SQGE: Support-query prototype guidance and enhancement for few-shot relational triple extraction.
Journal:
Neural networks : the official journal of the International Neural Network Society
Published Date:
Jan 22, 2025
Abstract
The current few-shot relational triple extraction (FS-RTE) techniques, which rely on prototype networks, have made significant progress. Nevertheless, the scarcity of data in the support set results in both intra-class and inter-class gaps in FS-RTE. Instances with restricted support sets make capturing the various features of target instances in the query set difficult, resulting in intra-class gaps. The support set lacks discernible target category characteristics, and the distances between data from various categories are insufficient, leading to intra-class gaps. In this paper, we propose an FS-RTE method based on support-query prototype guidance and enhancement (SQGE). It includes a support-query prototype guide module, which creates query prototypes based on the support prototype and combines the two prototypes. The fusion prototype can accurately capture the fundamental feature that aligns with the query set, suitably match the query features, and reduce the intra-class gap. Furthermore, to address the inter-class gap, we employ entity-level feature enhancement to improve the feature representation of target entities belonging to the same class. On the other hand, we construct positive and negative instances of the target class through contrastive learning, which not only strengthens the representation of the same target class but also distinguishes the feature space of the target class from other classes. Extensive experimental results on three datasets demonstrate the effectiveness of our approach. All the code and data are made available in https://github.com/gao929165733/SQGE_code.