IBRE: An Incremental Bootstrapping Approach for Chinese Appointment and Dismissal Relation Extraction

In the field of government affairs, Appointment and Dismissal Relation Extraction (ADRE) of officials from personnel news is crucial for updating government knowledge. However, ADRE faces great challenges, including extremely high accuracy demand, lack of data, tuple sparsity and conflict, and requiring incremental update. To address these challenges, we propose an Incremental Bootstrapping approach for Relation Extraction (IBRE) and apply it to real-time updating of personnel knowledge in government-affair Knowledge Graphs. IBRE starts with few seeds and trains with pattern generation, pattern evaluation, tuple prediction and seed augmentation in an iterative and incremental manner. First, we design a new strategy for defining seeds as document-tuple pairs to eliminate the effects of tuple sparsity and conflict. Then, a new definition of patterns with both word and part-of-speech is proposed to get high accuracy. Finally, we augment seeds with corrected tuples and apply incremental learning to continually improve performance with least training cost. We build a dataset called ADNP (Appointment and Dismissal News from People.cn) and compare our approach with baselines. Comparison results show that our approach performs the best. Moreover, experimental results demonstrate that incremental learning continuously improves the performance of IBRE.