Conversational recommendation identification with internal knowledgeable prompt-tuning

Abstract

Recommendation systems have evolved from static preference modeling to interactive paradigms, where conversational recommendation enable dynamic preference elicitation through multi-turn dialogue. Most existing works in conversational recommendation primarily focus on identifying what items to recommend during interactions, essentially framing the task as a form of static personalized recommendation. However, with the rapid development of Large Language Models (LLMs), user-service interactions increasingly involve casual, exploratory, or non-goal-oriented dialogue, which raises a critical and emerging challenge: determining when to provide recommendations during a conversation, a task known as conversational recommendation identification. Although there have already been some efforts on devoting prompt-tuning techniques to this task, most existing methods rely heavily on general-purpose external knowledge, which often fails to capture conversation-specific semantics and may suffer from domain mismatch, semantic ambiguity, or noise. To address these limitations, we propose a novel conversational recommendation identification method based on internal knowledgeable prompt-tuning, which enhances soft prompt construction by integrating a self-resource knowledge expansion mechanism that captures task-specific conversational features. Instead of depending on the implicit knowledge of pre-trained models or external resources, we retrieve contextual knowledge in a self-supervised manner to guide verbalizer construction. The resulting soft prompts are tailored to guide conversational recommendation identification, which effectively balance automatic template generation with improved recognition performance. Extensive experiments on both English and Chinese datasets demonstrate that our method consistently outperforms all baselines, including state-of-the-art LLMs, in terms of identification accuracy and robustness.