Manual Evaluation Matters: Reviewing Test Protocols of Distantly Supervised Relation Extraction

Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 1306–1318 August 1–6, 2021. ©2021 Association for Computational Linguistics 1306 Manual Evaluation Matters: Reviewing Test Protocols of Distantly Supervised Relation Extraction Tianyu Gao1, Xu Han2, Yuzhuo Bai2, Keyue Qiu2, Zhiyu Xie2, Yankai Lin3 Zhiyuan Liu2∗, Peng Li3, Maosong Sun2 and Jie Zhou3 1Department of Computer Science, Princeton University 2Department of Computer Science and Technology, Tsinghua University 3Pattern Recognition Center, WeChat AI, Tencent Inc tianyug@princeton.edu {hanxu17, qky18, byz18, xiezy19}@mails.tsinghua.edu.cn Abstract Distantly supervised (DS) relation extraction (RE) has attracted much attention in the past few years as it can utilize large-scale auto- labeled data. However, its evaluation has long been a problem: previous works either take costly and inconsistent methods to manually examine a small sample of model predictions, or directly test models on auto-labeled data— which, by our check, produce as much as 53% wrong labels at the entity pair level in the popular NYT10 dataset. This problem has not only led to inaccurate evaluation, but also made it hard to understand where we are and what’s left to improve in the research of DS- RE. To evaluate DS-RE models more credi- bly, we build manually-annotated test sets for two DS-RE datasets, NYT10 and Wiki20, and thoroughly evaluate several competitive mod- els, especially the latest pre-trained ones. The experimental results show that the manual eval- uation can indicate very different conclusions from automatic ones, especially some unex- pected observations, e.g., pre-trained models can achieve dominating performance while be- ing more susceptible to false-positives com- pared with previous methods. We hope that both our manual test sets and observations can help advance future DS-RE research.1 1 Introduction Relation extraction (RE) aims at extracting rela- tional facts between entities from the text. One crucial challenge for building an effective RE sys- tem is how to obtain sufficient annotated data. To tackle this problem, Mintz et al. (2009) propose distant supervision (DS) to generate large-scale auto-labeled data by aligning relational facts in knowledge graphs (KGs) to text corpora, with the *Corresponding author e-mail: liuzy@tsinghua.edu.cn 1Our code and data are publicly available at https:// github.com/thunlp/opennre. Musk unveiled the Tesla Cybertruck. Musk stood in front of a Tesla Model S. Musk owns 28.9M Tesla shares. CEO Shareholder Place of birth Capital Founder ������� Elon Musk ���������������Tesla Inc. ����������������� Figure 1: Typical errors made by DS evaluation. In the figure, DS labels the bag with only the relation CEO, while none of the sentences express the relation. Also, it misses a correct relation shareholder due to the incompleteness of the knowledge graphs. core assumption that one sentence mentioning two entities is likely to express the relational facts be- tween the two entities from KGs. As DS can bring hundreds of thousands of auto- labeled training instances for RE without any hu- man labor, DS-RE has been widely explored in the past few years (Riedel et al., 2010; Hoffmann et al., 2011; Zeng et al., 2015; Lin et al., 2016; Feng et al., 2018; Vashishth et al., 2018) …

Content truncated to 500 words