Deep Reinforcement Learning for Chinese Zero Pronoun Resolution

Deep neural network models for Chinese zero pronoun resolution learn semantic information for zero pronoun and candidate antecedents, but tend to be short-sighted---they often make local decisions. Ideally, modeling useful information of preceding potential antecedents is critical when later predicting zero pronoun-candidate antecedent pairs. In this paper, we show how to integrate local and global decision-making by exploiting deep reinforcement learning models. Experimental results on OntoNotes 5.0 show that our technique surpasses the state-of-the-art models.

Requirements

• Python 2.7
  • Pytorch(0.4.0)
  • CUDA

Training Instructions

• Experiment configurations could be found in conf.py
• Run ./setup.sh # it builds the data for training and testing from Ontonotes data.
  • It unzip data from ./data/zp_raw_data.zip and store it in ./data/zp_data
  • It devides the training dataset into the training and develpment set. The dataset is stored as train_data
• Run ./start.sh # train the model and get results.
  • It takes about 30 minutes for pre-training
  • about 60 minutes for reinforcement learning training on GeForce 1080 Ti

Other Quirks

• It does use GPUs by default. Please make sure that the GPUs are vailable.
  • The default device utilized is gpu0, to use other GPUs, please add -gpu $DEVICE_NUMBER to the script start.sh after pretraing.py and rl.py.