Abstract: Most prior work on Information Extraction typically predicts labels of individual instances (e.g., event triggers, relations, entities) independently regardless of their interactions. We propose a novel framework, HighIE, that aims to integrate high-order cross-subtask and cross-instance dependencies in both learning and inference. High-order inference on label variables is an NP-hard problem. To address it, we propose a high-order decoder that is unfolded from an approximate inference algorithm. The experimental results show that our approach achieves consistent improvement compared with prior work.
0 Replies
Loading