Query Expansion Techniques: Relevance Modeling vs. Generative Approaches and Future Directions

The article begins by recalling the two main families of query‑expansion technologies: relevance‑modeling, which builds a network of entities and concepts using Bayesian probability, and translation‑modeling, which treats expansion as a generation problem.

It explains that short queries containing a single entity are well‑served by relevance models, while longer queries with multiple entities or complex semantics benefit from generative approaches that preserve the overall meaning.

Three key challenges for complex queries are identified: (1) handling multiple entities and additional textual constraints, (2) dealing with non‑standard expressions that may miss relevant results, and (3) supporting users who only have a vague idea of the target, often expressed through a session of related queries.

The paper surveys recent works, noting two common task formulations: candidate‑ranking (scoring query‑candidate pairs) and direct generation of expanded queries or terms. It highlights the use of word embeddings, LSTM/Transformer encoders, attention, copy mechanisms, and adversarial examples to improve encoding.

Pre‑trained language models such as BERT and BART are discussed as powerful encoders that can be fine‑tuned for query suggestion, allowing better contextual understanding.

Beyond single‑query encoding, the article reviews session‑based modeling, where a series of user queries is encoded to capture context and reformulation behavior, often using multi‑task architectures that jointly perform candidate discrimination, query generation, and rewriting.

Feature‑rich approaches are also covered, including the incorporation of click‑through data and graph‑based representations (e.g., Node2Vec) to enrich query embeddings, as well as document‑side encoding to improve context awareness.

In summary, effective query expansion for long queries requires comprehensive query encoding, careful negative‑sample selection, attention/copy mechanisms, task‑specific objectives, and robust feature engineering, especially leveraging structured data and user interaction signals.