BDBRC: A Chinese military entity recognition model combining context contribution and residual dilatation convolutional networks

Abstract

Military information is gradually overloaded due to the diversity of sources and the exponential growth in quantity, which greatly affects the accuracy of intelligence personnel in extracting and analyzing military information. The modern warfare approach has also evolved from the traditional physical domain to the cognitive domain, and competing for advantages in the cognitive domain has become a key objective of combat. Therefore, constructing domain knowledge graphs and mining the relationships between data play an important role in cognitive domain analysis. In this paper, we propose a convolutional network recognition method based on improved two-layer bi-directional BiLSTM networks named the BERT-α BiLSTMs-RECNN-CRF (BDBRC) model. For the difficulties of military entities generally having long names and low extraction accuracy, as well as the existence of a large number of composite entities that are difficult to recognize, an improved two-layer BiLSTM model is devised first. In view of the fact that the BiLSTMs model always extracts features equally in long-distance text sequences without actually considering the different influences of different sentence contexts, contribution factor a is added to extract the contribution of the above and below to the target entity in different sentences respectively. Then, aiming at the strong problem of the domain of military news texts and the high level of inter-entity ambiguity, we propose a method that utilizes a modified convolutional network (RECNN) for partial feature extraction and jointly with a modified two-layer BiLSTM network for entity recognition. The experiment on the self-constructed dataset shows that the F1 value of the model proposed in this paper reaches 93.18%, and the F1 value, P, and H of our model are all better than the baseline model, which verifies the performance of the model. At the same time, we use public data sets MSRA and CLUB2020, and the experimental results show that the model proposed in this paper also has a good performance in the public data set, verifying the universality of the model. It can provide methodological support for the construction of the military knowledge graph.