When humans converse, we rarely speak in fully self-contained sentences. Instead, we rely on a shared linguistic shorthand. We use pronouns (anaphora), omit words entirely (ellipsis), and assume our conversational partner remembers what was said two minutes ago.
Traditional search engines and information retrieval (IR) systems, however, are built for isolated, ad-hoc queries. They expect a standalone statement of intent, like population of Paris. If a user follows up with how many people live there?, a standard keyword-based index fails because the word there is an ungrounded pointer.
Contextual Query Resolution (CQR) is the mechanism that bridges this gap. It transforms ambiguous, context-dependent dialogue turns into precise, self-contained representations that standard retrieval pipelines can understand.
The Core Challenge: The Context Gap
In a multi-turn conversation, the meaning of the current turn is a function of the entire dialogue history. Early foundational work by Scott Miller et al. in Hidden Understanding Models of Natural Language established that processing utterances in isolation discards the broader discourse state, leading to systemic failures in understanding user intent.
This problem manifests acutely in modern applications. For example, in e-commerce, Zhao Yan et al. noted in Building Task-Oriented Dialogue Systems for Online Shopping that a massive portion of real-world user interactions are "session-aware." If a user is looking at a shoe and says do you have this in red?, the system cannot rely on single-turn natural language understanding; it must resolve the reference to the specific product currently in focus.
To solve this, researchers have split into two primary paradigms: Explicit Query Reformulation (rewriting the query into text) and Implicit Contextual Embedding (encoding the context directly into vectors).
Paradigm 1: Explicit Query Reformulation
The most intuitive way to solve contextual dependency is to rewrite the messy follow-up query into a clean, standalone sentence before passing it to a search engine.
[Dialogue History]
Turn 1: What is photosynthesis?
Turn 2: How does it help plants grow?
│
▼ (Query Reformulation Model)
[Resolved Query]
"How does photosynthesis help plants grow?"
Early implementations of this paradigm used sequence-to-sequence architectures to handle linguistic shortcuts. Pushpendre Rastogi et al. demonstrated in Scaling Multi-Domain Dialogue State Tracking via Query Reformulation that a Pointer-Generator Network could effectively copy key entities from previous turns while generating the necessary structural words to resolve pronouns and elisions. This allowed downstream single-turn agents to function without modifying their internal logic.
As large language models (LLMs) and generative architectures matured, query rewriting became highly sophisticated. Instead of just resolving pronouns, modern systems perform joint rewriting and expansion:
- Generative Expansion: Fengran Mo et al. developed ConvGQR in RALI@TREC iKAT 2023: Generative Query Reformulation for Conversational Information Seeking, showing that combining query rewriting with a "knowledge infusion" training mechanism allows smaller, fine-tuned models (like T5) to outperform massive zero-shot models like GPT-4.
- Alignment with User Preferences: Rather than relying solely on supervised text targets, Zhongkai Sun et al. introduced PA-CQR in Improving Contextual Query Rewrite for Conversational AI Agents through User-preference Feedback Learning. They utilized a Direct Preference Optimization (DPO) variant to train the rewriter based on actual user-preference feedback, drastically reducing the omission of critical entities.
The primary benefit of explicit reformulation is interoperability. As demonstrated by Peiyuan Gong et al. in CoSearchAgent: A Lightweight Collaborative Search Agent with Large Language Models, rewriting vague queries into standalone text allows agents to seamlessly interact with standard, off-the-shelf web search APIs in collaborative environments like Slack.
Paradigm 2: Implicit Contextual Embedding
While explicit rewriting is highly interpretable, it introduces a computational bottleneck: you must generate text tokens before you can even begin the retrieval process. Furthermore, text generation can introduce cascading errors if the rewriter hallucinates or chooses the wrong synonym.
To bypass this, researchers developed Conversational Dense and Sparse Retrieval. Instead of rewriting the query into text, these models map the multi-turn dialogue history directly into a shared vector space where the context is resolved implicitly.
Dense Distillation
In Few-Shot Conversational Dense Retrieval, the authors introduced ConvDR, which uses a teacher-student framework. The teacher model is a standard ad-hoc retriever that embeds a manually produced, "perfectly rewritten" standalone query. The student model is fed the raw, multi-turn dialogue history and is trained to output an embedding as close to the teacher's embedding as possible. This allows the system to resolve conversational context entirely within the vector space, drastically speeding up inference.
[Perfect Standalone Query] ──► [Teacher Encoder] ──► [Target Vector Space]
▲
(Distillation Loss)
│
[Raw History + Current Turn] ──► [Student Encoder] ─────────┘
Sparse Contextualization
Alternatively, context can be injected into sparse representations. Nam Le Hai et al. explored this in MLIA-DAC@TREC CAsT 2022: Sparse Contextualized Query Embedding by extending the SPLADE sparse retrieval model. Instead of generating a natural language sentence, their model directly predicts a sparse vector of term weights over the entire vocabulary, effectively performing query expansion and context resolution simultaneously without an explicit text-generation step.
To make these models robust across different domains without massive human annotation, the author of Domain Adaptation for Dense Retrieval and Conversational Dense Retrieval through Self-Supervision by Meticulous Pseudo-Relevance Labeling. demonstrated that self-supervised domain adaptation—using cross-encoders to generate high-quality pseudo-labels—can successfully bridge the gap when deploying conversational dense retrievers to new, unseen topics.
Managing Context Drift and Flow
A naive approach to conversational context is to simply concatenate all previous turns into a long string. However, this introduces context drift—as a conversation progresses, older topics become irrelevant, cluttering the context window and confusing the retriever.
Selecting Relevant Turns
To combat this, Esteban A. Ríssola et al. proposed a neural turn-selection mechanism in Predicting Relevant Conversation Turns for Improved Retrieval in Multi-Turn Conversational Search. By using a BERT-based model to predict which historical turns are actually relevant to the current information need, they filter out conversational noise before query reformulation occurs. Similarly, Xiaochen Zuo et al. utilized a Key-Value Memory Network in RUCIR at TREC 2019: Conversational Assistance Track paired with a "freshness" weighting mechanism to prioritize recent context while maintaining access to recurring core themes.
Tracking Dialogue Flow
Rather than treating history as a static document, some architectures model the conversation as a dynamic progression. Hsin-Yuan Huang et al. developed FlowQA in FlowQA: Grasping Flow in History for Conversational Machine Comprehension. Instead of concatenating text, FlowQA introduces a mechanism that passes the internal hidden representations (the model's contextual reasoning) sequentially from one question to the next, allowing the model to naturally track the "flow" of dialogue over time.
For complex, multi-party environments, Changmao Li and Jinho D. Choi demonstrated in Transformers to Learn Hierarchical Contexts in Multiparty Dialogue for Span-based Question Answering that a hierarchical Transformer architecture is necessary. Their model uses specialized pre-training tasks to explicitly learn the structural relationships and causal dependencies between different speakers across disparate utterances.
The Frontier: System-Response Dependency
Initially, CQR benchmarks focused primarily on resolving dependencies pointing back to previous user queries. However, a major paradigm shift occurred with the realization that human search behavior is fundamentally reactive to what the system returns.
As highlighted by Jeff Dalton et al. in CAsT 2020: The Conversational Assistance Track Overview, retrieval performance drops significantly when a user's query depends on information provided in a previous system response rather than a previous user question.
User: Find me some good hiking trails in Oregon.
System: Silver Falls State Park is highly rated and features several waterfalls.
User: Are dogs allowed there? ◄── (Dependency is on the system's response, not the user's past turn)
Resolving this requires the system to parse its own generated output or retrieved passages to ground the user's follow-up intent. To handle this, the authors of University of Cambridge at TREC Cast 2022 designed a hybrid pipeline: they use raw, concatenated conversation history for broad initial document retrieval, and then apply a precise T5-rewritten query for precision re-ranking.
Modern Retrieval-Augmented Generation (RAG) systems, such as ChatQA developed by Zihan Liu et al. (ChatQA: Surpassing GPT-4 on Conversational QA and RAG), integrate these multi-turn context resolution capabilities directly into the LLM's fine-tuning process. This ensures that follow-up questions relying on previously retrieved snippets are correctly grounded, allowing conversational agents to maintain deep, multi-turn consistency across long open-domain interactions.