APCNN: Tackling Class Imbalance in Relation Extraction through Aggregated Piecewise Convolutional Neural Networks

Alisa Smirnova, Julien Audiffren, Philippe Cudré-Mauroux

Abstract: One of the major difficulties in applying distant supervision to relation extraction is class imbalance, as the distribution of relations appearing in text is heavily skewed. This is particularly damaging for the multi-instance variant of relation extraction. In this work, we introduce a new model called Aggregated Piecewise Convolutional Neural Networks, or APCNN, to address this problem. APCNN relies on the combination of two neural networks, a novel objective function as well as oversampling techniques to tackle class imbalance. We empirically compare APCNN to state-of-the-art approaches and show that it outperforms previous multi-instance approaches on two standard datasets.