A Prototype Network Enhanced Relation Semantic Representation for Few-shot Relation Extraction

Abstract Few-shot relation extraction is one of the current research focuses. The key to this fully extract semantic information through very little training data. Intuitively, raising semantics awareness in sentences can improve efficiency model features alleviate overfitting problem few-shot learning. Therefore, we propose an enhanced feature based on prototype network relations from texts. Firstly, design a multi-level embedding encoder with position and Transformer, which uses local text enhance representation. Secondly, encoded are fed into novel network, designs method that utilizes query prototype-level attention guide supporting prototypes, thereby enhancing prototypes representation better classify sentences. Finally, experimental comparison discussion, prove analyze effectiveness proposed encoder, stability model. Furthermore, our has substantial improvements over baseline methods.