Christopher Potts#225
Just finished watching a gem by @ChrisGPotts , "Finding linguistic structure in large language models", and I'm now properly convinced that *LLMs do learn linguistic structures*. I guess the only argument for neuro-symbolic models (compared to ...
Users recommend the video on LLMs learning linguistic structures because it is very informative with Chris Potts providing excellent explanations of the concepts.
Anyway, highly recommend watching the video as it is very informative, and Chris Potts does an amazing job in terms of explaining the concepts in a super nice and clear way (as always) : https://www.youtube.com/watch?v=DBorepHuKDM
e2e neural ones) would be that they can potentially learn these structures (rare or common) more efficiently (i.e. reaching threshold performance with fewer training data) given that some linguistic bias is embedded in them? 🤔
And maybe better continual learning and less hallucinations? Although, I'm not sure these have been reported for NeSy models. Also interpreting them would be easier?
@incrementaliser Thank you so much for the kind words and this thoughtful thread!
Christopher Potts#225Just finished watching a gem by @ChrisGPotts , "Finding linguistic structure in large language models", and I'm now properly convinced that *LLMs do learn linguistic structures*. I guess the only argument for neuro-symbolic models (compared to ...