Bayesian Surprise Predicts Human Event Segmentation in Story Listening

Area of Science:

• Cognitive psychology
• Computational neuroscience
• Natural language processing

Background:

• Event segmentation theory proposes that large prediction errors mark transitions between discrete experiences.
• Understanding the neural and computational mechanisms of event segmentation is crucial for cognitive science.

Purpose of the Study:

• To test event segmentation theory during story listening using computational models.
• To investigate the relationship between prediction error metrics and human-identified event boundaries.

Main Methods:

• Utilized GPT-2 language model to calculate word-level prediction error time series for three stories.
• Collected human participants' event boundary markings while they listened to the same stories.
• Employed regression models to correlate computational prediction error measures with human segmentation data.

Main Results:

• Event boundaries significantly correlated with transient increases in Bayesian surprise.
• A simpler prediction error measure, surprisal, did not show a significant association with event boundaries.
• Demonstrated a distinction between different operational definitions of prediction error in cognitive segmentation.

Conclusions:

• Findings support the role of prediction error as a key mechanism in event segmentation.
• Bayesian surprise emerges as a more relevant computational correlate of human event boundary detection than surprisal.
• Highlights the importance of precise definitions and measurements of prediction error in cognitive research.