[3S1-OS-7b] OS-7

[3S1-OS-7b-01] Emotion Analysis of Cancer Patients Using Interview Texts from Cancer Survivors

〇Soma Hisamura¹, Kyoko Sayama¹, Satoshi Watabe¹, Hayato Kizaki¹, Shungo Imai¹, Toru Kishida², Natsumi Fukuoka², Shuntaro Yada³, Eiji Aramaki³, Satoko Hori¹ (1. Keio University Faculty of Pharmacy, 2. GanNote, 3. Nara Institute of Science and Technology)

Keywords:Medical Natural Language Processing, Patient-Generated Text, Emotion Analysis

We aimed to develop an emotion classifier using a natural language processing model for cancer survivors' interview transcripts. Focusing on statements from GanNote-organized interviews, we utilized BERT and LUKE as pre-trained models. Training data included 1) cancer survivors' interview transcripts and 2) the WRIME dataset of social media posts with emotion labels. We built classifiers for 3-emotion multiclass and 8-emotion multilabel classifications based on Plutchik's Wheel of Emotions. The test data were cancer survivors' interview transcripts. The LUKE-trained model excelled in all tasks, scoring 0.76 for neutral in 3-emotion classification and above 0.60 for the other emotions. In 8-emotion classification, trust scored 0.62, sadness/fear/disgust/anticipation around 0.50, but joy/anger/surprise fell below 0.35. While some emotional classifications remain challenging, we succeeded in creating a classifier extracting three and most of the eight emotions from cancer survivor interviews.