Weakly Supervised Named Entity Tagging with Learnable Logical Rules

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4568–4581 August 1–6, 2021. ©2021 Association for Computational Linguistics 4568 Weakly Supervised Named Entity Tagging with Learnable Logical Rules Jiacheng Li1∗, Haibo Ding2, Jingbo Shang1, Julian McAuley1, Zhe Feng2 University of California, San Diego1 Bosch Research North America2 {j9li,jshang,jmcauley}@eng.ucsd.edu1 {haibo.ding, zhe.feng2}@us.bosch.com2 Abstract We study the problem of building entity tag- ging systems by using a few rules as weak supervision. Previous methods mostly focus on disambiguating entity types based on con- texts and expert-provided rules, while assum- ing entity spans are given. In this work, we propose a novel method TALLOR that boot- straps high-quality logical rules to train a neu- ral tagger in a fully automated manner. Specif- ically, we introduce compound rules that are composed from simple rules to increase the precision of boundary detection and generate more diverse pseudo labels. We further design a dynamic label selection strategy to ensure pseudo label quality and therefore avoid over- fitting the neural tagger. Experiments on three datasets demonstrate that our method outper- forms other weakly supervised methods and even rivals a state-of-the-art distantly super- vised tagger with a lexicon of over 2,000 terms when starting from only 20 simple rules. Our method can serve as a tool for rapidly building taggers in emerging domains and tasks. Case studies show that learned rules can potentially explain the predicted entities. 1 Introduction Entity tagging systems that follow supervised train- ing, while accurate, often require a large amount of manual, domain-specific labels, making them difficult to apply to emerging domains and tasks. To reduce manual effort, previous works resort to manual lexicons (Shang et al., 2018b; Peng et al., 2019) or heuristic rules provided by domain ex- perts (Fries et al., 2017; Safranchik et al., 2020; Lison et al., 2020b) as weak supervision. For ex- ample, LinkedHMM (Safranchik et al., 2020) can achieve performance close to supervised models using 186 heuristic rules in addition to a lexicon of over two million terms. However, it is challenging ∗Work done during an internship at Bosch Research. induce new rule If TokenString(x)==“Dallas”, then Label(x)=“Location” Ryn lives in Dallas. John lives in Dallas where he was born. He lives in Dallas this year. If POS(x)==“PROPN” and PreNgram(x)==“lives in”, then Label(x)=“Location” seed rule Fobes lives in Seattle. She lives in Vancouver. The man lives in California. v3 Figure 1: Examples of a seed logical rule and a newly induced rule from labeled data for recognizing loca- tions. ‘x’ denotes a token span from a given sentence. for experts to write complete and accurate rules or lexicons in emerging domains, which requires both a significant amount of manual effort and a deep understanding of the target data. How to build accurate entity tagging systems using less manual effort is still an open problem. In this work, we explore methods that can auto- matically learn new rules from unlabeled data and a small set of seed …

Content truncated to 500 words