Feature Distance-based Framework for Classification of Low-Frequency Semantic Relations

In the relation extraction of semantic relations, it is not uncommon to face settings in which the training data provides very few instances of some relation classes. This is mostly due to the high cost of producing such data and to the class imbalance problem, which may result in some classes presenting small frequencies even with a large annotated corpus. This work thus presents a semi-supervised bootstrapped method to expand this initial training dataset, using pattern matching to extract new candidate instances from the Web. The core of this process uses a multiview feature distance-based framework, which allows quantitative and qualitative analysis of intermediate steps of the process. Experimental results show that this framework provides better results in the relation classification task than the baseline, and the bootstrapped architecture improves the relation classification task as a whole for these low-frequency semantic relations settings.