Bootstrapping the lexicon: a computational model of infant speech segmentation

Area of Science:

• Cognitive Science
• Computational Linguistics
• Developmental Psychology

Background:

• Prelinguistic infants face the challenge of segmenting continuous speech streams into meaningful units.
• Understanding this process is crucial for theories of language acquisition and cognitive development.

Purpose of the Study:

• To introduce BootLex, a novel computational model for infant speech segmentation using distributional cues.
• To evaluate BootLex's performance across diverse language corpora and speech types.
• To compare BootLex with existing models, focusing on functional characteristics and cognitive plausibility.

Main Methods:

• Developed BootLex, a probabilistic algorithm leveraging distributional information to build a lexicon.
• Tested BootLex on English, Japanese, and Spanish corpora, including child-directed and adult-directed speech, and written texts.
• Quantitatively and qualitatively compared BootLex against three other computational models of infant segmentation.

Main Results:

• BootLex achieved significant speech segmentation results across various language and speech data.
• The model identified specific statistical input characteristics influencing segmentation performance.
• Comparisons revealed BootLex's functional characteristics and its similarity to human cognitive processes.

Conclusions:

• Distributional cues alone provide a powerful basis for infant speech segmentation.
• BootLex offers insights into the cognitive mechanisms underlying early language learning.
• The study contributes to computational cognitive modeling and theories of speech segmentation.