A method for military entity relation extraction based on two-level attention mechanism

To address the challenge of limited training data for entity relation extraction in the military domain, training instances within this domain are automatically generated employing the distant supervision method. Subsequently, to mitigate the problem of noise in corpus construction, a two-level attention model for entity relation extraction (RE) is introduced. Following the bidirectional gated recurrent unit (BI-GRU) network, instance-level and word-level attention mechanisms are incorporated. This allows the model to initially capture bidirectional semantic information of the training instances through BI-GRU. Subsequently, it employs word-level attention to identify crucial words within each training instance, and introduces instance-level attention across multiple training instances, emphasizing the key examples. Experiments conducted on the automatically constructed training corpus in the military domain illustrate the model's effective entity relation extraction capabilities. Furthermore, it mitigates the influence of noise data introduced through distant supervision, thereby enhancing the accuracy of relation extraction.