Document Expansion
Definition
Document Expansion
Document expansion enriches a document’s representation with additional terms (or term weights) at index time, before retrieval, to reduce the vocabulary mismatch between how authors write documents and how users phrase queries. The expanded document is then indexed by a standard retriever (e.g. BM25), so the cost is paid once during indexing and queries stay fast.
Intuition
Meet the query halfway
Lexical retrieval rewards exact term overlap, but a relevant document may simply never use the user’s words (“heart attack” vs “myocardial infarction”). Instead of expanding the query at search time (Query Expansion), document expansion expands the document ahead of time — predicting the words and questions a document could answer and folding them into its index entry. Because it happens offline, it adds no query-time latency.
Mathematical Formulation
There is no single equation; the operation augments the term set / term frequencies used by the scorer. For a generative expansion that appends predicted queries to document :
\text{score}(q,d) = \text{BM25}(q, d')$$ where: - $d'$ — the expanded document fed to the index - $\hat{q}_i$ — a synthetic query generated from $d$ (raising the term frequencies of useful, possibly novel terms) - $\Vert$ — concatenation For reweighting-based expansion the term set is unchanged but the per-term weights $w(t,d)$ are learned/predicted rather than taken from raw counts. ## Key Properties / Variants - **Generative (term addition):** [[DocT5Query]] / doc2query uses a seq2seq model (T5) to generate likely queries per document and appends them, adding *new* vocabulary. - **Reweighting (no new terms):** [[DeepCT]] and [[DeepImpact]] predict context-aware term weights for the terms already present — a bridge toward [[Learned Sparse Retrieval]]. - **Index-time vs query-time:** complementary to [[Query Expansion]] (which augments the query). - **Backbone-agnostic:** the expanded index is still a sparse [[Inverted Index]], so it inherits the efficiency and interpretability of lexical retrieval while recovering some semantic recall. ## Connections - Instances: [[DocT5Query]], [[DeepCT]], [[DeepImpact]] - Contrast with: [[Query Expansion]] (query-side) - Leads toward: [[Learned Sparse Retrieval]], [[SPLADE]] - Improves: [[BM25]] recall by mitigating vocabulary mismatch ## Appears In - [[Query Expansion]] - [[Learned Sparse Retrieval]] - [[IR-L07 - Learned Sparse Retrieval]] - [[IR-PTR Ch4 - Refining Query and Document Representations]]