Domain adaptation with augmented space method for multi-domain contact center dialogue summarization

In this paper we propose a method to improve the quality of extractive summarization for contact center dialogues in various domains by making use of training samples whose domains are different from that of the test samples. Since preparing sufficient numbers of training samples for each domain is too expensive, we leverage references from many different domains and employ the Augmented Space Method to implement domain adaptation. As the target of summarization, we take up contact center dialogues in six domains and summarize their transcripts. Our experiment shows that the proposed method achieves better results than the usual supervised learning approach.