/basic-supervised-ner

Primary LanguagePython

Basic Supervised NER

It is the realization of the basic supervised NER model based on bert-cased model. The model contains a BERT encoder and a linear classifier.

Structure

  • utils_ner.py
    • It contains the util function and class for data loading, specifically, it provides functions for two kinds of data types
      • read_cluener_example_from_file
        • for json datatype like cluener
      • read_examples_from_file
        • for standard two column data like CONLL, each row is "word label"
  • run_ner.py
    • It is the main python script to run ner model
    • important parameters
      • data_dir: the directory where training, dev, and test data are
      • labels: the file path that contains all labels
      • model_name_or_path: backbone model to use (bert-base-cased or bert-base-chinese)
      • output_dir: the directory to output training and test result
      • sample_type: which dataloading function to use ('conll' or 'cluener')
      • max_seq_length: max length
      • num_train_epochs
      • per_gpu_train_batch_size
      • seed: global random seed
      • do_train: whether to train
      • do_eval: whether to evaluate on dev set
      • do_predict: whether to predict on test set
      • overwrite_output_dir

Usage

  • Run pip install -r requirements
  • Refer to run.sh for training script. It is used to train on cluener dataset.
  • Make sure you have files named train.txt, dev.txt, test.txt in your data_dir when you use CONLL dataloading function
  • Make sure you have files named train.json, dev.json, test.json in your data_dir when you use cluener dataloading function