1DOD — One-Dimensional Object Detection for Streaming Text Segmentation

Vietnamese adaptation of 1DOD for meeting dialogue text segmentation.

Model: Qualcomm-AI-Research/BamiBERT — a RoBERTa-based Vietnamese language model with 2048-token context window. Trained on Vietnamese translated AMC (AliMeeting4MUG) data.

Finetuned checkpoint: bamibert-1dod-vi-v1/ (A=60.58, B=45.74)

Results on Vietnamese AMC (AliMeeting4MUG)

Metric Definitions (from the paper)

Key Findings

• A = 60.58 surpasses paper (59.71) despite using a different language (Vietnamese vs Chinese) and backbone (BamiBERT vs PoNet).
• B = 45.74 still lags behind paper (48.47) by ~2.7 pp — boundary precision remains the main weakness.
• S = 98.40 indicates near-perfect structural segmentation at the paragraph level.
• Smooth L1 loss (vs standard L1) and tail processing (t_stop=31, t_min=23) drove significant improvements.

Hyperparameters (matching the paper)

File Structure

1dod_vi/              # Vietnamese 1DOD (BamiBERT)
├── train_vi.py
├── models_vi.py
├── verify_vi.py
├── translate_vi.py
├── 1d_iou.py
├── similarityMetrics.py
└── helpers.py

1dod_cn/              # Original Chinese 1DOD (PoNet)
├── train.py
├── models.py
├── 1d_iou.py
├── similarityMetrics.py
├── helpers.py
└── generate_report.py

Training (Vietnamese)

cd 1dod_vi
python train_vi.py --mode train \
  --model_dir Qualcomm-AI-Research/BamiBERT \
  --data_dir ./Alimeeting4MUG \
  --lr 5e-5 --batch_size 2 --window_size 100 --epochs 3

Evaluation

python train_vi.py --mode test \
  --load_dir ./model_save_vi/{best_checkpoint} \
  --data_dir ./Alimeeting4MUG --window_size 100