A Novel Knowledge Enhanced Network for Few-Shot Combat Intention Recognition

To address the challenges of limited interpretability and scarce battlefield situational data in current deep learning-based models for combat intention recognition (CIR), this paper proposes a novel knowledge enhanced network (KEN). The proposed KEN formalizes expert knowledge using knowledge graph. By integrating the knowledge representation module and knowledge alignment module, the proposed network can introduce knowledge into any recognition model. Leveraging the prior knowledge, our KEN alleviates the performance limitations resulting from data scarcity. Furthermore, the knowledge alignment module could present an intuitive visualization of the degree of attention that samples allocate to different knowledge elements, thereby greatly enhancing the interpretability of the model. Experimental results show the superiority of our KEN, e.g. achieving around 95.63% accuracy on typical combat scenarios, outperforming other baseline models.