Study: Humans mimic greedy sampling in constrained language tasks

Researchers explored how humans and computational models produce language under vocabulary constraints, using a limited set of 250 words in some scenarios. They found that human language production generally aligns more with a greedy sampling approach than a globally optimal one, though more skilled individuals exhibit non-greedy revision behaviors. The study also noted that both humans and models tend to rely on semantically lighter words when faced with severe vocabulary limitations, with implications for understanding resource-rational cognition and language impairments. AI

IMPACT Investigates cognitive processes in language generation, offering insights into how LLMs might better mimic human-like communication under constraints.

RANK_REASON Academic paper detailing a study on language production mechanisms. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• humans
• L2 communication
• language production
• large language models
• psycholinguistics
• resource-rational models
• Sequential Monte Carlo
• vocabulary constraints

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →