This is a LUKE (Japanese version) Base model for emotion analysis in Japanese additionally fine-tuned for emotion detection and classification.
The model was based on
studio-ousia/luke-japanese-large, and later fine-tuned on a dataset containing 10 labels of emotional texts from Japanese dictionaries, blogs and Twitter.
The dataset was composed of about 1,700 sentences, with about 100~250 sentences each for each emotion category.
emotion_mapping = { 0: 'amaze/surprise', 1: 'anger', 2: 'dislike', 3: 'excitement', 4: 'fear', 5: 'joy', 6: 'like/fondness', 7: 'relief', 8: 'sad/gloom', 9: 'shame' }
emotion_mapping = { 0: '驚き', 1: '怒り', 2: 'いや', 3: '昂り', 4: '怖がり', 5: '喜び', 6: '好き', 7: '安らぎ', 8: '悲しみ', 9: '恥ずかしい' }
Evaluation result with our test data
| Accuracy | Macro F1 |
|---|
| 95.57% | 95.42% |
I will continue to fine-tune this model soon.
