Relevance Feedback

A user-driven loop where the searcher manually marks specific results as relevant or non-relevant after an initial retrieval. The system extracts discriminating terms from these judged documents to augment the original query vector.

• Mechanism: Uses Robertson-Spärck Jones term weighting to reweight query terms based on their distribution in the feedback set.
• Key Advantage: Provides the highest quality signal because it relies on direct human judgment.
• Trade-off: Introduces significant cognitive load and friction, making it impractical for high-volume web search but valuable in legal e-discovery and patent search.

A passive observation technique that infers relevance from user behavior without requiring explicit judgment. The system monitors click-through rates, dwell time, and scroll depth to identify which results satisfy the user's information need.

• Behavioral Signals: Clicks with long dwell times indicate relevance; quick back-clicks (pogo-sticking) signal non-relevance.
• Query Reformulation: The system can automatically add terms from clicked documents to the query or reweight the original terms.
• Cold Start Problem: Requires sufficient interaction data to build a reliable behavioral model.

An automatic technique that assumes the top-k documents from an initial BM25 or dense retrieval run are relevant, without any user input. Terms from these pseudo-relevant documents are extracted and used to expand the query for a second retrieval pass.

• Term Selection: Typically selects the top n terms by TF-IDF or Robertson-Spärck Jones weight from the feedback set.
• Query Drift Risk: If the initial retrieval is poor, the feedback loop amplifies noise, causing the second retrieval to drift further from the actual information need.
• Parameter Sensitivity: Performance heavily depends on the number of feedback documents and expansion terms selected.

A mathematical safeguard that balances the original query terms against the expansion terms extracted from feedback documents. Without a decay factor, the original user intent can be overwhelmed by the statistical noise of the feedback set.

• Rocchio Algorithm: Uses parameters α (original query weight), β (relevant document weight), and γ (non-relevant document weight) to control the vector shift.
• Interpolation: The reformulated query is a weighted linear combination of the original query vector and the centroid of the relevant document vectors.
• Stability: A high α value anchors the query to the user's original expression, preventing semantic drift.

The process of identifying which terms from the feedback document set are most useful for distinguishing relevant from non-relevant content. Not all frequent terms in relevant documents are good expansion candidates.

• Offer Weight: Measures the term's prevalence in relevant documents versus the entire collection.
• Chi-Square Feature Selection: Statistically tests the independence of a term's occurrence and document relevance.
• Stop Word Filtering: High-frequency function words are excluded even if they appear disproportionately in relevant documents, as they lack semantic discriminative power.

In dense retrieval systems, relevance feedback operates on embedding vectors rather than sparse lexical terms. The system modifies the query embedding by moving it closer to the centroid of relevant document embeddings and away from non-relevant ones.

• Query Embedding Refinement: The reformulated query vector is computed as a weighted average of the original query embedding and the mean embedding of relevant documents.
• Contrastive Adjustment: Non-relevant document embeddings can be used to push the query vector away from unhelpful regions of the semantic space.
• Computational Cost: Requires re-encoding the aggregated feedback signal, which is significantly more expensive than sparse term reweighting.

An automatic technique that assumes the top-k results are relevant without explicit user judgment. The system extracts high-weight terms from these pseudo-relevant documents and appends them to the original query.

• Blind Feedback: No user interaction required.
• Risk: Can cause query drift if the initial top results are actually non-relevant.
• Common Usage: Often used as a baseline before implementing true relevance feedback.

The general process of augmenting a user's raw query with additional terms to overcome the vocabulary mismatch problem. Relevance feedback is a specific method of achieving this.

• Global Methods: Use thesauri or WordNet (static, context-free).
• Local Methods: Use the top retrieved documents (dynamic, context-aware).
• Goal: Increase recall by matching documents that use synonyms or related concepts.

A classic vector-space model for relevance feedback that reformulates the query vector by moving it closer to the centroid of relevant documents and away from the centroid of non-relevant documents.

• Formula: Q_new = α * Q_orig + β * Centroid(Rel) - γ * Centroid(NonRel)
• Parameters: α, β, and γ control the weight of the original query, positive feedback, and negative feedback respectively.
• Limitation: Assumes spherical, well-separated clusters.

The mechanism inside relevance feedback that adjusts the importance of specific query terms based on their distribution in the feedback set. Terms prominent in relevant documents get a boost, while terms common in non-relevant documents are suppressed.

• Signal: Uses term frequency in the feedback set vs. the collection.
• Outcome: Sharpens the query to focus on highly discriminative terms.

Inferring user intent from behavioral signals rather than explicit ratings. This allows the system to learn continuously without interrupting the user flow.

• Positive Signals: Click-throughs, dwell time, saving, printing.
• Negative Signals: Quick back-clicks, skipping a result, ignoring a suggestion.
• Advantage: Generates massive training data passively.

A temporal weighting mechanism that gives higher importance to recent user interactions compared to older ones. This prevents stale preferences from dominating the query reformulation.

• Concept: Applies an exponential or linear decay to the weight of older feedback.
• Use Case: Critical for session-based search where user intent shifts rapidly.
• Implementation: Often integrated directly into the term-weighting formula.