FIRM: A Fused Information Re-optimization Method applied to the classification of defense science and technology indicators

Recent years have witnessed a growing interest in applying few-shot learning algorithms to personalized recommendation, and text classification tasks. In this paper, we propose a Fused Information Re-optimization Method based few-shot learning algorithm (FIRM) which can realize the fusion of label description information and sample instances. With the fusion strategy, FIRM can fully use the commonality between categories, the characteristic information of each category, and training model parame-ters with both generalization ability and classification characteristics. Further, to evaluate the validity of the proposed algorithm on the defense domain, a few-shot dataset (i.e., the defense science and technology indicators dataset) is constructed. Experimental results on the dataset validate our proposal and indicate the superior capability in classification tasks than a prototype network and MAML. Our main contributions are: 1) A few-shot dataset in the field of defense science and technology is con-structed by manually extracting the indicators from a large number of defense science and technology project evaluation documents. 2) A method of fusing information re-optimization is proposed to solve the indicator few-shot classification problem. 3) The effectiveness of FIRM is verified through comparative experiments on the constructed dataset, against with the prototype network, MAML algorithm.