DGMED: A Novel Document-Level Graph Convolution Network for Multi-Event Detection

Abstract

Online news documents can contain thousands of characters and tens of events. To detect events in these documents, it is important to construct long-range context information. Such information, however, is not effectively created in existing event detection methods including DMBERT, MOGANED. As a result, these methods show poor event detection accuracy in production where long documents are common. To address this, this paper proposes a Document-level Graph convolution network for Multi-event Detection (DGMED). DGMED represents each sentence in a long document as a graph, and it interconnects these graphs using novel cross-sentence global neural network nodes. These nodes allow DGMED further construct accurate document-level contextual information, thus accurately extracting multiple events as required. We evaluate DGMED using a public event extraction dataset (i.e., Maven) and a production large-scale dataset (named AML). Evaluation results show that DGMED can out-perform state-of-the-art methods BERT+CRF and BiLSTM+CRF up to 0.7% in Maven and 5.7% in AML.

Tasks

Reproductions