Decomposed Meta-Learning for Few-Shot Relational Triple Extraction

Relational triple extraction is an important task in the field of knowledge extraction, which aims at automatically identifying relational triples from plain text. Traditional supervised learning methods usually require a large amount of labeled data, and their performance degrades significantly in a few-shot learning setting. To address this issue, we propose the Decomposed Meta-Learning framework for few-shot relational triple extraction (DMFTE). Firstly, to classify the relations occurred in a sentence, the representation of the support instances is strengthened by an attention module, which later is leveraged for capturing the connection between the original prototypes and the query instances. And then based on Prototype Network algorithm, we utilize the information of the relations to provide semantic guidance to the original prototypes to enhance the representations of the prototypes. After that, we train a few shot entity span detector introducing the model-agnostic meta-learning (MAML) algorithm to find a more suitable parameter setting for fast adaption to new entities. We show that integrating semantics of relational representation for prototypes and MAML enhanced entity span detector can significantly improve the performance on few-shot relational triple extraction and extensive experiments show that our method outperforms many prior methods by at least 6.93 % F1 on FewRel 1.0.