Low-resource Deep Entity Resolution with Transfer and Active Learning

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5851–5861 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 5851 Low-resource Deep Entity Resolution with Transfer and Active Learning Jungo Kasai♥∗ Kun Qian♣ Sairam Gurajada♣ Yunyao Li♣ Lucian Popa♣ ♥Paul G. Allen School of Computer Science & Engineering, University of Washington ♣IBM Research – Almaden jkasai@cs.washington.edu {qian.kun,Sairam.Gurajada}@ibm.com {yunyaoli,lpopa}@us.ibm.com Abstract Entity resolution (ER) is the task of iden- tifying different representations of the same real-world entities across databases. It is a key step for knowledge base creation and text mining. Recent adaptation of deep learn- ing methods for ER mitigates the need for dataset-specific feature engineering by con- structing distributed representations of entity records. While these methods achieve state- of-the-art performance over benchmark data, they require large amounts of labeled data, which are typically unavailable in realistic ER applications. In this paper, we develop a deep learning-based method that targets low- resource settings for ER through a novel com- bination of transfer learning and active learn- ing. We design an architecture that allows us to learn a transferable model from a high- resource setting to a low-resource one. To fur- ther adapt to the target dataset, we incorporate active learning that carefully selects a few in- formative examples to fine-tune the transferred model. Empirical evaluation demonstrates that our method achieves comparable, if not bet- ter, performance compared to state-of-the-art learning-based methods while using an order of magnitude fewer labels. 1 Introduction Entity Resolution (ER), also known as entity matching, record linkage (Fellegi and Sunter, 1969), reference reconciliation (Dong et al., 2005), and merge-purge (Hern´andez and Stolfo, 1995), identifies and links different representa- tions of the same real-world entities. ER yields a unified and consistent view of data and serves as a crucial step in downstream applications, includ- ing knowledge base creation, text mining (Zhao et al., 2014), and social media analysis (Campbell ∗Work done during summer internship at IBM Research – Almaden. et al., 2016). For instance, seen in Table 1 are citation data records from two databases, DBLP and Google Scholar. If one intends to build a system that analyzes citation networks of publica- tions, it is essential to recognize publication over- laps across the databases and to integrate the data records (Pasula et al., 2002). Recent work demonstrated that deep learning (DL) models with distributed representations of words are viable alternatives to other machine learning algorithms, including support vector machines and decision trees, for performing ER (Ebraheem et al., 2018; Mudgal et al., 2018). The DL models provide a universal solution to ER across all kinds of datasets that alleviates the ne- cessity of expensive feature engineering, in which a human designer explicitly defines matching functions for every single ER scenario. However, DL is well known to be data hungry; in fact, the DL models proposed in Ebraheem et al. (2018); Mudgal et al. (2018) achieve state-of-the-art per- formance by learning from thousands of labels.1 Unfortunately, realistic ER tasks have limited …

Content truncated to 500 words