Emotion Classification in Japanese Cancer Survivor Interview Narratives Using Sentiment Polarity and Plutchik Emotion...

Area of Science:

• Natural Language Processing (NLP)
• Computational Linguistics
• Oncology Support

Background:

• Cancer survivors experience complex, coexisting emotions during and after treatment.
• Limited research exists on multidimensional emotion classification in cancer survivor narratives.
• Understanding these emotions is crucial for improving survivorship care.

Purpose of the Study:

• Develop and evaluate NLP-based emotion classification models for Japanese cancer survivor narratives.
• Assess the complementary perspectives of polarity and multidimensional emotion labels.
• Examine the effectiveness of domain-specific fine-tuning versus transfer learning.

Main Methods:

• Analyzed 15 Japanese cancer survivor interview transcripts.
• Annotated text chunks with 3-class sentiment polarity and 8 Plutchik emotion labels.
• Fine-tuned Japanese BERT and LUKE models for classification tasks.
• Evaluated performance using precision, recall, F1-score, and Hamming loss.
• Conducted domain-transfer analysis using the WRIME dataset.

Main Results:

• Interview-trained models outperformed WRIME-trained transfer models.
• Japanese BERT achieved a micro-F1 of 0.696 for polarity; LUKE achieved a macro-F1 of 0.427 for 8 emotions.
• Neutral polarity and trust were the most frequent labels; anger was least frequent.
• Sadness and trust frequently co-occurred, indicating complex emotional states.

Conclusions:

• Domain-specific emotion classification for cancer survivor narratives is feasible.
• Fine-tuned models generally outperformed transfer models, though optimal architecture varied.
• Polarity and 8-emotion labels offer complementary insights into survivor experiences.
• Further research with larger datasets and imbalance-aware methods is needed for clinical application.