ColBERT
A BERT-based late interaction retrieval model that retains multiple token-level vectors per document and improves recall precision through fine-grained matching. It is a foundational method in multi-vector embedding retrieval.
Timeline (last 90 days)
Late interaction is a modeling paradigm in information retrieval, popularized by the ColBERT line of work, which uses operations such as MaxSim to compute the sum of maximum similarities between query and document tokens as the relevance score.
ViDoRe adopts a ColBERT-style late interaction mechanism: documents are split into patches, each patch generates a dense vector, queries are likewise encoded as multi-vectors, and scoring is done via MaxSim by summing the per-vector maximum similarities.
ColBERT (Contextualized Late Interaction over BERT) was proposed by Stanford University in 2020 and is the foundational work of the late-interaction architecture.
ColBERT's core innovation is the MaxSim operation: for each token vector in a query, it finds the most similar token vector among all token vectors in the document, and sums these maximum similarity scores to obtain the final matching score.
ColBERT was proposed by Stanford University in 2020, with the core idea that the query side and document side each independently encode to generate token-level vector sequences, and relevance scores are obtained via MaxSim operations during retrieval.
ColBERT (Contextualized Late Interaction over BERT) is an efficient neural retrieval model proposed by Omar Khattab et al. at Stanford University in 2020, which computes fine-grained matching scores at the token level via the MaxSim operation.
Multi-vector retrieval represented by ColBERT retains an independent vector for each token in a document and achieves more fine-grained matching through a late interaction mechanism.
All Facts (7)
Late interaction is a modeling paradigm in information retrieval, popularized by the ColBERT line of work, which uses operations such as MaxSim to compute the sum of maximum similarities between query and document tokens as the relevance score.
50%UnverifiedViDoRe adopts a ColBERT-style late interaction mechanism: documents are split into patches, each patch generates a dense vector, queries are likewise encoded as multi-vectors, and scoring is done via MaxSim by summing the per-vector maximum similarities.
50%UnverifiedColBERT (Contextualized Late Interaction over BERT) was proposed by Stanford University in 2020 and is the foundational work of the late-interaction architecture.
50%UnverifiedColBERT's core innovation is the MaxSim operation: for each token vector in a query, it finds the most similar token vector among all token vectors in the document, and sums these maximum similarity scores to obtain the final matching score.
50%UnverifiedColBERT was proposed by Stanford University in 2020, with the core idea that the query side and document side each independently encode to generate token-level vector sequences, and relevance scores are obtained via MaxSim operations during retrieval.
50%UnverifiedColBERT (Contextualized Late Interaction over BERT) is an efficient neural retrieval model proposed by Omar Khattab et al. at Stanford University in 2020, which computes fine-grained matching scores at the token level via the MaxSim operation.
50%UnverifiedMulti-vector retrieval represented by ColBERT retains an independent vector for each token in a document and achieves more fine-grained matching through a late interaction mechanism.
50%