NatCS: Eliciting Natural Customer Support Dialogues

Summary

AWS AI Labs dataset of synthetic H2H customer-service conversations collected with discourse/spoken-form complexities observed in real calls. Two methods: NatCSSelf (written-as-spoken self-dialogues) and NatCSSpoke (paired recorded then transcribed). Closer than MultiWOZ/SGD/MultiDoGO/Taskmaster to real retail/finance transcripts on turn length, perplexity, intent diversity, and human realism/spoken-likeness. Subset labeled with TOD dialogue acts (InformIntent/ElicitSlot/etc.) and open intent/slot schemas. DA classifiers trained on NatCS transfer to real data better than SGD (F1 54.1 vs 31.6). Also used as DSTC11 Track 2 intent-induction resource.

Keywords

natcs · customer-support · spoken-dialogue · dstc11 · intent-induction · dialogue-acts · aws

Topics

task-oriented dialogue, customer support, spoken conversation

Research notes

• Primary: arxiv abs/html (cs.CL, Accepted to Findings of ACL 2023). Data+baselines https://github.com/amazon-science/dstc11-track2-intent-induction (Apache-2.0). Related DSTC11 overview arXiv 2304.12982. HF paper page links unofficial splevine/dstc11-intent (4205 rows); not an official NatCS card. Discord posted PDF. Dataset not in datasets_local.csv.