Traditional information retrieval operates on a stateless assumption: each query entered into a search bar is treated as an isolated, independent event. As Downey et al. point out in Models of Searching and Browsing: Languages, Studies, and Application, this narrow focus fails to capture the iterative nature of how humans actually look for information. When a user enters a short, ambiguous query like "Apple", a stateless system must guess whether they want the fruit, the technology company, or a local orchard.
However, search is rarely a one-shot action; it is a journey. As Radlinski et al. demonstrate in Query chains: learning to rank from implicit feedback, users act as intelligent reformulators, modifying their search terms over multiple attempts to narrow down results based on what they find. If the user's previous queries in the same session were "iPhone price" and "MacBook Pro", the ambiguity of "Apple" disappears. By breaking the query independence assumption, session-aware models leverage this temporal context to decode latent user intent.
Representing the Session State
To move beyond isolated query processing, retrieval systems must maintain a persistent session state that tracks the user's trajectory. The simplest way to build this context is through representation aggregation. In THUIR2 at NTCIR-16 Session Search (SS) Task, Li et al. augment the current query by concatenating the chronological history of previous queries and the titles of documents clicked within that session. Similarly, the authors of ECNU at CLEF PIR 2018 : Evaluation of Personalized Information Retrieval show that combining all queries within a single session into a unified representation mitigates the narrowness of individual search terms, providing a more comprehensive picture of the user's information-seeking process.
Beyond simple concatenation, modern architectures model sessions as sequential trajectories or graphs to capture fine-grained behavioral dynamics:
- Temporal Trajectories: In A Context-aware Time Model for Web Search, Borisov et al. utilize Recurrent Neural Networks (RNNs) to track session history, demonstrating that a user's dwell time on a current result is heavily biased by their prior actions in the session.
- Session Graphs: In e-commerce settings where queries are notoriously fragmented, Zuo et al. leverage session graphs and graph attention mechanisms in Context-Aware Query Rewriting for Improving Users' Search Experience on E-commerce Websites to automatically rewrite vague queries based on persistent intent signals from previous interactions.
Distributing Intent Across Keystrokes
The intuition behind session-aware modeling extends even to early-stage search interactions, such as Query Autocomplete (QAC). When a user types a single character or prefix, predicting their final query based solely on those keystrokes is highly speculative.
However, as Everaert et al. demonstrate with the AmazonQAC: A Large-Scale, Naturalistic Query Autocomplete Dataset, incorporating session history dramatically improves generative query completion. This confirms that search intent is frequently distributed across a sequence of interactions over time, rather than being contained entirely within the current partial input.
The Challenge of Contextual Noise
While integrating session history provides powerful disambiguation signals, it introduces a major challenge: context noise. Not all search sessions are linear or focused on a single topic. Users frequently experience sudden topic shifts or multitask within a single browsing window.
As Ma et al. observe in MM6 at the NTCIR-16 Session Search Task, blindly incorporating historical session context can sometimes degrade search accuracy compared to treating queries independently. If a user abruptly switches from researching a work project to looking up a dinner recipe, forcing the historical context onto the new query creates harmful interference. Thus, the frontier of session-aware retrieval lies not just in capturing context, but in adaptively deciding when a new query has broken free from the past.