DUET
What it is
DUET (Mitra et al., 2017) uses two parallel sub-networks: a local model that scores based on exact query-term matches in the document (similar to TF-IDF counting), and a distributed model that uses learned dense representations for semantic matching. The scores are combined linearly. DUET was one of the first models to explicitly formalize that good retrieval requires both lexical precision and semantic recall — a tension that later hybrid search methods continue to address.
[illustrate: Query + document → local sub-network (exact match counts) + distributed sub-network (dense embeddings) → combined score]
How it works
-
Local sub-network:
- Interaction matrix: 1 where query term appears in document, 0 otherwise
- CNN over the binary interaction matrix
- Captures exact-match and proximity signals
-
Distributed sub-network:
- Character trigram representations (like DSSM)
- Learned dense embeddings for query and document
- Element-wise product and MLP for interaction
-
Combination:
- Weighted sum of local and distributed scores
- Weights learned during training
Variants and history
DUET (2017, Microsoft Research) anticipated the hybrid retrieval paradigm — the combination of sparse (BM25-like) and dense signals that later became standard in production IR. The DUET v2 paper analyzed which components matter most. Modern hybrid search systems (combining BM25 + dense retrieval with RRF or learned interpolation) are conceptual descendants.
When to use it
Primarily historical. The explicit lexical + semantic combination is now better handled by:
- Hybrid search (BM25 + dense retrieval with learned interpolation)
- SPLADE (learns sparse representations that capture both)