Role-Wise Spatially-Aware Joint Information Extraction with Vocabulary Expansion

Information extraction is the process of transforming unstructured text into structured information. We have observed that entities, events, and relationships within a sentence can be effectively decoded using only the starting and ending pointers. Therefore, we introduce a novel paradigm that treats information extraction as a unified task: span extraction, where relationships are identified by the starting position of one span and the ending position of another. Based on this, we propose one pointer for joint information extraction, which models all tasks using only the head-to-tail pointer. Additionally, we incorporate vocabulary expansion to address the class imbalance problem in the training set, and implicitly models multiple tasks by introducing role-wise spatially-aware subspaces. Experimental results demonstrate that our approach outperforms both task-specific and previous unified models, confirming the effectiveness of our method.