SHSK0118/BERT-basedDomainClassification_ComplaintTexts_ja

BERT-based Domain Classification for Japanese Complaint Texts

A BERT-based Japanese text classification model trained for domain classification of complaint texts.

---

Model Details

• Architecture: BERT for Sequence Classification
• Language: Japanese
• Task: Multi-class domain classification
• Framework: Hugging Face Transformers

---

Training Data

Training corpus:

BERT-basedDomainClassification_ComplaintTexts_ja Dataset

Dataset split:

• Train: 90%
• Validation: 5%
• Test: 5%

---

Evaluation

Test Accuracy: 73.0%

---

Performance Discussion

The model was trained on primarily formal written text (Wikimedia-derived corpus), while evaluation was conducted on complaint-style texts.

The domain gap between formal and conversational language likely contributed to reduced performance.

---

Intended Use

• Educational purposes
• Research prototyping
• Domain classification experiments

---

Limitations

• No domain adaptation applied
• Performance sensitive to genre distribution

---

Author

Independent implementation by Shota Tokunaga.