RAPS: A Novel Few-Shot Relation Extraction Pipeline with Query-Information Guided Attention and Adaptive Prototype Fusion: RAPS for Few-Shot RE

作者：Yuzhe Zhang, Min Cen, Tongzhou Wu, Hong Zhang · 年份：2024 · DOI：10.1145/3677779.3677803 · 被引用次数：2 · 研究领域：Topic Modeling、Natural Language Processing Techniques、Domain Adaptation and Few-Shot Learning

Few-shot relation extraction (FSRE) aims at recognizing unseen relations by learning with limited annotated instances. Prototypical network has been proven to be successful owing to its simplicity and effectiveness for the FSRE task. However, unnecessary and even destructive parameters are unavoidably introduced when implicitly introducing external relation information to assist prototype representation learning under the prototypical network-based frameworks. Furthermore, simply averaging the feature representations of support examples in each relation category might not be efficient since the contribution of the support instances are generally heterogeneous. To solve these issues, we propose a novel pipeline for FSRE task on the ground of queRy-information guided Attention and adaptive Prototype fuSion, RAPS for short. To be more precise, our approach involves using a query-information guided attention module to generate a relationship prototype. This module leverages extensive interactive information between support and query examples, enabling us to obtain more precise initial prototype representations. Then, we elaborately combine the derived initial prototype with the relation information by an adaptive prototype fusion mechanism to obtain an integrated prototype for both training and prediction. Experiments conducted on two well-known benchmark datasets demonstrate a remarkable improvement of RAPS against previous state-of-the-art FSRE methods. Our code is available at...