[[
wikihub
]]
Search
⌘K
Explore
People
For Agents
Sign in
Explore
People
For Agents
Sign in
@jemoka / Jemoka Knowledge Base / raw/concept/kbhdomains_of_generalization.md
Suggest edit
Cancel
Submit suggestion
Title
Name
Note
--- title: "Generalization" source: https://www.jemoka.com/posts/kbhdomains_of_generalization/ tags: [emnlp2024] --- Compositionality Getting the right contents. semantic capacity tested — knowing the propositional content operationalization — form => meaning mapping measure of success — generalizing to the right meaning representation for novel expressions (this is non-trivial; multiple compatible generalization maybe applicable depending on context) task Given that a model can map certain expressions to their meaning representations, can they also do this for new expressions? results lexical generalization (fill in new words/pairs) isn’t too hard for new NN but structural generalization (permute syntactical forms) is very hard subclause decomposition phrase id iterative proposional phrase and noun phrase annotation verb phrase normalization limitations “structure” requires some definition by committing to a particular formalism PTing makes exposing these stems hard Entity Tracking Getting the right entity through a conversation. task “box world”: putting stuff into boxes results models that did have a lot of code in pretraining generalized better than those that didn’t this is even shown in minimal pairs of the same architecture with more code