RoRED: Bootstrapping labeling rule discovery for robust relation extraction

Labeling rules can be leveraged to produce training data by matching the sentences in the corpus. However, the robustness of the relation extraction is reduced by noisy labels generated from incorrectly matched and missing sentences. To address this problem, we propose the bootstrapping labeling rule discovery method for robust relation extraction (RoRED). Specifically, we first define PN-rules to filter incorrectly matched sentences based on positive (P) and negative (N) rules. Second, we design a semantic-matching mechanism to match missing sentences based on semantic associations between rules, words, and sentences. Moreover, we present a co-training-based rule verification approach to refine the labels of matched sentences and improve the overall quality of bootstrapped rule discovery. Experiments on a real-world dataset indicate that RoRED achieves at least a 20% gain in F1 score compared to state-of-the-art methods.