Reposted by @psydelle.bsky.socialCentre for Language Evolution @uoe-cle.bsky.social · 25/07/2025Different ML datasets use wildly different definitions of “systematic generalization”. To REALLY tell how well a model generalizes, we must look for systematicity of the model's representations. (1/2) 👇 121
Reposted by @psydelle.bsky.socialCentre for Language Evolution @uoe-cle.bsky.social · 25/07/2025Semi-compositional language like “kick the habit” takes longer for humans to process than both fully-compositional AND fully-idiomatic units! Also, a memory model with sBERT embeddings fails more on the same items. What's so special about these units? (1/2) 👇edin.acWhat does memory retrieval leave on the table? Modelling the Cost...Despite being ubiquitous in natural language, collocations (e.g., kick+habit) incur a unique processing cost, compared to compositional phrases (kick+door) and idioms (kick+bucket). We confirm this... 121
