The abundant information resources available on the internet about defense technology are of vital importance as data sources for obtaining high-value military intelligence. The aim of open information extraction in the field of defense technology is to extract structured triplets containing subject, predicate, object, and other arguments from the massive amount of information available on the internet. This technology has important implications for ontology induction and the construction of knowledge graphs in the defense technology domain. However, while information extraction experiments in the general domain yield good results, open information extraction in the defense technology domain faces several challenges, such as a lack of domain annotated data, arguments overlapping unadaptability, and unrecognizable long entities.

In this paper, an annotation strategy is proposed based on the entity boundaries, and an annotated dataset in the defense technology field combined with the experience of domain experts was constructed. Furthermore, a two-stage open information extraction method is proposed in the defense technology field that utilizes a pretrained language model-based sequence labeling algorithm to extract predicates and a multihead attention mechanism to learn the prediction of argument boundaries. In the first stage, the input sentence was converted into an input sequence<[CLS], input sentence[SEP]>, and the input sequence was encoded using a pretrained language model to obtain an implicit state representation of the input sequence. Based on this sentence representation, a conditional random field (CRF) layer was used to predict the position of the predicates, i.e., to predict the BIO labels of the words. In the second stage, the predicated predicates from the first stage were concatenated with the original sentence and converted into an input sequence<[CLS], predicate[SEP], and input sentence[SEP]>, which was encoded using a pretrained language model to obtain an implicit state representation of the input sequence. This representation was then fed to a multihead pointer network to predict the position of the argument. The predicted position was tagged with the actual position to calculate the cross-entropy loss function. Finally, the predicates and the arguments predicted by the predicate and argument extraction models were combined to obtain the complete triplet.

The experimental results from the extensive experiments conducted on a self-built annotated dataset in the defense technology field reveal the following. (1) In predicate extraction, our method achieved a 3.92% performance improvement in the F1 value as compared to LSTM methods and more than 10% performance improvement as compared to syntactic analysis methods. (2) In argument extraction, our method achieved a considerable performance improvement of more than 16% in the F1 value as compared to LSTM methods and about 11% in the F1 value as compared to the BERT+CRF method.

The proposed two-stage open information extraction method can overcome the challenge of arguments overlapping unadaptability and the difficulty of long-span entity extraction, thus improving the shortcomings of existing open information extraction methods. Extensive experimental analysis conducted on the self-built annotated dataset proved the effectiveness of the proposed method.