Jane Li 🦖

🦀New preprint! (w/ @najoung.bsky.social)🦞 Is grammaticality a major organizing principle of NLM representations? We show that many NLMs exhibit abstract rep. separation for grammaticality. We believe this work addresses debates about confounds in measuring model gram. knowledge. [1/10]

The Q of whether NLMs can have genuine distinctions of grammaticality is debated (Dentella et al., Hu et al. 2023). We believe the methods applied in this work acts as a complementary lens to the study of NLM gram. knowledge, and contribute a new perspective on these debates.

🦀New preprint! (w/ @najoung.bsky.social)🦞 Is grammaticality a major organizing principle of NLM representations? We show that many NLMs exhibit abstract rep. separation for grammaticality. We believe this work addresses debates about confounds in measuring model gram. knowledge. [1/10]

Studying under what data or architectural conditions we see detectability or generalizability in rep. space, and tying this to existing work on syntactic representations in NLMs is a clear next step. [7/10]

🦀New preprint! (w/ @najoung.bsky.social)🦞 Is grammaticality a major organizing principle of NLM representations? We show that many NLMs exhibit abstract rep. separation for grammaticality. We believe this work addresses debates about confounds in measuring model gram. knowledge. [1/10]

Sentence representations can also encode far more than grammaticality, of course. We find by crossing string probability and grammaticality during probe fitting, that the representational separation is driven by grammaticality rather than probability. [5/10]

🦀New preprint! (w/ @najoung.bsky.social)🦞 Is grammaticality a major organizing principle of NLM representations? We show that many NLMs exhibit abstract rep. separation for grammaticality. We believe this work addresses debates about confounds in measuring model gram. knowledge. [1/10]

We show that gram distinctions live in sentence representation space of a range of high-capacity models, detectable by low-complexity linear probes. They are: -Generalizable across phenomena and even langs -Deconfounded from string probs -Distinct from other sentential confounds

🦀New preprint! (w/ @najoung.bsky.social)🦞 Is grammaticality a major organizing principle of NLM representations? We show that many NLMs exhibit abstract rep. separation for grammaticality. We believe this work addresses debates about confounds in measuring model gram. knowledge. [1/10]