Relevance Feedback

The mechanism relies on binary or graded relevance assessments provided directly by the user. Unlike implicit signals (clicks, dwell time), explicit feedback removes ambiguity.

• Binary Feedback: User marks results as simply 'relevant' or 'non-relevant'.
• Graded Feedback: User assigns a score (e.g., 1-5 stars) to indicate degrees of relevance.
• Signal Clarity: This direct labeling provides a high-confidence training signal for the algorithm, distinguishing it from Pseudo-Relevance Feedback which assumes top results are relevant.

The classic vector space model for implementing relevance feedback. It reformulates the query vector by moving it towards the centroid of relevant documents and away from the centroid of non-relevant documents.

• Formula: Q_new = α * Q_orig + β * (Σ Relevant / |Rel|) - γ * (Σ Non-Relevant / |NonRel|)
• Parameters: α, β, and γ control the influence of the original query, relevant documents, and non-relevant documents respectively.
• Effect: This geometrically shifts the search vector in the embedding space to increase the rank of documents similar to those marked relevant.

The core operation where the importance of specific keywords is adjusted based on their distribution in the feedback set.

• Positive Expansion: Terms that appear frequently in relevant documents but were absent from the original query are added with a high weight.
• Negative Suppression: Terms that appear frequently in non-relevant documents are down-weighted or removed to prevent false positives.
• Dynamic Thesaurus: This process effectively builds an on-the-fly thesaurus of related terms specific to the user's current session context.

Relevance Feedback is not a single-step process; it is designed for recursive refinement. Each iteration provides the system with more data, progressively narrowing the semantic gap.

• Session Context: The system maintains state across multiple query reformulations within a single search session.
• Diminishing Returns: After 2-3 iterations, the marginal gain in precision typically plateaus as the user's information need is fully captured.
• Stopping Criteria: The process ends when the user stops providing feedback, usually because the desired information has been found.

While both techniques modify the initial query, their triggers and methods differ fundamentally.

• Relevance Feedback: Triggered by post-hoc user judgments on retrieved results. It is an interactive, supervised loop.
• Automatic Query Expansion: Triggered pre-hoc by a static knowledge base (like WordNet) or global co-occurrence statistics. It is a blind, unsupervised process.
• Precision vs. Recall: Relevance Feedback is typically superior for increasing precision in a specific session, while blind expansion often sacrifices precision for higher recall.

The primary barrier to adoption in modern web search is the cold-start problem and user patience.

• Cognitive Load: Requiring a user to explicitly judge results adds friction to the search experience.
• Latency: Re-running a modified query with new term weights must happen in real-time (< 100ms) to feel seamless.
• Modern Adaptation: In enterprise and e-discovery platforms, where precision is paramount and users are highly motivated, explicit relevance feedback remains a critical tool for complex information retrieval tasks.

An automatic variant that assumes the top-k initially retrieved documents are relevant without explicit user input. The system extracts high-frequency terms from these pseudo-relevant documents and appends them to the original query. While computationally efficient, this blind approach risks query drift if the initial retrieval set contains off-topic results. Key parameters include the number of documents sampled and the number of expansion terms selected.

The classic vector-space formulation for relevance feedback, developed by Joseph Rocchio in 1971. It reformulates a query vector by adding the centroid of relevant document vectors and subtracting the centroid of non-relevant document vectors, weighted by tunable parameters α, β, and γ. The resulting vector moves closer to relevant clusters and away from irrelevant ones in the embedding space. This algorithm remains foundational in modern dense retrieval systems.

The broader user-driven process of iteratively modifying a search query based on observed results. Unlike automatic expansion, reformulation involves explicit human decisions—adding quotes for phrase matching, switching keywords, or applying filters. Session context tracking enables systems to learn from these reformulation chains and suggest improvements. It is the manual counterpart to algorithmic relevance feedback.

A modern approach using large language models to generate expansion terms or full alternative queries from a prompt. Unlike statistical methods, LLMs can produce contextually rich synonyms, paraphrases, and even hypothetical answers. The model is prompted with the original query and instructed to produce terms that relevant documents would likely contain. This technique bridges the vocabulary gap between short keyword queries and long-form document language.

A critical failure mode in relevance feedback where expansion terms shift the query's semantic focus away from the user's original intent. This occurs when irrelevant documents are mistakenly judged as relevant, or when pseudo-relevance feedback samples noisy top-k results. Mitigation strategies include:

• Term weighting to de-emphasize expansion terms
• Query anchoring to preserve original terms
• Selectivity filters that reject low-confidence expansion candidates

Enriches queries using session history, user location, device type, and behavioral signals rather than document content alone. For example, a query for 'java' from a user with a programming search history expands differently than one from a coffee enthusiast. This technique personalizes relevance feedback by incorporating implicit relevance signals—clicks, dwell time, and scroll depth—alongside explicit judgments.