PROJECT TITLE :

Gradual Machine Learning for Entity Resolution

ABSTRACT:

Entity resolution (ER), which is typically thought of as a classification problem, can pose a great deal of difficulty when applied to real data due to the prevalence of dirty values. The most advanced solutions for ER were constructed using a variety of learning models, the most prominent of which were Deep Neural Networks. These models require a large amount of training data that has been accurately labeled. Unfortunately, high-quality labeled data typically require a significant amount of expensive manual labor, and as a result, they are not easily accessible in many real-world scenarios. In this paper, we propose a new learning paradigm for ER that we call gradual machine learning. The goal of this learning paradigm is to enable efficient machine labeling without the need for any manual labeling effort to be made. It starts off by labeling some of the easier instances in a task, which can be done automatically by the machine with a high level of accuracy. After that, it labels some of the more difficult instances by using iterative factor graph inference. When performing a task using gradual machine learning, the difficult instances in the task are labeled gradually in small stages based on the estimated evidential certainty provided by the instances that have already been labeled as easier. The results of our in-depth experiments on real data have demonstrated that the performance of the proposed method is noticeably superior to that of its unsupervised alternatives, and it is highly competitive in comparison to the most advanced supervised techniques currently available. We show, by using ER as a test case, that gradual machine learning is a promising paradigm that could potentially be applied to other difficult classification tasks that require extensive labeling effort.