Computational Modeling of the Segmentation of Sentence Stimuli From an Infant Word-Finding Study

Area of Science:

• Cognitive Science
• Developmental Psychology
• Computational Linguistics

Background:

• Infant word segmentation models traditionally use speech transcriptions.
• Recent advancements allow direct analysis of auditory speech stimuli.
• Evaluating models against infant behavioral data is crucial for benchmarking.

Purpose of the Study:

• To propose and demonstrate a method for testing computational word segmentation models on actual auditory stimuli from infant experiments.
• To benchmark computational models against infant learning outcomes in psycholinguistic studies.
• To enhance the accuracy and relevance of computational models for infant language acquisition.

Main Methods:

• Applied the DP-Parser computational model to auditory stimuli from infant psycholinguistic experiments.
• DP-Parser processes raw speech, generating overlapping embeddings to identify potential words.
• Utilized dynamic programming to segment utterances based on maximizing constituent segment frequency.

Main Results:

• The DP-Parser model successfully mimicked American English learners' word segmentation performance on Italian sentences.
• The model favored segmenting words with high syllabic transitional probability, aligning with human learning patterns.
• Demonstrated the feasibility of computational analysis directly on experimental speech stimuli.

Conclusions:

• Computational models analyzing actual speech stimuli provide a more accurate benchmark for infant word segmentation.
• This methodology can refine models to better reflect human infant language acquisition.
• Future research should leverage direct speech analysis for developing more effective computational models of infant learning.