BanglaShop-CRS：面向用户中心的孟加拉语对话推荐数据集
BanglaShop-CRS: A User-Centric Bangla Dataset for Conversational Recommendation
- University of Vermont（佛蒙特大学）
- Santa Fe Institute（圣塔菲研究所）
- Complexity Science Hub（复杂性科学中心）
机构由 AI 辅助整理，请以论文原文为准。
AI总结：
针对孟加拉语对话推荐资源匮乏问题，构建了基于真实电商行为的大规模合成数据集BanglaShop-CRS，通过整合用户历史与反馈生成多轮对话，验证了对话上下文和微调对推荐质量的提升，并提供人工与自动评估基准。
AI中文摘要：
对话推荐系统（CRS）使用户能够通过自然语言交互表达偏好、约束和反馈。然而，现有的CRS资源主要集中在英语和其他高资源语言，导致孟加拉语及孟加拉语-英语混合代码切换场景的代表性不足。为解决这一差距，我们引入了BanglaShop-CRS，一个基于真实电子商务行为的大规模用户中心合成孟加拉语对话推荐数据集。该数据集包含27,178个多轮对话、274,802个话语和360万词元，覆盖10个产品领域。我们的生成流程整合了用户购买历史、正面和负面反馈以及评论文本，以保持对话内容与用户偏好之间的一致性。我们在目录受限和开放词汇推荐协议下评估了BanglaShop-CRS。结果表明，对话上下文能提升推荐质量，而微调在各类模型中带来进一步提升。由五位母语为孟加拉语的标注者进行的人工评估确认了对话的流畅性、信息量、逻辑性和连贯性，且在所有维度上具有显著一致性。事实基础评估进一步显示，与正确用户记录的对齐强于打乱后的记录，具有显著的标注者间一致性（κ=0.65），且人类与GPT-5.1的判断相当。BanglaShop-CRS为推进孟加拉语对话推荐提供了一个可扩展的基准。
英文摘要：
Conversational recommender systems~(CRS) enable users to express preferences, constraints, and feedback through natural language interaction. However, existing CRS resources are concentrated in English and other high-resource languages, leaving Bangla and code-mixed Bangla--English settings underrepresented. To address this gap, we introduce BanglaShop-CRS, a large-scale user-centric synthetic Bangla conversational recommendation dataset grounded in real e-commerce behavior. It contains 27,178 multi-turn dialogues, 274,802 utterances, and 3.6M tokens across 10 product domains. Our generation pipeline incorporates user purchase histories, positive and negative feedback, and review texts to maintain consistency between dialogue content and user preferences. We evaluate BanglaShop-CRS under catalog-constrained and open-vocabulary recommendation protocols. Results show that dialogue context improves recommendation quality, while fine-tuning yields further gains across models. Human evaluation by five native Bangla-speaking annotators confirms the fluency, informativeness, logicality, and coherence of the dialogues, with significant agreement across all dimensions. Factual-grounding evaluation further shows stronger alignment with correct than shuffled user records, with substantial inter-annotator agreement ($κ=0.65$) and comparable human and GPT-5.1 judgments. BanglaShop-CRS provides a scalable benchmark for advancing conversational recommendation in Bangla.
