An Exploration of Prompt-Based Zero-Shot Relation Extraction Method

An Exploration of Prompt-Based Zero-Shot Relation Extraction Method Jun Zhao1∗, Yuan Hu1∗, Nuo Xu1, Tao Gui1†, Qi Zhang1†, Yunwen Chen2, Xiang Gao2 1 School of Computer Science, Shanghai Key Laboratory of Intelligent Information Processing, Fudan University, Shanghai, China 2 DataGrand Information Technology (Shanghai) Co., Ltd., Shanghai, China {zhaoj19,yuanhu20,tgui,qz}@fudan.edu.cn xun22@m.fudan.edu.cn {chenyunwen,gaoxiang}@datagrand.com Abstract Zero-shot relation extraction is an important method for dealing with the newly emerging relations in the real world which lacks labeled data. However, the mainstream two-tower zero-shot methods usually rely on large-scale and in-domain labeled data of predefined relations. In this work, we view zero-shot relation extraction as a semantic matching task optimized by prompt-tuning, which still maintains superior generalization performance when the labeled data of predefined relations are extremely scarce. To maximize the efficiency of data exploitation, instead of directly fine-tuning, we introduce a prompt-tuning technique to elicit the existing relational knowledge in pre-trained language model (PLMs). In addition, very few relation descriptions are exposed to the model during training, which we argue is the performance bottleneck of two-tower methods. To break through the bottleneck, we model the semantic interaction between relational instances and their descriptions directly during encoding. Experiment results on two academic datasets show that (1) our method outperforms the previous state-of-the-art method by a large margin with different samples of predefined relations; (2) this advantage will be further amplified in the low-resource scenario. 1 Introduction Relation extraction (RE) aims to extract the relation between entity pairs from unstructured text. The extracted relation facts can benefit various downstream applications such as knowledge graph completion (Wang et al., 2014), web search (Xiong et al., 2017) and dialog systems (Madotto et al., 2018). However, many effective RE methods (Wu and He, 2019; Du et al., 2018) work within predefined relation sets. They failed to deal with a real-world environment where new relations will emerge after the training phase. These fast-growing new relations make it impossible for us to gather labeled training data for all of them. To recognize the newly emerging relations lacking labeled data, zero-shot RE is of the utmost practical interest. Despite the great potential of zero-shot RE in real-world applications, there have been relatively few studies focusing on this challenging task. To enable models to predict unseen relations, previous works usually model zero-shot relation extraction as a well-designed task form. Levy et al. (2017) consider relation extraction as a machine reading comprehension. They first associate a few question templates for each relation and then determine which relation satisfies the given sentence and question by model prediction. However, a reasonable and effective question template usually needs careful design, which cannot meet the extraction needs of rapidly growing new relations (Chen and Li, 2021). Therefore, instead of manually constructing question templates, subsequent works (Obamuyide and Vlachos, 2018; Chen and Li, 2021) take advantage of the readily available textual description to represent the new relations, and formulate zero-shot RE as a semantic matching task achieving superior results. However, current methods usually require a large number of in-domain labeled …

Content truncated to 500 words