Relation-Guided Cascading Generation for Low-Shot Relational Triplet Extraction

Relational triplet extraction aims at constructing knowledge graphs, in the format of entity-relation-entity, from unstructured texts. Previous works mainly focus on an ideal scenario with sufficient labeled data, but fail to consider the cases with only a few (few-shot) or even no (zero-shot) labeled samples. Recently, although several methods have considered the few-shot relational triplet extraction, they can only tackle the one-relational triplet extraction without taking the multiple-triplet extraction into consideration. Meanwhile, these approaches overlook the accumulation of errors in the entity and relation extraction. Moreover, the approaches for zero-shot triplet extraction ignore the impact of fully utilizing relation labels and integrating negative samples to benefit the model train. In this paper, we carefully analyze the problems existing in the low-shot relational triplet extraction task (including both few-shot and zero-shot settings) and explore how to extract the triplets from the generative perspective to address the severe limitation of most previous methods, which only extract one relational triplet per sentence. Specifically, we propose a relation-guided cascading generation framework (CasGen) to take full advantage of the relation information to guide the entity extraction and then leverage the extracted entities to facilitate the relation extraction. Besides, our proposed cascading strategy also involves negative samples to make the model learn fine-grained representations. The experimental results show that our CasGen achieves better performance than many state-of-the-art methods by 19.5 (44.7%$\uparrow$) F1 score in the few-shot setting and 7.6 (30.9%$\uparrow$) F1 score in the zero-shot setting on average.