[3Rin4] Interactive 1

[3Rin4-64] Urgency Prediction Based on Document Classification Considering Annotation Noise for Medical Consultation Texts

〇Yusuke Fukasawa¹, Naoya Hashimoto¹ (1.Kids Public)

Keywords:Medical, Document Classification, Annotation Noise

In teleconsultation service, it is crucial to obtain as much information as possible from the pre-interview text provided by the user. In this study, we constructed a model that predicts medical urgency from the pre-interview text. Data from the teleconsultation service "Online Pediatricians" were used as training data. Because the teaching signal is given by different doctors, it may contain inconsistent noise. To solve this problem, we propose a method based on Paired Softmax Divergence Regularization, in which data are given different labels during learning and training is performed separately for data with different labels. In experiments, the proposed method was compared against previous methods and shown to be useful.