Information retrieval systems often struggle with the "vocabulary gap"—the discrepancy between the specific words a user types and the different but relevant words used in documents. To bridge this gap, researchers use Query Expansion (QE) and Pseudo-Relevance Feedback (PRF), techniques that automatically add related terms to a search. However, these methods frequently introduce a destructive side effect known as query drift or semantic drift.
Semantic drift occurs when the expansion process inadvertently shifts the focus of a search away from the user’s original intent. As Zhao et al. note, this typically happens when expansion candidates are selected based on ambiguous terms without sufficient contextual filtering. The result is a system that might improve average performance across many queries but fails spectacularly on difficult ones, a trade-off that Ding et al. identify as a major hurdle for system robustness.
The Mechanics of Intent Displacement
The intuition behind semantic drift is simple: if the initial search results are even slightly off-topic, using them to "improve" the query will only lead the system further astray. Collins-Thompson describes this as an instability rooted in the assumption that top-ranked documents are relevant; when they are not, the expansion process amplifies the error.
This drift manifests in several ways:
- Ambiguity and Polysemy: When a query contains a word with multiple meanings, expansion can pull in terms from the wrong context. El-Ganainy et al. observe that ambiguous terms often introduce contextually unrelated information from external sources, degrading precision. In visual tasks, Gan et al. show that keyword-based expansion can return results that match the text but fail the visual concept, such as retrieving static objects when the user intended a dynamic action.
- Over-Expansion and Dilution: Adding too many terms, even synonyms, can dilute the core intent. Gerber et al. argue that broad expansion using general lexical databases often introduces common, generic words that drown out specific concepts. Similarly, Zhang and Lu found that retrieval performance often peaks with a very small number of high-quality expansions, as excessive variations introduce semantic noise.
- Segment Imbalance: For complex queries, the system might over-emphasize one part of the request at the expense of the whole. Li et al. describe this as "expansion drift," where the model prioritizes documents relevant to a sub-component but irrelevant to the holistic meaning of the search.
Strategies for Grounding Expansion
To combat drift, researchers focus on "grounding" the expansion process—ensuring that new terms remain anchored to the original query’s context.
One approach involves adaptive weighting. Instead of treating all expansion terms equally, models like the one proposed by Bendersky et al. learn to prioritize both the explicit words in a query and the latent concepts added during expansion. Zhai and Lv further suggest that the balance between the original query and feedback documents should be adjusted for every individual search to prevent the system from losing focus.
Another strategy uses external knowledge to provide a cleaner signal than the noisy dataset being searched. Egozi et al. leverage Wikipedia to map queries to high-level concepts, using unsupervised filtering to avoid unrelated topical noise. Weerkamp et al. demonstrate that expanding queries using high-quality external collections like news archives can bridge the vocabulary gap without the topic drift associated with searching noisy, user-generated content.
Modern Neural and LLM Perspectives
The rise of Large Language Models (LLMs) and dense retrieval has introduced new facets to the drift problem. While LLMs are excellent at reformulating queries, they can still introduce extraneous terms. Hui et al. address this in conversational search by using "Retrieval Fusion," combining multiple LLM-generated variations to filter out noise through a "Chorus Effect."
In neural retrieval, where queries are mapped to latent vector spaces, the risk is that these representations become uninterpretable or disconnected from the original intent. Lin and Lin note that latent embeddings can incorporate distracting terms that are semantically related but irrelevant to the specific user intent. To solve this, the authors of QA4PRF treat expansion as a machine reading comprehension task, using attention mechanisms to ensure that expansion terms capture deep contextual meaning rather than just statistical co-occurrence.