Contextual-Aware Information Extractor with Adaptive Objective for Chinese Medical Dialogues
Gangqiang Hu, Shengfei Lyu, Xingyu Wu, Jinlong Li, Huanhuan Chen · ACM Transactions on Asian and Low-Resource Language Information Processing · 2022
Electronic Medical Records ( EMRs ) are the foundation of modern medical information systems. Despite the benefits of EMRs, the exhausting process of constructing EMRs decreases the efficiency of medical consultation. Therefore, it becomes an emerging research field to automatically extract EMRs from medical dialogues. In Chinese medical dialogues, the phenomena of omission and reference are extremely common, leading to strong contextual relevance among utterances. However, recent studies on converting Chinese medical dialogues to EMRs lack a reliable mechanism to effectively exploit the contextual relevance information among utterances. Moreover, they neglect the frequency imbalance of different items and treat these items indiscriminately, which eventually degrade the overall system performance. In this article, we proposed a Contextual-Aware Information Extractor ( CANE ), which employs a local-to-global mechanism over utterances to model the contextual relevance among utterances. Furthermore, an adaptive objective is introduced to alleviate the frequency imbalance of items by dynamically assigning weights to each sample. Experimental results indicate that CANE outperforms previous state-of-the-arts with considerable improvements (+6.11% and +3.39% on F1-score).