IRWOZ 2.0 is a comprehensive upgrade of the original IRWOZ dataset, targeting higher accuracy in industrial human‑robot interaction (HRI) dialogue state tracking. The first version suffered from substantial noise in states and utterances, which limited model performance.
The upgrade leverages large language models (Mistral and Claude‑3.5) for automated generation, complemented by manual corrections and automated typo removal. The resulting corpus contains 390 complete dialogues across four industrial domains: Assembly, Delivery, Positioning, and Relocation.
Benchmark experiments on dialogue state tracking using GPT‑2 show the BLEU‑4 score rising from 0.1651 in the original set to 0.5604 after the improvements, demonstrating the strong impact of higher‑quality annotations.
IRWOZ 2.0 is publicly released on IEEE DataPort, and can be accessed at https://ieee-dataport.org/documents/irwoz-20-large-language-model-driven-dialogue-dataset-industrial-robot-conversations.
Review: By combining LLM‑driven generation with meticulous human refinement, the dataset markedly reduces noise and offers a robust benchmark for industrial robot dialogue research.