中国科学院自动化研究所机构知识库

The dialogue system is widely used in many application scenarios, while the construction of the dialogue system always faces the difficulty of zero-resource training data. To alleviate that, we propose a knowledge transfer framework to build a dialogue system based on existing machine translators and training data in data-rich language. Specifically, we first generate various kinds of pseudo data with cyclic translation procedure and different data combinations. Then we propose a noise injection method and a multi-task training method for the pipeline system and end-to-end system, respectively. The noise injection method optimizes each module by incorporating machine translation noises into the pipeline process to handle the error propagation problem, thus improving the whole system's robustness. The multi-task training method combines cross-lingual dialogue, monolingual dialogue, and machine translation into the end-to-end dialogue system's training process, thus reducing the impact of noises in pseudo data. The extensive experiments on a real-world e-commerce dataset demonstrate that our methods can achieve remarkable improvements over strong baselines.