Shib Sankar Dasgupta
Researcher in representation learning, lead author on two key steps of the Box Embedding lineage:
- Gumbel Box (Dasgupta et al., NeurIPS 2020, Improving Local Identifiability in Probabilistic Box Embeddings) — models box corners as Gumbel random variables to fix the local-identifiability weakness of Smoothed Box.
- Word2Box (Dasgupta et al., ACL 2022, Word2Box: Capturing Set-Theoretic Semantics of Words using Box Embeddings) — unsupervised word boxes trained CBOW-style with Gumbel-box intersection.
Key Contributions
- Box Embedding — Gumbel Box optimization
- Word2Box — unsupervised set-theoretic word embeddings
- Set-Theoretic Embeddings
Articles
-
Express Words in a Box - Understanding Box Embedding from the Basics — covers his Gumbel Box and Word2Box work
-
Answering Compositional Queries with Set-Theoretic Embeddings — with Andrew McCallum, Steffen Rendle, Li Zhang; box embeddings for compositional item–attribute queries
Related People
- Luke Vilnis — originated the box-embedding line (Box Lattice)
- Andrew McCallum — co-author on set-theoretic / box-embedding work
- Steffen Rendle — co-author on compositional-query box embeddings