Enhancing Prototype Networks for Relation Classification with Multi-Feature and Confusing Loss Function in Low-resource scenarios

Few-shot relation classification (FSRC) is a task that aims to recognize novel relations using a limited set of labeled instances. Numerous approaches have employed meta-learning to tackle the challenge. Despite the impressive results achieved through these approaches, existing models remain deficient when tasked with confusing relations that are closely resemble each other. We consider that this is largely owing to the existing models not being able to capture the informative features in the instance well. In this paper, we obtain Multi-Feature information by utilizing relation name and relation description information to obtain better prototype network representation. Meanwhile, we introduce a useful loss function that aims to enhance the model's capacity to differentiate between real relation and those that lead to confusion. Our proposed model was subjected to extensive experimentation on the FewRel1.0 dataset to assess its performance. The results demonstrate that our approach achieves comparable, and in some cases, significantly better accuracy than strong baselines. Overall, this study shows the outstanding performance of our approach and highlights its potential to improve relation classification tasks.