detie-cut.yaml

# @package _global_
model:
  name: TripletsExtractorBERTOnly # the model variant used in the paper
  lang: en
  use_syntax_features: False # using extra syntax-based features (not used in the paper)
  postprocess_adp: False # extra work with adpositions (not used in the paper)
  join_is: True # appending [is] [of] [from] to the text (OpenIE6-inspired language-dependent trick used for English)

  pretrained_encoder: bert-base-multilingual-cased # tried: xlm-roberta-large, bert-base-uncased, bert-large-cased
  word_dropout: 0

  stanza_emb_size: 32 # syntax features embeddings size (not used in the paper)

  # Training process details
  max_epochs: 100
  batch_size: 32
  unfreeze_epoch: 0
  unfreeze_layers_from_top: 4
  syntetic_data_after_epoch: ${model.max_epochs}

  # Loss params
  matching: iou # Intersection-over-Union or 'dice' or 'dice_squared'
  disable_bg: True # (not) taking the background 'O' labels into account
  focal_gamma: 0
  class_weights: [1, 1, 1, 1]

  # Transformer params
  num_layers: 2
  hid_size: 512
  n_classes: 4 # S (source), R (relation),T (target), O (other) ~ arg1, rel, arg2, o
  num_detections: 20 # a number of possible simultaneous 'detections' (sequences of labels) for each sentence

  best_version: 276 # an ID of the model