Chinese Dialogue Corpora + DSTC9

Good data is an environment. Each of these corpora was built so that other people could train against it, evaluate on it, and compare. Six years on they are still in use: as of late 2026 the Hugging Face copies of LCCC, KdConv, and CrossWOZ are each still downloaded hundreds to a thousand times a month.

• CrossWOZ (TACL 2020) : the first large-scale Chinese cross-domain Wizard-of-Oz task-oriented dataset. 6K dialogues, 102K utterances, 5 domains (hotel, restaurant, attraction, metro, taxi), dialogue state and dialogue act annotation on both user and system side. About 60% of dialogues have cross-domain goals. Ships with a user simulator and pipelined benchmark models, so it doubles as an RL environment for dialogue policy.
• KdConv (ACL 2020) : a multi-domain knowledge-driven dialogue dataset grounding 4.5K conversations (86K utterances, 19 turns on average) in knowledge graphs across film, music, and travel. Comes with benchmark models that show how much background knowledge helps.
• LCCC and CDial-GPT (NLPCC 2020, best student paper) : a 6.8M / 12M dialogue corpus produced by a rigorous cleaning pipeline (rules plus a classifier trained on 110K manually annotated pairs), with pre-trained dialogue models released on Hugging Face. This is the most starred of the three.
• DSTC9 Track 2 (AAAI 2021 workshop, TASLP 2024 overview): with Microsoft Research and Tsinghua, I co-organized the multi-domain task-completion challenge, which added an end-to-end task-completion task and a cross-lingual dialogue state tracking task built on CrossWOZ and MultiWOZ.

What I did. Worked across the whole pipeline with the team: collected the initial seed data from scratch, took part in the human data collection, and helped clean the corpora and train the benchmark models. Second author on CrossWOZ, co-author on KdConv and LCCC, and co-organizer of DSTC9 Track 2 with Microsoft Research. LCCC received the NLPCC 2020 Best Student Paper Award.