AI Papers Reader

Personalized digests of latest AI research

View on GitHub

Synthetic Conversations from Reviews Can Jump-Start Recommender Chatbots

Conversational recommender systems are supposed to suggest movies or products while chatting, but building them usually requires domain-specific dialogue logs that are costly to collect, privacy-sensitive, and simply missing in most new categories. In a study posted on arXiv, researchers asked whether those logs can be bootstrapped from material that is abundant: product reviews, item metadata, and records of which users liked which items.

The team calls the approach zero-data bootstrapping. From Amazon’s Movies & TV category, they encoded candidate items, selected a subset using one of two information-theoretic criteria—Jensen-Shannon diversity, which favors seed items spread across the data, or Fisher information, which favors items expected to add new parameter-level information—and compared those with random sampling and popularity. For each selected item, the teacher model GPT-4o read sampled reviews, wrote a user query in the informal style of Reddit recommendation threads, and then produced a list of 20 suggested titles. Those synthetic query–recommendation pairs were used to fine-tune smaller, open-weight models—Qwen2.5-1.5B, Llama3.2-3B, and Qwen3-4B—which were tested on the ReDial and INSPIRED movie benchmarks.

A synthetic query in the paper illustrates the format: a user asks for “a quirky, whimsical movie with a great ensemble cast and a bit of adventure,” says The Brothers Bloom sounds appealing, and receives The Grand Budapest Hotel, Big Fish, and others in reply.

The main experimental result is that grounding synthetic dialogues in domain signals matters. Fully fine-tuned Qwen2.5-1.5B improved Recall@1, the rate at which the correct recommendation is ranked first, on INSPIRED by 207.8% relative to zero-shot prompting, while naĂŻve GPT-generated data without seed grounding improved only 18.8%. Smaller models gained more: on ReDial the same model improved Recall@1 by 40.6% under full fine-tuning and 41.8% with LoRA, a low-rank adaptation method, while the larger Qwen3-4B gained 18.3% and 9.0%, respectively.

Active selection also improved data efficiency: both criteria beat random and popularity-based choice at the same teacher-call budgets, and adding metadata or collaborative-filtering signals helped further. The benefits carry outside LLMs: a neighborhood-based recommender, NBCRS, scored 0.39% Recall@5 on ReDial with raw seed data but 11.52% after being trained on the synthetic conversations.

The synthetic data was especially valuable when real dialogues were scarce. On INSPIRED’s roughly 1,000 dialogues, a model trained only on synthetic data outperformed one trained on the real conversations; combining synthetic with real data gave further gains. On the larger ReDial corpus, however, adding synthetic data slightly reduced performance.

Caveats remain. The experiments are movie-only, so transfer to music or e-commerce is not established, and the authors note that LLM-generated dialogue can contain hallucinations or stylistic artifacts. In fact, naïve synthetic lists matched known movie titles more often than grounded lists, 92.3% versus 76.5%, even though they trained worse models—a reminder that surface plausibility is not the same as useful supervision.

The work does not claim synthetic data should replace abundant real conversations. Its significance is narrower and practical: with systematic selection and domain grounding, non-conversational signals may be enough to jump-start a recommender in a domain where no one has yet transcribed a single dialogue.