2019/11/26 by Bo-Hsiang Tseng, Tseng, Bo-Hsiang, Marek Rei +9
Computer Science · #AI in Service Interactions #Computation and Language (cs.CL) #FOS: Computer and information sciences #Natural Language Processing Techniques #Speech and dialogue systems #Topic Modeling
paper · pdf · doi:10.48550/arxiv.1911.11672
openalex publication_date 2019/11/26 · openalex created_date 2022/07/26 · openalex updated_date 2026/07/28
Dialogue systems benefit greatly from optimizing on detailed annotations,\nsuch as transcribed utterances, internal dialogue state representations and\ndialogue act labels. However, collecting these annotations is expensive and\ntime-consuming, holding back development in the area of dialogue modelling. In\nthis paper, we investigate semi-supervised learning methods that are able to\nreduce the amount of required intermediate labelling. We find that by\nleveraging un-annotated data instead, the amount of turn-level annotations of\ndialogue state can be significantly reduced when building a neural dialogue\nsystem. Our analysis on the MultiWOZ corpus, covering a range of domains and\ntopics, finds that annotations can be reduced by up to 30 % while maintaining\nequivalent system performance. We also describe and evaluate the first\nend-to-end dialogue model created for the MultiWOZ corpus.\n