PDTB-discourse-relation-classifier

• A simple PDTB discourse relation classifier built on top of the RoBERTa model.
• It classifies a pair of two sentences or phrases into one of the 11 classes as defined below:
• The model can be trained from the preprocessed dataset in data.

Requirements

• Python 3.6+
• Pytorch
• Huggingface
• Pytorch Lightning
• TorchMetrics

Training

python discourse_baseline.py

Download pre-trained model

• Link

Inference

• Prepare two sentences (or phrases) that a discourse relation is to be classified, and save them to first_sent_path and second_sent_path respectively.
• Run python prepare_silver_dt.py n_parts split_ind first_sent_path second_sent_path
  • n_parts: the number of data splits for the data.
  • split_id: current index of the data splits
• So if you want to split the corpus into three parts, and conduct inference on the first part, it would be:
  • python prepare_silver_dt.py 3 0 path/to/first_sent path/to/second_sent
  • The second and third parts can be run separately on different SLURM servers