Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Asking Clarifying Questions for Preference Elicitation With Large Language Models

Ali Montazeralghaem, Guy Tennenholtz, Craig Boutilier, Ofer Meshi

TL;DR

This work tackles eliciting user preferences in conversational recommendations when user history is sparse. It introduces a diffusion-inspired two-stage framework that forward-corrupts a complete user profile by sequentially asking and removing information and then reverses this process by training an LLM to reconstruct the profile through effective clarifying questions, producing a funnel dialogue. The method relies on two LLMs, a Questioner and a User Simulator, and leverages Sequential Question Answering to generate training data that guides domain-general funnel questioning; experiments on Movielens show significant gains in elicitation quality (BLEU/ROUGE) and demonstrate a clear shift from broad to detailed questions. The results highlight the importance of fine-tuning and incorporating question history, offering a scalable approach to enhance preference elicitation across domains with LLMs.

Abstract

Large Language Models (LLMs) have made it possible for recommendation systems to interact with users in open-ended conversational interfaces. In order to personalize LLM responses, it is crucial to elicit user preferences, especially when there is limited user history. One way to get more information is to present clarifying questions to the user. However, generating effective sequential clarifying questions across various domains remains a challenge. To address this, we introduce a novel approach for training LLMs to ask sequential questions that reveal user preferences. Our method follows a two-stage process inspired by diffusion models. Starting from a user profile, the forward process generates clarifying questions to obtain answers and then removes those answers step by step, serving as a way to add ``noise'' to the user profile. The reverse process involves training a model to ``denoise'' the user profile by learning to ask effective clarifying questions. Our results show that our method significantly improves the LLM's proficiency in asking funnel questions and eliciting user preferences effectively.

Asking Clarifying Questions for Preference Elicitation With Large Language Models

TL;DR

This work tackles eliciting user preferences in conversational recommendations when user history is sparse. It introduces a diffusion-inspired two-stage framework that forward-corrupts a complete user profile by sequentially asking and removing information and then reverses this process by training an LLM to reconstruct the profile through effective clarifying questions, producing a funnel dialogue. The method relies on two LLMs, a Questioner and a User Simulator, and leverages Sequential Question Answering to generate training data that guides domain-general funnel questioning; experiments on Movielens show significant gains in elicitation quality (BLEU/ROUGE) and demonstrate a clear shift from broad to detailed questions. The results highlight the importance of fine-tuning and incorporating question history, offering a scalable approach to enhance preference elicitation across domains with LLMs.

Abstract

Large Language Models (LLMs) have made it possible for recommendation systems to interact with users in open-ended conversational interfaces. In order to personalize LLM responses, it is crucial to elicit user preferences, especially when there is limited user history. One way to get more information is to present clarifying questions to the user. However, generating effective sequential clarifying questions across various domains remains a challenge. To address this, we introduce a novel approach for training LLMs to ask sequential questions that reveal user preferences. Our method follows a two-stage process inspired by diffusion models. Starting from a user profile, the forward process generates clarifying questions to obtain answers and then removes those answers step by step, serving as a way to add ``noise'' to the user profile. The reverse process involves training a model to ``denoise'' the user profile by learning to ask effective clarifying questions. Our results show that our method significantly improves the LLM's proficiency in asking funnel questions and eliciting user preferences effectively.

Paper Structure

This paper contains 19 sections, 10 equations, 7 figures, 2 algorithms.

Table of Contents

  1. Introduction
  2. Background
  3. Text Generation
  4. Profiling Process
  5. Diffusion Models
  6. Proposed Model
  7. Profile Reconstruction by Asking Questions
  8. Profile Corruption
  9. User Simulation
  10. Experiments
  11. Evaluation Process
  12. Results and Analysis
  13. Effect of fine tuning
  14. Effect of Number of Questions
  15. Effect of Adding Question History
  16. ...and 4 more sections

Figures (7)

  • Figure 1: Our model for addressing corrupted user profiles and reconstruction through clarifying questions.
  • Figure 2: An example of our framework for user profile processing. Starting from a complete user profile in textual form, the forward process converts it into structured JSON and sequentially generates elicitation questions while progressively removing information. The reverse process then reconstructs the profile by iteratively answering the elicitation questions
  • Figure 3: (a) BLEU and ROUGE scores for four models, showing the performance of the non-fine-tuned and fine-tuned Questioners with different user simulators. (b) Percentage of unanswered questions for models.
  • Figure 4: BLEU (left) and ROUGE (right) scores vs. number of questions.
  • Figure 5: Effect of adding questions along with answers to partial user profiles
  • ...and 2 more figures