Dependency-aware Prototype Learning for Few-shot Relation Classification

Proceedings of the 29th International Conference on Computational Linguistics, pages 2339–2345 October 12–17, 2022. 2339 Dependency-aware Prototype Learning for Few-shot Relation Classification Tianshu Yu 1,2, Min Yang 1∗, Xiaoyan Zhao 1 1 Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences 2 University of Chinese Academy of Sciences ts.yu@siat.ac.cn, min.yang@siat.ac.cn, xy.zhao@siat.ac.cn Abstract Few-shot relation classification aims to classify the relation type between two given entities in a sentence by training with a few labeled in- stances for each relation. However, most of existing models fail to distinguish multiple re- lations that co-exist in one sentence. This pa- per presents a novel dependency-aware proto- type learning (DAPL) method for few-shot re- lation classification. Concretely, we utilize de- pendency trees and shortest dependency paths (SDP) as structural information to complement the contextualized representations of input sen- tences by using the dependency-aware embed- ding as attention inputs to learn attentive sen- tence representations. In addition, we introduce a gate controlled update mechanism to update the dependency-aware representations accord- ing to the output of each network layer. Exten- sive experiments on the FewRel dataset show that DAPL achieves substantially better perfor- mance than strong baselines. For reproducibil- ity, we will release our code and data upon the publication of this paper at https:// github.com/publicstaticvo/DAPL. 1 Introduction Relation classification, which aims to classify the relation between two entities in a sentence, is a fundamental task for information retrieval (Kadry and Dietz, 2017), knowledge graph construction (Shen et al., 2020; Ji et al., 2021) and question answering (Luo et al., 2018). Most of existing relation classification methods (Wang et al., 2016; Guo et al., 2019; Shen et al., 2020; Tian et al., 2021; Zhao et al., 2022a) focus on the supervised scenario where sufficient labeled training data is available. However, it is time-consuming and labor-intensive to collect large-scale labeled data in many real- world applications, especially in the low-resource settings (Geng et al., 2019, 2020; Fan et al., 2021; Zhao et al., 2022b). *Min Yang is corresponding author. Recently, few-shot relation classification (FSRC), which explores relation extraction methods by training with a few labeled examples in each relation, has become a hot research topic (Gao et al., 2019; Qu et al., 2020; Gao et al., 2020; Wang et al., 2020; Xu and Xiang, 2021; Ding et al., 2021; Fan et al., 2021). For instance, Han et al. (2018) introduce a large-scale FSRC dataset and implement several well-known few-shot learning techniques (Finn et al., 2017; Snell et al., 2017) for FSRC. Qu et al. (2020) propose a Bayesian meta learning approach for FSRC, which learns the posterior distributions of prototype vectors among different relations. Despite the remarkable progress of FSRC meth- ods, there is still a technical challenge which is not addressed well in prior work. Specifically, there can be multiple relations that co-exist in a sentence, while only one relation corresponds to the given en- tity pairs. The other existed relations may mislead the classifier to the wrong relation class, which is called the misleading relation. …

Content truncated to 500 words