A Student-Teacher Architecture for Dialog Domain Adaptation Under the Meta-Learning Setting

Numerous new dialog domains are being created every day while collecting data for these is extremely costly since it involves human interactions. Therefore, essential to develop algorithms that can adapt different efficiently when building data-driven models. Most recent research on domain adaption focuses giving the model a better initialization, rather than optimizing adaptation process. We propose an efficient adaptive task-oriented system model, which incorporates meta-teacher emphasize impacts between generated tokens with respect context. first train our base and adversarially in meta-learning setting rich-resource domains. The learns quantify importance of under contexts across During adaptation, guides focus important order achieve efficiency. evaluate two multi-domain datasets, MultiWOZ Google Schema-Guided Dialogue, state-of-the-art performance.