Adversarial Domain Adaptation For Chinese Semantic Dependency Graph Parsing

The Chinese Semantic Dependency Graph (CSDG) Parsing reveals the deep and fine-grained semantic relationship of Chinese sentences, and the parsing results have a great help to the downstream NLP tasks. However, most of the existing work focuses on parsing in a single domain. When transferring to other domains, the performance of the parser tends to drop dramatically. And the target domain often lacks the annotated data, so it is difficult to train the parser directly in the target domain. To solve this problem, we propose a lightweight yet effective domain adaptation component for CSDG parsing that can be easily added to the architecture of existing single domain parser. It contains a data sampling module and an adversarial training module. Furthermore, we present CC SD, the first Chinese Cross-domain Semantic graph Dependency dataset. Experiments show that with the domain adaptation component we proposed, the model can effectively improve the performance in the target domain. On the CCSD dataset, our model achieved state-of-the-art performance with significant improvement compared to the strong baseline model.