This paper focuses on the task of emotion-based CRSs (conversational recommendation systems), which aim to infer users’ implicit preferences from their emotional conversations without a clearly defined objective, in order to recommend better items that satisfy users’ needs. Previous work mostly studies factoid-based CRSs, which address cases with explicit needs and capture users’ preferences based on expressed entities or attributes. However, in real-world applications, most users express their needs through implicit emotions and feelings, without clearly stated entities or attributes. Existing entity-matching-based methods struggle in this scenario of vague requirements. To address this problem, we propose a novel model that is capable of inferring users’ preferences from their subjective expressions. Specifically, we first apply a multifaceted augmentation technique to supplement the missing background knowledge in the conversation, including relevant subjective and objective facts and relations, in order to fully grasp users’ implicit needs. Based on this enhanced knowledge, we then construct a user preference tree to capture the relationships between emotions and item attributes. The tree is updated gradually based on users’ feedback in each round of conversation. Experiments on two popular datasets demonstrate the effectiveness of our approach.