Unsupervised Domain Adaptation for Dialogue Sequence Labeling Based on Hierarchical Adversarial Training

This paper presents a novel unsupervised domain adaptation method for dialogue sequence labeling. Dialogue sequence labeling is a supervised learning task that estimates labels for each utterance in the given dialogue document, and is useful for many applications such as topic segmentation and dialogue act estimation. Accurate labeling often requires a large amount of labeled training data, but it is difficult to collect such data every time we need to support a new domain, such as contact centers in a new business field. In order to solve this difficulty, we propose an unsupervised domain adaptation method for dialogue sequence labeling. Our key idea is to construct dialogue sequence labeling using labeled source domain data and unlabeled target domain data so as to remove domain dependencies at utterance-level and dialogue-level contexts. The proposed method adopts hierarchical adversarial training; two domain adversarial networks, an utterance-level context independent network and a dialogue-level context dependent network, are introduced for improving domain invariance in the dialogue sequence labeling. Experiments on Japanese simulated contact center dialogue datasets demonstrate the effectiveness of the proposed method.