Relabeling Distantly Supervised Training Data for Temporal Knowledge Base Population

摘要

We enhance a temporal knowledge base population system to improve the quality of distantly supervised training data and identify a minimal feature set for classification. The approach uses multi-class logistic regression to eliminate individual features based on the strength of their association with a temporal label followed by semi-supervised relabeling using a subset of human annotations and lasso regression. As implemented in this work, our technique improves performance and results in notably less computational cost than a parallel system trained on the full feature set.