BERT-DST

• Dialogue State Tracking

• Chao, Guan-Lin, and Ian Lane. "Bert-dst: Scalable end-to-end dialogue state tracking with bidirectional encoder representations from transformer." arXiv preprint arXiv:1907.03040 (2019). [Paper]

• Dataset

  • KLUE-DST

    • domain-key-value => slot-key(domain-key), slot-value(value)
    • slot-key => Class defined in Ontology (45 classes in total)
    • slot-value => Values in the dataset

• Experiment

  • Classification Module (using [CLS] token)

    • 45 slot-keys defined in KLUE-DST and 5 slot-types(none, dontcare, yes, no, span) defined for experiments
    • When predicted with the rest of the classes except none and span, the slot-value is predicted with the rest of the classes (dontcare, yes, no)
    • When predicted as a class with span, the slot-value is determined by the Span Prediction Module
    • When predicted as a class with none, it means there is no slot-key and no slot-value

  • Span Prediction Module (using sequence tokens)

    • When span is predicted in the Classification Module, the slot-value in the text is predicted through start token and end token

  • Result through Validation set

    • Slot F1: 0.908588
    • Joint-Goal Accuracy: 0.473283