Moving from limited-domain natural language generation (NLG) to open domainis difficult because the number of semantic input combinations growsexponentially with the number of domains. Therefore, it is important toleverage existing resources and exploit similarities between domains tofacilitate domain adaptation. In this paper, we propose a procedure to trainmulti-domain, Recurrent Neural Network-based (RNN) language generators viamultiple adaptation steps. In this procedure, a model is first trained oncounterfeited data synthesised from an out-of-domain dataset, and then finetuned on a small set of in-domain utterances with a discriminative objectivefunction. Corpus-based evaluation results show that the proposed procedure canachieve competitive performance in terms of BLEU score and slot error ratewhile significantly reducing the data needed to train generators in new, unseendomains. In subjective testing, human judges confirm that the procedure greatlyimproves generator performance when only a small amount of data is available inthe domain.
展开▼
