MapRE: An Effective Semantic Mapping Approach for Low-resource Relation Extraction

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2694–2704 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 2694 MapRE: An Effective Semantic Mapping Approach for Low-resource Relation Extraction Manqing Dong, Chunguang Pan, and Zhipeng Luo DeepBlue Technology (Shanghai) Co., Ltd dongmanqing@gmail.com, luozp@deepblueai.com Abstract Neural relation extraction models have shown promising results in recent years; however, the model performance drops dramatically given only a few training samples. Recent works try leveraging the advance in few-shot learn- ing to solve the low resource problem, where they train label-agnostic models to directly compare the semantic similarities among con- text sentences in the embedding space. How- ever, the label-aware information, i.e., the re- lation label that contains the semantic knowl- edge of the relation itself, is often neglected for prediction. In this work, we propose a framework considering both label-agnostic and label-aware semantic mapping informa- tion for low resource relation extraction. We show that incorporating the above two types of mapping information in both pretraining and fine-tuning can significantly improve the model performance on low-resource relation extraction tasks. 1 Introduction Relation Extraction (RE), which aims at discover- ing the correct relation between two entities in a given sentence, is a fundamental task in NLP (Gao et al., 2019). The problem is generally regarded as a supervised classification problem by training on large-scale labelled data (Zhang et al., 2017). Neural models, e.g. RNN-based methods (Zhou et al., 2016), or more recently, BERT-based meth- ods (Soares et al., 2019; Peng et al., 2020), have shown promising results on RE tasks, where they achieve state-of-the-art performance or even com- parable with human performance on several public RE benchmarks. Despite the promising performance of the exist- ing neural relation classification frameworks, re- cent studies (Han et al., 2018) found that the model performance drops dramatically as the number of instances for a relation decreases, e.g., for long-tail Support Instances It is approximately 9 km away from Mount Korbu, the tallest mountain of the (a) mountain Titiwangsa Mountains. range The San Ysidro Mountains are part of the Peninsular Ranges System. (b) head of government One of the umpires was Edmund Barton, who became Australia's first prime minister. Keith Burdette was the Secretary of Commerce for the state of West Virginia under the administration of Governor Earl Ray Tomblin. Senator Patrick Leahy and Vermont Governor Phil Scott. Query Instance (a) or (b) ? Figure 1: Example for a 2-way 2-shot relation extrac- tion task. The entities with underlines are head entities, and the entities in bold are tail entities. The target is to predict the relation between the head and the tail enti- ties for a given query instance. relations. An extreme condition is few-shot rela- tion extraction, where only few support examples are given for the unseen relations, see Figure 1 as an example. A conventional way to solve the data deficiency problem of RE is distant supervision (Mintz et al., 2009; Hu et al., 2019), which assumes same entity- pairs have same relations …

Content truncated to 500 words