Relation extraction from imbalanced data by independently estimating linguistic expressions

Abstract

Supervised Relation Extraction (RE) often comes with imbalanced datasets. Although advanced classifiers are achieving great performance on well-prepared datasets, it is hard to expect equivalent results on severely skewed data

underrepresented relation classes can be neglected. Data imbalance problem is more severe when a dataset contains many relations because most relations are in the long tail. This paper proposes an RE algorithm to learn from a dataset with imbalance. The algorithm independently estimates a probability (sample precision) of a pattern of dependency graph to express a certain relation. For efficiency, lattices are expanded with reasonable stopping conditions to collect dependency patterns. Since probabilities are independently estimated, it can be more resistant against the data imbalance problem. This paper includes the system description and an experiment to compare resistance with a standard classifier.