Language models and humans differ in sentence surprise

Researchers have investigated why language models exhibit less surprise than humans when processing ambiguous sentences. They tested the hypothesis that language models can consider more interpretations simultaneously than humans. By adjusting the number of parses used in recurrent neural network grammars, they found that reducing simultaneous parses increased predicted garden path effects, but not enough to match human reading times. This suggests that the difference in parse multiplicity alone does not explain the discrepancy in surprise levels. AI

影响 Investigates fundamental differences in how language models and humans process linguistic ambiguity, potentially informing future model design.

排序理由 Academic paper published on arXiv detailing a hypothesis and experimental results regarding language model behavior. [lever_c_demoted from research: ic=1 ai=1.0]

在 arXiv cs.CL 阅读 →

• humans
• language models
• Recurrent Neural Network Grammars

• 论文
• 其他

AI 生成摘要 · Google Gemini · 来自 1 个来源。 我们如何撰写摘要 →