Multi-Label Text Classification by Graph Neural Network with Mixing Operations

Anonymous

Abstract: Multi-label text classification is one of the fundamental tasks in natural language processing. Recently, the graph convolution network (GCN) is leveraged to boost the performance of such a task. However, the best way for label correlation modeling and feature learning with label system awareness is still unclear. This paper proposes Mix-GCN, a graph network with two mixing operations, to improve the conventional GCN framework for multi-label text classification in the following two steps. Firstly, we model the label correlations by mixing the graph built from statistical co-occurrence information and the graph constructed from prior knowledge. Secondly, we propose a mixing operation to continuously inject GCN embedding into LSTM representation learning for better label-aware representation. Experimental results on four benchmarks demonstrate that Mix-GCN significantly outperforms the state-of-the-art models and performs better in long-tail label cases.