mhc27@cornell.edu |