STIL -- Simultaneous Slot Filling, Translation, Intent Classification, and Language Identification: Initial Results using mBART on MultiATIS++

Slot-filling, Translation, Intent classification, and Language\nidentification, or STIL, is a newly-proposed task for multilingual Natural\nLanguage Understanding (NLU). By performing simultaneous slot filling and\ntranslation into a single output language (English in this case), some portion\nof downstream system components can be monolingual, reducing development and\nmaintenance cost. Results are given using the multilingual BART model (Liu et\nal., 2020) fine-tuned on 7 languages using the MultiATIS++ dataset. When no\ntranslation is performed, mBART's performance is comparable to the current\nstate of the art system (Cross-Lingual BERT by Xu et al. (2020)) for the\nlanguages tested, with better average intent classification accuracy (96.07%\nversus 95.50%) but worse average slot F1 (89.87% versus 90.81%). When\nsimultaneous translation is performed, average intent classification accuracy\ndegrades by only 1.7% relative and average slot F1 degrades by only 1.2%\nrelative.\n