ColBERT

ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT

Omar Khattab and Matei Zaharia · SIGIR 2020

Read paper · arXiv:2004.12832

Preserve token vectors and sum each query token's best document match.

What you will implement

Cosine MaxSim, masks and pairwise softmax with synthetic query-vector optimization; no BERT training or ANN index.

• 66. ColBERT: Token-Level MaxSim and Pairwise Ranking

  Preserve per-token matching evidence and optimize positive/negative document ranking with a stable pairwise loss.

  Supplied projected token vectors, cosine MaxSim and pairwise softmax; no BERT encoder, tokenizer or ANN index training.