Improving Semantic Dependency Parsing with Higher-Order Information Encoded by Graph Neural Networks

Higher-order information brings significant accuracy gains in semantic dependency parsing. However, modeling higher-order is non-trivial. Graph neural networks (GNNs) have been demonstrated to be an effective tool for encoding many graph learning tasks. Inspired by the success of GNNs, we investigate improving parsing with encoded multi-layer GNNs. Experiments are conducted on SemEval 2015 Task 18 dataset three languages (Chinese, English, and Czech). Compared previous state-of-the-art parser, our parser yields 0.3% 2.2% improvement average labeled F1-score English in-domain (ID) out-of-domain (OOD) test sets, 2.6% Chinese ID set, 2.0% 1.8% Czech OOD sets. Experimental results show that outperforms best one languages. The outstanding performance demonstrates GNNs exceedingly beneficial SDP.