Joint entity and relation extraction with table filling based on graph convolutional Networks

Information extraction involves extracting structured information from text, which can be categorized into named entity recognition (NER) and relation extraction (RE). Recent successful works address the dependency between NER and RE using a filling table approach, but they overlook the long-distance dependencies between cells in the table, which may not adequately understand the semantics between distant elements in the text. To address this limitation, we introduce an innovative approach known as Table Filling with Graph Convolutional Networks (TableF-GCN) for joint NER and RE. First, TableF-GCN constructs a graph that captures the interaction between cells within the table. Additionally, it leverages GCNs to encode the long-distance dependency within the graph through multi-hop propagation, thereby enabling the acquisition of local and global contextualized embeddings for each node in the graph. Experimental results are conducted on joint NER and RE extraction, ablation study, hyper parameter setting, and a case study on three widely utilized datasets. The comparative analysis demonstrates TableF-GCN outperforms the baselines in recall, but not necessarily in precision. Fortunately, when consolidating the outcomes from the ConLL04 and ADE datasets, TableF-GCN showcases overall improvement through statistical analysis.