How much does prosody help word segmentation? A simulation study on infant-directed speech.

Area of Science:

• Cognitive Science
• Developmental Psychology
• Computational Linguistics

Background:

• Infants acquire numerous word forms by age two, a process potentially aided by segmenting continuous speech.
• Prosodic boundaries in speech have been hypothesized to facilitate this word segmentation task.

Purpose of the Study:

• To computationally model infant word segmentation with and without prosodic information.
• To evaluate the impact of expert-annotated versus automatically derived prosodic boundaries on segmentation algorithms.

Main Methods:

• Utilized five cognitively-based word segmentation algorithms (sub-lexical vs. lexical, heuristic vs. ideal learner).
• Applied models to a corpus of infant-directed speech, both with and without prosodic boundary information.
• Compared performance gains when prosodic information was expert-annotated versus automatically detected.

Main Results:

• Prosodic information did not uniformly benefit all segmentation models; sub-lexical algorithms showed the most improvement, while lexical ones showed minimal gain.
• Automatically detected prosodic boundaries, incorporating known infant acoustic sensitivities, yielded smaller positive effects or even negative effects for some algorithms due to detection errors.
• The utility of prosodic breaks for word segmentation is contingent on accurate detection and the specific segmentation strategy employed.

Conclusions:

• While infants may possess the capacity to use prosodic breaks for word segmentation, its practical benefit is not guaranteed.
• Incorporating prosody into word segmentation strategies may not be optimal for infants when faced with realistic, potentially error-prone prosodic signals.
• Future research should consider the reliability of prosodic cues in naturalistic infant-directed speech for word segmentation models.