Graph Representation of Chinese EMRs and its Short-text Classification Application1

Chinese Electrical Medical Records (EMRs) include not only text entities features but also relationship features. However, the remaining representation solutions of Traditional Chinese Electrical Medical Records (EMRs) just focuses on entities and context without other relationships, lacking details and limiting the text analysis performance of NLP models. Therefore, a graph representation solution that combines entities features and relationship features is introduced. Specifically, Point-wised Mutual Information (PMI) is leveraged to represent relationships among text entities, and Word2Vec to represent text entities themselves. As a presentation and evaluation of NLP application, GCN is applied to construct word2vec+GCN text classification solution that classifies short text-data from Chinese EMRs in text classification evaluation of TCM medical text records. As a result, word2vec-GCN achieved 89.51% of average precision, 89.83% of average recall, 89.06% of Macro F1 88.90% of Kappa, and 98.90% AUC values, outperforming other solutions based on ordinary NLP models. Meanwhile, the ablation experiment was introduced to validate the hypotheses that the performance improvement would be caused by the embedding of relationships features, and the hypotheses was approved. It could be concluded that the graph representation based on word2vec and PMI is available, improving the qualification of text representation and performance of NLP tasks like text classification, even if this graph representation is utilized by ordinary NLP models.