Joint Extraction of Entities and Relations from Ancient Chinese Medical Literature

Abstract

Extracting important knowledge from literature is a challenging Natural Language Processing (NLP) task involving two subtasks, Named Entity Recognition (NER) and Relation Extraction (RE). Traditionally these two sub-tasks are served as a pipeline model and the error propagation is inevitable. In this paper, considering of the characteristics of ancient Chinese medical books, we propose a new joint model that performs NER and RE simultaneously. This model also involves a Multi-Head Attention layer to extract long distance related features and adversarial training method to improve robustness and stability. Experiments show that our model performs well on multi-context joint entity and relation extraction in TCM (Traditional Chinese Medicine) dataset.