mix nasty.train.e2e_coref (Nasty v0.3.0)

Train end-to-end span-based coreference resolution models.

Usage

mix nasty.train.e2e_coref \
  --corpus data/ontonotes/train \
  --dev data/ontonotes/dev \
  --output priv/models/en/e2e_coref \
  --epochs 25 \
  --batch-size 16 \
  --learning-rate 0.0005

Options

• --corpus - Path to training data directory (required)
• --dev - Path to development data directory (required)
• --output - Base path for saving models (required)
• --epochs - Number of training epochs (default: 25)
• --batch-size - Training batch size (default: 16)
• --learning-rate - Learning rate (default: 0.0005)
• --hidden-dim - LSTM hidden dimension (default: 256)
• --dropout - Dropout rate (default: 0.3)
• --patience - Early stopping patience (default: 3)
• --max-span-width - Maximum span width (default: 10)
• --top-k-spans - Keep top K spans per sentence (default: 50)
• --span-loss-weight - Weight for span detection loss (default: 0.3)
• --coref-loss-weight - Weight for coreference loss (default: 0.7)