Pseudo-Relevance Feedback

PRF operates on a local feedback principle, treating the top-k documents (typically 10-50) from an initial retrieval as a pseudo-relevant set. Unlike explicit relevance feedback, no human judgment is required. The core assumption is that the initial retrieval algorithm, such as BM25, is precise enough that the top results are mostly on-topic. This assumption is fragile; if the initial retrieval fails and returns off-topic documents, the expansion terms will introduce noise, causing query drift and degrading recall.

Once the pseudo-relevant set is identified, PRF must extract the most useful expansion terms. Common strategies include:

• Robertson/Sparck Jones weight: A probabilistic weighting function that scores terms based on their frequency in the pseudo-relevant set versus the entire collection.
• Kullback-Leibler divergence: Measures the divergence between the term's distribution in the pseudo-relevant set and its background distribution.
• Chi-squared feature selection: Ranks terms by their statistical dependence on the pseudo-relevant class. Typically, 10-30 top-weighted terms are added to the original query with a lower boost weight than the original query terms.

PRF inherently defines a two-stage retrieval pipeline:

1. Initial Retrieval: A fast, often sparse retrieval method like BM25 retrieves the top-k candidate documents.
2. Term Extraction & Re-Retrieval: Expansion terms are mined from the candidates, appended to the original query, and a second retrieval pass is executed against the full index. This architecture adds latency, making it challenging for real-time search. Modern implementations often cache the expansion terms for frequent queries or use a lightweight initial retriever to minimize the performance hit.

While classically a sparse retrieval technique, PRF has been adapted for dense retrieval systems:

• Vector-based PRF: Instead of extracting discrete terms, the embeddings of the top-k pseudo-relevant documents are averaged to create a centroid vector. The original query vector is then interpolated with this centroid to pull it closer to the relevant cluster in the embedding space.
• ANCE (Approximate Nearest Neighbor Negative Contrastive Estimation): Uses an asynchronously updated index to select hard negatives for training, a concept related to PRF's self-training loop.
• Generative PRF: Large language models read the top-k pseudo-relevant documents and generate new, abstractive query terms or entirely reformulated queries.

PRF is not universally beneficial and can severely degrade performance in specific scenarios:

• Query Drift: The most common failure, where the expansion terms shift the query's focus away from the original information need. This is frequent with ambiguous queries.
• Mitigation via Selectivity: Modern PRF systems use classifiers to predict if a query will benefit from expansion. Features include the clarity score of the initial result set and the query's historical performance.
• Mitigation via Anchoring: Forcing the original query terms to have a higher weight than the expansion terms, or restricting expansion terms to those that co-occur frequently with the original query terms in the pseudo-relevant set.

The key distinction lies in the source of the relevance signal:

• Explicit Relevance Feedback: Relies on a human user marking documents as 'relevant' or 'not relevant'. The Rocchio algorithm is a classic method that refines the query vector by moving it toward the centroid of relevant documents and away from non-relevant ones. This is highly accurate but requires user effort.
• Pseudo-Relevance Feedback: Automates this process by assuming the top-k are relevant. It requires no user effort but is inherently less accurate and prone to noise amplification. PRF is often used as a baseline against which more sophisticated, interactive relevance feedback systems are measured.

The primary failure mode of pseudo-relevance feedback. Query drift occurs when the initial top-k documents contain a significant number of non-relevant results, causing the expansion terms extracted from them to shift the query's semantic focus away from the user's original information need. For example, a query for 'python' might drift toward 'reptile care' if the initial retrieval is ambiguous. Mitigation strategies include using a selective feedback approach, where only a subset of the most reliable top terms or documents is used for expansion.

A modern neural approach that contrasts with the term-based statistical extraction of pseudo-relevance feedback. Instead of simply counting frequent terms in top documents, this technique uses deep language models like BERT to read the query and each top document, generating expansion terms that are semantically appropriate for the query's specific context. This method can produce abstractive expansions—terms that capture the concept but may not literally appear in the top documents—providing a more robust defense against vocabulary mismatch than purely extractive methods.

Leverages large language models, such as T5 or GPT, to generate expansion terms or full alternative queries. Unlike pseudo-relevance feedback, which is extractive and relies on an initial retrieval pass, generative expansion can be zero-shot, prompting the model to produce synonyms, related entities, or paraphrases directly from the query string. However, a hybrid approach is emerging where the top-k documents from an initial retrieval are fed as context into the LLM prompt, grounding the generative expansion in the same pseudo-relevance assumption but using a more powerful, abstractive generation mechanism.

A document expansion technique that inverts the logic of pseudo-relevance feedback. Instead of expanding the query at search time, doc2query uses a sequence-to-sequence model to generate potential queries that a document might answer and appends them to the document's index during ingestion. This bridges the vocabulary gap by making documents more likely to match the varied phrasing of user queries. While pseudo-relevance feedback enriches the query from top documents, doc2query enriches all documents with predicted queries, shifting the computational cost to indexing time.