Hybrid graph structure learning for improving semantic dependency parsing with robust graph neural networks

Graph neural networks (GNN) can learn powerful representations of graphs, but their effectiveness is heavily influenced by the quality of the graph structure used. Graph structure learning (GSL) was developed to extract useful graph structures from data and learn better graph representations using GNN. GSL creates an adjacency matrix with two multilayer perceptrons and a biaffine attention network. However, it does not directly incorporate the principles of homophily, sparsity, and degree distribution, which are important in many real-life situations. Differentiable graph structure learning neural networks (DGSLN), which adhere to these basic principles, have been proven to improve GNN performance. This study proposes a GNN-based semantic dependency parsing model that incorporates the GSL adapted from the DGSLN approach to generate an initial graph. For the graph representation learning stage, we use GNN variants, GIN and GATv2, which have shown robust performance in previous research. The proposed model significantly improves previous best models, achieving an average F1 score of 93.77% on in-domain and 92.27% on out-of-domain datasets. However, our model's parsing performance is slower than prior top-performing models. The model utilizing GIN exhibits the highest level of performance and shows exceptional performance in semantic dependency parsing on English datasets.