Traditionally, dialogue systems are implemented using modular/pipeline architectures. However, this modular approach suffers from several drawbacks such as error accumulation, lower flexibility, or the need for multiple-level data annotation. To tackle these issues, dialogue system research has been focusing on end-to-end neural architectures in recent years. The introduction of pre-trained language models such as BERT or GPT-2 revolutionized the NLP field, including dialogue modeling. However, the pre-trained models also present considerable risks for task-oriented dialogue, such as lack of knowledge grounding or diversity. To address these issues, we introduce modified training objectives for language model finetuning, and we employ massive data augmentation via back-translation to increase the diversity of the training data. We further examine the possibilities of combining data from multiples sources to improve performance on the target dataset. We carefully evaluate our contributions with both human and automatic methods. Our model achieves state-of-the-art performance on the MultiWOZ dataset and is competitive with the DSTC9 shared task's best submissions.
***The talk will be streamed via Zoom. For details how to join the Zoom meeting, please write to sevcikova et ufal.mff.cuni.cz***