Abstract: Event extraction is one of the important precedent tasks for knowledge graphs, while as a sub-task, event argument extraction has a significant impact on the quality of event extraction. The existing pipelined event extraction approaches usually ignore the relationships between triggers and arguments, or among arguments, which leads to low quality of event extraction. To solve this issue, this paper proposes a bidirectional gated recurrent neural network (Bi-GRU)-based event argument extraction approach. The proposed approach considers Bert-based word vector, word part-of-speech, word position, and trigger types as features, applies Bi-GRU to encode these features of each word in a sentence to get a word vector, leverages the improved multi-attention mechanism to assign weights to different parts of the sentence, and finally identify arguments and their roles in the sentence by a full-connection layer. Experiments are conducted on a benchmark dataset, and experimental results show that argument recognition and role classification tasks achieve 69.2% and 61.6% in F1-score respectively, and are better than compared state-of-the-art approaches.