A new paper argues that Surprisal Theory, which posits that human processing difficulty of language is directly related to its surprisal within a language model, is a tautology. The author contends that without additional constraints, any measure of difficulty can be fitted to some language model, rendering the theory unfalsifiable. This ambiguity, previously masked by the assumption that corpus-generated models predict human behavior, has been challenged by recent findings showing that improved corpus models can worsen predictions of processing difficulty. The paper suggests that to make Surprisal Theory falsifiable, the relevant language model must be derived from non-empirical considerations of the comprehender, such as memory or processing goals, rather than solely from behavioral data. AI
IMPACT Challenges the theoretical underpinnings of how language models are evaluated against human processing.
RANK_REASON Academic paper published on arXiv discussing linguistic theory.
Read on Hugging Face Daily Papers →
- alphaXiv
- arXiv
- Bibliographic Explorer
- CatalyzeX Code Finder for Papers
- Connected Papers
- CORE Recommender
- DagsHub
- Gotit.pub
- Hugging Face
- Influence Flower
- Litmaps
- ScienceCast
- scite Smart Citations
- Surprisal Theory
- Corpus
- language model
- psycholinguistics
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
