Source: Khattab & Zaharia 2020: ColBERT

SIGIR 2020 paper introducing late interaction: per-token embeddings matched cheaply at query time.

Contents

Summary

Sits between bi-encoders and cross-encoders: query and document are encoded independently into token-level vectors (so documents can be precomputed), then a cheap interaction step scores the match.

Key claims

  • Document representations are computed offline, so only the query is encoded at search time.
  • Two orders of magnitude faster and four orders of magnitude fewer FLOPs per query than BERT cross-encoder ranking.
  • Competitive with BERT-based rankers in effectiveness and better than all non-BERT baselines on two passage-search benchmarks.

Feeds

Reranking

Suggest changes →