Evaluating models of robust word recognition with serial reproduction

Area of Science:

• Cognitive Science
• Computational Linguistics
• Psycholinguistics

Background:

• Spoken communication is inherently noisy, with environmental interference, speaker variability, and linguistic ambiguity.
• Effective spoken word recognition and language processing depend on listeners' prior knowledge to resolve uncertainty.
• Probabilistic generative language models offer a framework for understanding how prior knowledge influences language comprehension.

Purpose of the Study:

• To compare the effectiveness of various probabilistic generative language models in capturing human linguistic expectations.
• To investigate how abstract representations of linguistic context influence spoken language processing and transmission.
• To identify which language model features best predict changes in spoken utterances during serial reproduction.

Main Methods:

• Utilized the serial reproduction paradigm, akin to the "Telephone" game, to collect spoken utterances from successive English-speaking adult participants.
• Evaluated a suite of broad-coverage probabilistic generative language models against the collected utterance chains.
• Employed a logistic regression model to pinpoint words susceptible to loss or alteration during spoken transmission.

Main Results:

• Language models incorporating abstract representations of linguistic context, specifically phrase structure, demonstrated superior predictive accuracy for human-generated changes in spoken utterances.
• The findings indicate that phrase structure is a key factor in how listeners anticipate and process spoken language.
• The logistic regression analysis confirmed the predictive power of contextual representations in understanding word alterations during transmission.

Conclusions:

• Probabilistic generative language models that leverage abstract linguistic context, such as phrase structure, are more effective at modeling human linguistic expectations.
• Understanding the role of prior knowledge and contextual representations is crucial for explaining robust spoken language processing in noisy conditions.
• These findings contribute to a deeper understanding of the interplay between memory constraints and linguistic representations in language comprehension and transmission.