Abstract: Jungo Kasai, Kun Qian, Sairam Gurajada, Yunyao Li, Lucian Popa. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics.
2019.
Introduction
Entity Resolution (ER)—the process of identifying whether two distinct data records refer to the same real-world entity—is a fundamental challenge in data integration. While modern Deep Learning (DL) models have significantly advanced the field by eliminating the need for manual feature engineering, they typically require vast amounts of labeled data to achieve high performance. In many practical scenarios, labeled examples are scarce, and the data is heavily imbalanced, with non-matching pairs vastly outnumbering matches. This paper proposes a framework that combines adversarial transfer learning with a partition-based active learning strategy to enable high-performance ER in low-resource settings, requiring up to 10x fewer labels than traditional DL methods.
Methodology
The authors introduce a two-pronged approach to address data scarcity: a transferable model architecture and a specialized active learning strategy.
Transferable Architecture
To facilitate knowledge transfer across datasets with different schemas, the model utilizes fastText embeddings for robust word representation and Bi-directional Gated Recurrent Units (BiGRUs) to generate attribute-level representations.
A key innovation is the attribute-agnostic design. Instead of concatenating attribute similarity vectors—which would tie the model to a specific database schema—the authors sum the similarity vectors. This "summing trick" ensures the final classification layer receives a vector of constant dimensionality regardless of the number of attributes in the source or target datasets. Furthermore, the model employs adversarial training via a Gradient Reversal Layer. This forces the network to learn features that are effective for matching while remaining invariant to the specific characteristics of the source or target datasets.
Partition-Based Active Learning
Standard active learning often struggles with ER due to the extreme class imbalance; random or uncertainty-based sampling frequently selects "easy" non-matches. The authors propose partition sampling, which splits the unlabeled pool into predicted matches and non-matches. By sampling the most uncertain examples from both partitions, the model is forced to learn from likely false positives and false negatives. Additionally, the framework incorporates pseudo-labeling of high-confidence examples to stabilize the network during fine-tuning.
Experimental Results
The framework, termed DTAL (Deep Transfer Active Learning), was evaluated across multiple domains, including citation data (DBLP, ACM, Cora), restaurant listings, and software products.
Key Findings
- Label Efficiency: DTAL consistently achieved state-of-the-art F1 scores using an order of magnitude fewer labels. On the Cora dataset, which features complex schemas with 8 attributes, DTAL reached a 97.68 F1 score with only 1,000 labels. In contrast, standard deep learning models achieved only 84.35 F1 with the same amount of data.
- Transfer Success: The adversarial transfer component proved highly effective within similar domains. In the restaurant genre, transferring knowledge from Zomato-Yelp to Fodors-Zagats improved the initial F1 score from 11.76 to 70.13 before any active learning occurred. With just 100 additional active labels, the model achieved a perfect 100 F1 score.
- DL vs. Traditional Baselines: The research quantifies the "data-hungry" nature of DL. Without the proposed transfer and active learning techniques, traditional models like SVMs often outperform standard DL models when training data is limited.
Conclusion
This work demonstrates that the combination of adversarial transfer learning and partition-based active learning can overcome the data bottleneck in Deep Entity Resolution. By creating a universal similarity representation and ensuring balanced uncertainty sampling, the DTAL framework makes deep learning a viable and efficient solution for ER even in low-resource environments. Future work includes extending these techniques to cross-domain transfers and non-English datasets.
Summary updated