How do language models organize concepts and their properties? Do they use taxonomies to infer new properties, or infer based on concept similarities? Apparently, both! 🌟 New paper with my fantastic collaborators @amuuueller.bsky.social and @kanishka.bsky.social - ThreadSky

juand-r.bsky.social • 110 days ago

How do language models organize concepts and their properties? Do they use taxonomies to infer new properties, or infer based on concept similarities? Apparently, both!

🌟 New paper with my fantastic collaborators @amuuueller.bsky.social and @kanishka.bsky.social

Comments

domodonn.bsky.social•101 days ago

📌

juand-r.bsky.social•110 days ago

1/ 🧠 Property inheritance (PI) entails projecting traits from broader categories to specific instances. If birds 🐦can fly, we might assume sparrows can too. Taxonomic relations between categories can explain PI, but Sloman '98 showed humans are sensitive to concept similarities.

juand-r.bsky.social•110 days ago

2/ 🧩 For example, humans are more likely to project properties to concepts which are more similar (e.g., from dog to corgi) than less similar (e.g., from dog to shar pei).
🤔 How does this materialize in LMs?
Do LMs rely on taxonomic relations or similarity relations (or both)?

juand-r.bsky.social•110 days ago

3/ 🤖 To investigate, we construct a dataset of concept pairs from the THINGS database, and—armed with two similarity metrics—we use 𝗯𝗲𝗵𝗮𝘃𝗶𝗼𝗿𝗮𝗹 and 𝗰𝗮𝘂𝘀𝗮𝗹 interpretability methods to disentangle how LMs represent property inheritance relations. 🕵️‍♀️

juand-r.bsky.social•110 days ago

4/ 🔄 We find that LMs rely on both taxonomic structure and similarity in property inheritance. They’re more likely to project new properties to concepts that are 𝘣𝘰𝘵𝘩 𝘵𝘢𝘹𝘰𝘯𝘰𝘮𝘪𝘤𝘢𝘭𝘭𝘺 𝘳𝘦𝘭𝘢𝘵𝘦𝘥 𝘢𝘯𝘥 𝘩𝘪𝘨𝘩𝘭𝘺 𝘴𝘪𝘮𝘪𝘭𝘢𝘳. But, we can say even more than this:

juand-r.bsky.social•110 days ago

5/ 🔄 We use Boundless DAS (Zhengxuan Wu, Atticus Geiger, Thomas Icard, Christopher Potts, @noahdgoodman.bsky.social) to localize property inheritance in LMs. By using ambiguous train sets with generalization test sets, we test whether the subspaces are sensitive to similarity effects.

juand-r.bsky.social•110 days ago

6/ 🔍We find that the subspaces learned by B-DAS are sensitive to similarity effects. This sensitivity persists even when the training data’s labels can only be explained by taxonomic relations. This means similarity and taxonomy are fundamentally entangled in the subspaces! 🤯

moppety.bsky.social•110 days ago

Really interesting! And here's a call for papers with a special session on ontology and LLMs that might interest you... deadline November 25. https://bsky.app/profile/oasis-ontology.bsky.social/post/3l7w75nkytp24

juand-r.bsky.social•110 days ago

Thank you! I'll take a look

simonesprenger.bsky.social•81 days ago

📌

Comments

Posting Rules

Reply