Grammaticality may help organize neural language models’ internal representations
A new preprint’s authors report abstract separation by grammaticality in many neural language models. They believe the work addresses debates over how to measure models’ grammatical knowledge.
TLDR
The preprint examines whether grammaticality—whether language follows grammatical rules—is a major organizing principle in neural language models’ internal representations. Its authors report that many models show abstract separation by grammaticality, and believe the work addresses debates about confounding factors in measuring models’ grammatical knowledge.
Combined views
2
1 Source, first seen ago