From Reviews to Dialogues: Active Synthesis for Zero-Shot LLM-based Conversational Recommender System

TL;DR

This paper tackles the data-scarcity challenge in conversational recommender systems by introducing an active data augmentation framework that leverages non-conversational seed data, active sample selection, and synthetic data generation via a black-box LLM. By prioritizing informative and diverse seeds and using query templates, the approach produces high-quality synthetic dialogues used to fine-tune smaller CRS models, addressing no- or low-resource domains. Across ReDial and INSPIRED benchmarks, the method outperforms zero-shot and GPT-generated baselines and demonstrates that external signals (metadata and collaborative interactions) further boost performance, while offering budget-efficient data generation strategies. The work shows that domain-specific synthetic data can meaningfully substitute or complement in-domain data, enabling scalable, privacy-conscious CRS deployment.

Abstract

Conversational recommender systems (CRS) typically require extensive domain-specific conversational datasets, yet high costs, privacy concerns, and data-collection challenges severely limit their availability. Although Large Language Models (LLMs) demonstrate strong zero-shot recommendation capabilities, practical applications often favor smaller, internally managed recommender models due to scalability, interpretability, and data privacy constraints, especially in sensitive or rapidly evolving domains. However, training these smaller models effectively still demands substantial domain-specific conversational data, which remains challenging to obtain. To address these limitations, we propose an active data augmentation framework that synthesizes conversational training data by leveraging black-box LLMs guided by active learning techniques. Specifically, our method utilizes publicly available non-conversational domain data, including item metadata, user reviews, and collaborative signals, as seed inputs. By employing active learning strategies to select the most informative seed samples, our approach efficiently guides LLMs to generate synthetic, semantically coherent conversational interactions tailored explicitly to the target domain. Extensive experiments validate that conversational data generated by our proposed framework significantly improves the performance of LLM-based CRS models, effectively addressing the challenges of building CRS in no- or low-resource scenarios.

Figures (8)

• Figure 1: Comparison of CRS under different settings.
• Figure 2: Overview of our proposed Active Data Augmentation framework for conversational recommendation. Our pipeline starts with a seed dataset $\mathcal{D}$, which contains metadata, reviews, and user-collaborative signals. Given a selection budget $B$, we first extract features and then apply active learning techniques (see \ref{['alg:active_interface']}) to select the most diverse and representative samples. These samples are then used together with a query template set $\mathcal{Q}_{\mathrm{tmp}}$ to generate synthetic conversational data $\mathcal{S}$ using LLM $\mathcal{M}$ (see \ref{['alg:synthetic_data_creation']}). Finally, the generated synthetic data $\mathcal{S}$ is employed to fine-tune a language model $\pi_\theta$ using supervised fine-tuning( \ref{['sec:finetuning']}).
• Figure 3: Performance metrics for LlaMA3-1B across Redial and Inspired datasets.
• Figure 4: Performance metrics for LlaMA3-3B across Redial and Inspired datasets.
• Figure 5: Performance metrics for Gemma2-2B across Redial and Inspired datasets.