Pseudo-Relevance Feedback

PRF operates on a critical heuristic: the top k documents retrieved by a base model (like BM25) are treated as relevant without human judgment. This bypasses the need for user interaction but introduces risk. If the initial retrieval drifts, the feedback loop amplifies noise.

• Risk: Query drift occurs when top documents are off-topic, poisoning the expanded query.
• Mitigation: Selecting a small k (e.g., 3-10) and filtering low-quality terms reduces noise.

Not all terms from pseudo-relevant documents are useful. PRF algorithms score candidate expansion terms using metrics like Robertson Selection Value (RSV) or divergence from the collection.

• Local Analysis: Terms are ranked by their prominence in the feedback set vs. the whole corpus.
• Parameter Tuning: Typically, 10–30 top-ranked terms are added to the original query with adjusted weights.

The Relevance Model 3 (RM3) is the standard PRF implementation in modern search. It interpolates the original query model with a relevance model built from feedback documents.

• Formula: P(w|Q') = (1-λ) * P(w|Q) + λ * P(w|R)
• Lambda (λ): Controls the interpolation weight. A typical value is 0.5, balancing the original intent with the expansion signal.
• Usage: Found in research tools like Anserini and Pyserini.

PRF significantly boosts recall by adding synonyms and related terms, but can harm precision if the initial top documents are non-relevant.

• Recall Boost: Captures documents using different vocabulary (e.g., expanding 'car' to 'automobile', 'vehicle').
• Precision Risk: A query for 'Java' (the island) might drift toward 'Java' (the language) if top results are ambiguous.
• Metric: Often improves Mean Average Precision (MAP) by 10-20% on standard test collections.

Unlike explicit feedback, PRF requires no user effort, making it suitable for zero-interaction search systems. However, it lacks the accuracy of human judgments.

• Explicit Feedback: Users mark documents as relevant/irrelevant; highly precise but low adoption.
• Implicit Feedback: Clicks and dwell time are weak signals but noisy.
• PRF Advantage: Fully automatic and transparent to the user, enabling real-time query modification.

PRF is a natural fit for sparse retrieval systems like BM25 because it operates directly on the term space. The expanded query is simply a new weighted set of terms executed against an inverted index.

• Efficiency: No re-encoding of dense vectors is required; the expanded query runs as a standard bag-of-words search.
• Contrast: In dense retrieval, PRF requires averaging or interpolating dense embeddings, a process known as Pseudo-Relevance Feedback for Dense Retrieval (PRF-d) .

The supervised predecessor to pseudo-relevance feedback. This iterative technique requires a human user to explicitly mark documents as relevant or non-relevant from an initial result set. The system then uses this signal to reformulate the query—typically by adding terms from relevant documents and subtracting terms from non-relevant ones—to refine subsequent retrieval.

• Requires explicit user interaction, unlike the automatic assumption in pseudo-relevance feedback
• Often uses Rocchio's algorithm to adjust query vector weights
• Provides a stronger relevance signal but introduces friction in the search experience

A family of methods that automatically augment a user's original query with additional terms to improve recall and overcome vocabulary mismatch. Unlike pseudo-relevance feedback, which derives terms from top-ranked documents, query expansion can leverage external knowledge sources.

• Synonym expansion: Adding equivalent terms from a thesaurus or WordNet
• Hypernym/Hyponym expansion: Broadening or narrowing queries using hierarchical relationships
• Spelling correction: Fixing typographical errors before retrieval
• Word embedding expansion: Adding semantically similar terms from dense vector spaces

The core retrieval problem that pseudo-relevance feedback aims to solve. This occurs when a relevant document describes a concept using terminology that differs from the user's query terms, causing a lexical matching failure. For example, a query for 'heart attack' may miss a highly relevant document that uses the clinical term 'myocardial infarction'.

• A primary driver of poor recall in bag-of-words retrieval systems
• Pseudo-relevance feedback bridges this gap by extracting the document's own vocabulary
• Dense retrieval models address this differently by operating in a semantic embedding space

The retrieval paradigm within which pseudo-relevance feedback and BM25 operate. Sparse retrieval represents documents and queries as high-dimensional vectors where most dimensions are zero, with non-zero values corresponding to explicit term weights like TF-IDF or BM25 scores. This enables efficient matching via inverted indices.

• Contrasts with dense retrieval, which uses low-dimensional continuous embeddings
• Pseudo-relevance feedback is a classic technique for improving sparse retrieval recall
• The expanded query terms are directly added to the sparse vector representation for re-execution

The probabilistic ranking function that typically generates the initial result set for pseudo-relevance feedback. BM25 scores documents based on term frequency saturation, inverse document frequency, and document length normalization. The top-k documents ranked by BM25 are assumed relevant by the pseudo-relevance feedback algorithm.

• The quality of pseudo-relevance feedback depends heavily on the precision of this initial ranking
• If BM25 returns noisy top results, the feedback loop will amplify the noise
• Parameters k1 and b control term saturation and length normalization respectively

A modern retrieval architecture that combines results from sparse retrieval (like BM25) and dense retrieval (like DPR) using fusion algorithms such as Reciprocal Rank Fusion (RRF). Pseudo-relevance feedback can be applied to the sparse component independently or to the fused result set.

• Allows the precision of semantic search to complement the recall of lexical feedback loops
• Query expansion terms from pseudo-relevance feedback can be injected into both retrieval paths
• Addresses the limitation of pure pseudo-relevance feedback by providing a more robust initial candidate set