Few-shot relation extraction method based on self-supervised learning prototypical networks

Few-shot relation extraction (FSRE) is a key task in Information extraction that aims to train models using a small number of labeled samples to identify relation categories in sentences. Existing research widely adopts meta-learning approaches, with prototypical networks being popular due to their simplicity and efficiency. However, these methods mainly focus on distinguishing samples of different categories, establishing inter-class boundaries based solely on simple calculations of relation category samples, while ignoring the differences between instances and failing to fully utilize instance feature information. To address this issue, this paper proposes self-supervised learning prototypical networks (SSLPN), a self-supervised learning-based prototypical network method for FSRE. This model utilizes self-supervised learning to enhance sample discriminability and effectively learn category representations. SSLPN consists of three stages, and this three-stage architecture synergistically enhances sample discriminability, mitigates overfitting, and optimizes model performance for FSRE tasks. The FewRel1.0 dataset was used to evaluate SSLPN, and the outcomes demonstrated that the proposed method considerably enhanced the FSRE performance, confirming its efficacy.