Large language models can segment narrative events similarly to humans

Area of Science:

• Cognitive Science
• Computational Linguistics
• Artificial Intelligence

Background:

• Human event perception involves segmenting continuous experience into discrete occurrences.
• Quantifying event boundaries typically relies on aggregating multiple human annotations.
• Automated methods are needed to efficiently analyze event structures in large datasets.

Purpose of the Study:

• To investigate the feasibility of using a large language model (GPT-3) for automated event boundary detection in narrative text.
• To compare GPT-3's event annotations with human-derived annotations.
• To assess GPT-3's potential as a tool for studying human event perception.

Main Methods:

• GPT-3 was employed to segment continuous narrative text into distinct events.
• Event boundaries identified by GPT-3 were statistically correlated with those identified by human observers.
• GPT-3's annotations were compared against a consensus solution derived from averaging human annotations.

Main Results:

• GPT-3 successfully segmented narrative text into meaningful events.
• GPT-3-derived event annotations showed significant correlation with human annotations.
• GPT-3 identified event boundaries that closely approximated the human consensus, outperforming individual human annotators on average.

Conclusions:

• Large language models like GPT-3 offer a viable computational approach for automated event annotation.
• GPT-3's performance in event segmentation suggests parallels between its predictive capabilities and human cognitive processes.
• Automated event annotation using GPT-3 could accelerate research into the principles of human event perception.