Applied Deep Learning 深度學習之應用 by Vivian Chen 陳縕儂
Surpassed strong baseline for all three assignments (Final grade: 100/100)
⭐Please consider starring this project if you find my code useful.⭐
For more details, refer to the reports.
- HW1: Report
- Intent Classification ← CNN+LSTM
- Accuracy: 0.9027
- Slot Tagging ← CNN+LSTM
- Joint accuracy: 0.8060
- Intent Classification ← CNN+LSTM
- HW2: Report
- Chinese Question Answering with Multiple Paragraph
- Divided into two stage inference
- Paragraph Selection ← Chinese Macbert Large
- Accuracy: 0.970
- Chinese QA ← Chinese Macbert Large
- Extact Match: 0.778
- Chinese Question Answering with Multiple Paragraph
- HW3: Report
- Chinese News Summarization (Title Generation) ← mT5 base
- Without RL finetuning:
- Rouge 1: 0.2691
- Rouge 2: 0.1081
- Rouge 3: 0.2393
- With policy gradient RL finetuning:
- Rouge 1: 0.2712 (+0.8%)
- Rouge 2: 0.1070 (-1.0%)
- Rouge 3: 0.2411 (+0.8%)
- Without RL finetuning:
- Chinese News Summarization (Title Generation) ← mT5 base
- Final Project:
- Topic: MULTIMODAL PROMPT TUNING: REAL-WORLD USAGE ON NTU TREE CLASSIFICATION
- Report
- Presentation Slide
- 10min Video
- Achieved 85% accuracy under 8-shot setting with < 2min training time