Source: 05_RF_QE.pdf
Most users can’t write the perfect query on the first try, so the first query is treated as an initial attempt and refined. Two ideas drive this: relevance feedback (use info about which documents are relevant) and query expansion (add related terms to the query). Feedback can be explicit (user marks relevant docs) or implicit (system infers from results/clicks/thesauri), and the source can be local (top retrieved docs) or global (whole collection).
The classic explicit technique is the Rocchio Method: move the query vector toward the centroid of relevant documents and away from non-relevant ones, using tuning constants α, β, γ. It’s simple and effective. Evaluation should use the residual collection (excluding the docs the user already saw) to avoid inflated results. On the Web, asking users for judgments is too costly, so clicks become the feedback signal — but clicks reflect preferences, not direct relevance, and are skewed by trust/position bias (users over-click the top result). Skip-Above and Skip-Previous strategies convert click patterns into preference relations (agreeing with human judgments ~80% of the time), extendable across query chains, and can train a ranking function (e.g., via SVMs).
When no user input is available, implicit feedback mines correlations automatically. Local analysis clusters terms from the retrieved set — association clusters (co-occurrence), metric clusters (co-occurrence weighted by distance), and scalar clusters (similar neighborhoods) — then adds neighbor terms to the query; Local Context Analysis blends local + global using noun-group concepts over passages. Global analysis builds a similarity thesaurus (term-term relationships in a concept space) or a statistical thesaurus (clustering documents, then grouping low-frequency discriminating terms) to expand queries. Query expansion tends to improve recall but can hurt precision, so it needs careful tuning.
| Term | Meaning | Use Case |
|---|---|---|
| Relevance feedback | Use known-relevant docs to reformulate the query | Improving precision iteratively |
| Query expansion | Add related terms to the query | Improving recall |
| Explicit vs. implicit feedback | User-provided vs. system-inferred info | Lab vs. web search |
| Local vs. global analysis | Top-retrieved docs vs. whole collection | Source of expansion terms |
| Rocchio Method | Move query vector toward relevant, away from non-relevant | Classic RF in vector model |
| α, β, γ | Rocchio weights for original/relevant/non-relevant | Tuning feedback strength |
| Residual collection | All docs minus feedback docs already seen | Fair RF evaluation |
| Clickthrough data | Clicks as implicit preference signals | Web-scale feedback |
| Trust/position bias | Users over-trust top-ranked results | Caution interpreting clicks |
| Skip-Above / Skip-Previous | Rules turning clicks into preferences | Learning to rank |
| Query chain | Sequence of queries for one task | Cross-query preferences |
| Association cluster | Terms grouped by co-occurrence | Synonym expansion |
| Metric cluster | Co-occurrence weighted by in-doc distance | Proximity-aware expansion |
| Scalar cluster | Terms with similar neighborhoods | Indirect/induced synonymy |
| Local Context Analysis | Noun-group concepts from top passages | Hybrid expansion |
| Similarity thesaurus | Term-term concept space (terms indexed by docs) | Global expansion |
| Statistical thesaurus | Doc clusters → classes of low-freq terms | Global expansion |
flowchart TD
A[Original Query q] --> B[Initial Retrieval]
B --> C{Feedback source?}
C -->|Explicit| D[User marks relevant docs
or clicks]
C -->|Implicit-Local| E[Cluster terms from top docs
assoc/metric/scalar]
C -->|Implicit-Global| F[Similarity / Statistical Thesaurus]
D --> G[Rocchio: move query vector]
E --> H[Add neighbor terms]
F --> H
G --> I[Modified Query qm]
H --> I
I --> J[Re-retrieve]
J --> K[Improved results]
K -->|new query in chain| A