Multilingual Intent Classification in Customer Service (MICCS)
收藏资源简介:
该数据集由J&T Express与上海科技大学联合构建,是一个基于真实物流客服日志的多语言意图分类基准,包含约3万条经过脱敏处理的独立用户查询。数据源自60万条历史记录,经规则过滤、大语言模型辅助质量控制及人工验证,构建为包含13个父类别和17个子类别的双层分类体系,涵盖英语、西班牙语、阿拉伯语等主要语言,并支持印尼语、中文等语言的零样本评估。其独特价值在于保留原生查询的噪声分布和语言特性,通过配对提供机器翻译与原生测试集,直接量化合成数据与真实场景的评估差距,主要应用于多语言客户服务系统的意图理解与路由优化。
This dataset was co-developed by J&T Express and ShanghaiTech University, and serves as a multilingual intent classification benchmark based on real logistics customer service logs. It contains approximately 30,000 anonymized independent user queries derived from 600,000 historical records after undergoing rule-based filtering, LLM-assisted quality control, and manual verification. It is structured into a two-tier classification system with 13 parent categories and 17 sub-categories, covering major languages including English, Spanish and Arabic, and supports zero-shot evaluation for languages such as Indonesian and Chinese. Its unique value lies in retaining the noise distribution and linguistic features of native user queries, and provides paired machine-translated and native test sets to directly quantify the evaluation gap between synthetic data and real-world scenarios. It is primarily applied to intent understanding and routing optimization for multilingual customer service systems.




