Minimally Supervised Instance Matching: An Alternate Approach
Instance matching concerns identifying pairs of instances that refer to the same underlying entity. Current state-of-the-art instance matchers use machine learning methods. Supervised learning systems achieve good performance by training on significant amounts of manually labeled samples. To alleviate the labeling effort, this poster (The work presented herein is also being published as a full conference paper at ESWC 2015. This poster provides a more high-level overview and discusses supplemental experimental findings beyond the scope of the material in the full paper.) presents a minimally supervised instance matching approach that is able to deliver competitive performance using only 2 % training data. As a first step, a committee of base classifiers is trained in an ensemble setting using boosting. Iterative semi-supervised learning is used to improve the performance of the ensemble classifier even further, by self-training it on the most confident samples labeled in the current iteration. Empirical evaluations on real-world data show that, using a multilayer perceptron as base classifier, the system is able to achieve an average F-Measure that is within 2.5 % of that of state-of-the-art supervised systems.
KeywordsInstance Matching Boosting Semi-supervision Self-training
- 2.Ngomo, A.-C.N., Lyko, K.: Unsupervised learning of link specifications: deterministic vs. non-deterministic. In: OM, pp. 25–36 (2013)Google Scholar