TLDR: ICR (Iterative Clarification and Rewriting) is a new framework for conversational search that improves query rewriting by iteratively generating clarification questions and refining queries. Unlike single-step rewriting, ICR uses large language models to progressively resolve ambiguities, trains models to avoid overthinking or underthinking, and fuses retrieval results from all refined queries. It achieves state-of-the-art performance on conversational search datasets and shows strong zero-shot generalization.
Conversational search has become an integral part of how we interact with information, allowing users to find answers through multi-turn dialogues. However, a common challenge arises when users’ queries are vague or omit crucial details, making it difficult for search systems to provide accurate results. This is where Conversational Query Rewriting (CQR) comes into play, aiming to transform ambiguous conversational queries into clear, self-contained ones that can be effectively used for retrieval.
Traditional approaches to CQR often employ an “end-to-end” rewriting method, where a model attempts to rewrite a query in a single step. While this has seen some success, it struggles with queries that contain multiple unclear expressions. It’s challenging for a model to identify and resolve all ambiguities simultaneously in one go, much like a human might not get the perfect answer on their first try.
Introducing ICR: A New Approach to Conversational Search
To overcome these limitations, researchers Zhiyu Cao, Peifeng Li, and Qiaoming Zhu have introduced a novel framework called ICR (Iterative Clarification and Rewriting). ICR proposes an iterative rewriting scheme that uses clarification questions as a central pivot. Instead of a single rewrite, the model alternates between generating clarification questions and then rewriting the query based on the answers or inferred information. This process allows for a progressive refinement of the query, tackling ambiguities one step at a time.
Imagine asking a search engine, “What was the purpose of this program?” In a conversational context, “this program” is highly ambiguous. ICR would first ask a clarification question like, “What program does ‘this program’ refer to?” Once that’s clarified, it might then ask, “What specific aspect of the purpose are you interested in?” Each clarification helps to build a more complete and precise query, leading to better search results.
How ICR Works: Four Key Modules
The ICR framework is built upon four main components:
1. Clarification-Rewriting Data Generation (CRDG): This module uses large language models (LLMs) to simulate the iterative process of asking clarification questions and rewriting queries. It generates a sequence of clarification questions and rewritten queries, stopping when the quality of the rewritten query (measured by retrieval performance) no longer improves. This creates a rich dataset of iterative refinement examples.
2. Progressive Supervised Fine-Tuning (PSFT): Asking a good clarification question and rewriting a query are distinct skills. PSFT trains the model in stages: first focusing on generating clarification questions, then on rewriting based on clarifications, and finally combining both skills. This progressive training helps the model learn each skill effectively without getting confused.
3. Direct Preference Optimization for Clarification-Rewriting Process (DPO-CRP): LLMs can sometimes “overthink” (ask too many unnecessary questions) or “underthink” (stop clarifying too early). DPO-CRP addresses these issues by training the model with preference data. It learns to prefer trajectories that are neither too long nor too short, and that break down complex ambiguities into individual, manageable steps. This ensures the model asks appropriate questions at the right time.
4. Process-aware Reciprocal Rank Fusion (PRRF): During the iterative process, multiple rewritten queries are generated. Instead of just using the final query, PRRF intelligently combines the retrieval results from all rewritten queries. It gives more weight to queries generated in later iterations, as these are typically more refined and accurate, thereby maximizing the benefit of the iterative process.
Also Read:
- Sticker-TTS: A New Framework for Smarter AI Reasoning Through Historical Experience
- Making Sense of AI Actions: TalkToAgent’s Approach to Explaining Reinforcement Learning
Impressive Results and Generalization
Experiments conducted on popular conversational search datasets like TopiOCQA and QReCC demonstrate that ICR significantly outperforms existing state-of-the-art methods. For instance, on the TopiOCQA dataset, ICR showed notable improvements in metrics such as MRR, NDCG, Recall@10, and R@100 for dense retrieval. This indicates that ICR not only retrieves more relevant passages but also ranks them higher.
Furthermore, ICR exhibits excellent zero-shot generalization capabilities, meaning it performs well on new, unseen datasets like CAsT-19, CAsT-20, and CAsT-21 without specific training for them. This highlights the robustness and adaptability of the iterative clarification and rewriting framework.
The research also revealed that the retrieval performance of queries improves rapidly in the initial iterations, then slows down as the query becomes more complete. This confirms the effectiveness of the iterative refinement process. The DPO-CRP module was particularly crucial in ensuring that the model’s iterative steps were effective, leading to a higher success rate in improving query quality.
While ICR marks a significant advancement in conversational search, the authors acknowledge areas for future improvement. For example, exploring ways to directly utilize the information within the clarification questions themselves, rather than just the rewritten queries, could further enhance retrieval. Additionally, refining the fusion process to discard rewritten queries that significantly deviate from the original intent could prevent noise introduction.
For a deeper dive into the methodology and experimental details, you can read the full research paper available here.
