conf/experiment/ontonotes_pseudo.yaml

# @package _global_

# This configuration trains ontonotes using the speaker information and pseudo singletons.
# This is the best OntoNotes configuration in our CRAC 2021 work.
# Note that this configuration doesn't assume other upstream features such as document genre

# Model name in CRAC 2021: longdoc^S ON + PS 60K


defaults:
  - override /datasets: ontonotes
  - override /trainer: train.yaml
  - override /model: model.yaml

trainer:
  log_frequency: 250
  max_evals: 20
  patience: 10
  eval_per_k_steps: null

model:
  doc_encoder:
    add_speaker_tokens: True

datasets:
  ontonotes:
    singleton_file: ontonotes/ment_singletons_longformer_speaker/30.jsonlines