Knowledge-enhanced meta-prompt for few-shot relation extraction

Abstract

Few-shot relation extraction (RE) aims to identity and extract the relation between head and tail entities in a given context by utilizing a few annotated instances. Recent studies have shown that prompt-tuning models can improve the performance of few-shot learning by bridging the gap between pre-training and downstream tasks. The core idea of prompt-tuning is to leverage prompt templates to wrap the original input text into a cloze question and map the output words to corresponding labels via a language verbalizer for predictions. However, designing an appropriate prompt template and language verbalizer for RE task is cumbersome and time-consuming. Furthermore, the rich prior knowledge and semantic information contained in the relations are easily ignored, which can be used to construct prompts. To address these issues, we propose a novel Knowledge-enhanced Meta-Prompt (Know-MP) framework, which can improve meta-learning capabilities by introducing external knowledge to construct prompts. Specifically, we first inject the entity types of head and tail entities to construct prompt templates, thereby encoding the prior knowledge contained in the relations into prompt-tuning. Then, we expand rich label words for each relation type from their relation name to construct a knowledge-enhanced soft verbalizer. Finally, we adopt the meta-learning algorithm based on the attention mechanisms to mitigate the impact of noisy data on few-shot RE to accurately predict the relation of query instances and optimize the parameters of meta-learner. Experiments on FewRel 1.0 and FewRel 2.0, two benchmark datasets of few-shot RE, demonstrate the effectiveness of Know-MP.