Frame-shifting and Conceptual Blending: What do large language models have to say?
summary
Argues that EEG/ERP data (N400, P600) from 160 joke stimuli (subjects see 80 jokes, 80 straight endings) align with language-model surprisal measures: GPT-3, BERT, RoBERTa and ALBERT explain N400 variance beyond human cloze and GPT-3 even outperforms cloze on N400 prediction. However, jokes still produce residual N400/P600 effects that surprisal and contextual dissimilarity don’t fully explain, and models miss many joke inferences and fail comprehension checks. Tools: EEG, single-trial regressions, mixed-effects models, modern LLMs.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:04Semantics and symbol grounding
- 4:27Cognitive semantics models
- 7:26Case study of conceptual blending
- 12:47Frame shifting in humor
- 18:16Measuring brain activity
- 28:13Experimental analysis of jokes
- 38:07Predictive coding and LLMs
- 46:06Refining brain response models
- 53:04Future of language modeling
related talks
40mThe Future of Language Models
51mTowards a Pragmatic Understanding of Large Language Models
30mLarge Language Models and Prompt Engineering for Developers
54mKeynote: Can you trust your (large language) model?
44mAI Intuition: Exploring Language Model Latent Space
57mNatural Language Processing 2020 Discussion
01h 00mInside GPT – Large Language Models Demystified
40mIs my Large Language Model a Strange Loop?