Skip to content

Latest commit

 

History

History
62 lines (50 loc) · 4.5 KB

config_files_readme.md

File metadata and controls

62 lines (50 loc) · 4.5 KB
Raw

Configuration files

In this system, experiments are configured using JSON files. This readme explains the main attributes of three JSON configuration files:

  • build_features_config.json
  • train_config.json
  • test_config.json

Configuration file for feature extraction (build_features_config.json):

Used with the script src/features/build_features.py . To extract features for event and entity mentions, it requires two types of input files for each split (train/dev/test):

  • A json file contains its mention objects (e.g. train_event_mentions).
  • text file contains its sentences (e.g. train_text_file).

Notes:

  • The provided build_features_config.json file is configured to extract joint features for event and entity mentions (with predicate-argument structures extraction).
  • SwiRL system's output on the ECB+ corpus is provided with this repo (its directory should be assigned to the srl_output_path attribute).
  • ELMo's files (options_file, weight_file) can be downloaded from - https://allennlp.org/elmo (we used Original 5.5B model files).

Configuration file for training (train_config.json):

Used with the script src/all_models/train_model.py. The provided train_config.json file is configured to train joint model for cross-document entity and event coreference.

Most of the attributes are self-explained (e.g. batch_size and lr) , but there are few who need to be explained:

  • char_pretrained_path/char_vocab_path - initial character embeddings (provided in this repo at data/external/char_embed). The original embeddings are available at https://github.com/minimaxir/char-embeddings.
  • char_rep_size - the character LSTM's hidden size.
  • feature_size - embedding size of binary features.
  • glove_path - path to pre-trained word embeddings. We used glove.6B.300d which can be downloaded from https://nlp.stanford.edu/projects/glove/.
  • train_path/dev_path - path to the pickle files of the train/dev sets, created by the build_features script (and can be downloaded from https://drive.google.com/open?id=197jYq5lioefABWP11cr4hy4Ohh1HMPGK).
  • dev_th_range - threshold range to tune on the validation set.
  • entity_merge_threshold/event_merge_threshold - merge threshold during training (for entities/events).
  • merge_iters - for how many iterations to run the agglomerative clustering step (during both training and testing). We used 2 iterations.
  • patient - for how many epochs we allow the model continue training without an improvement on the dev set.
  • use_args_feats - whether to use argument/predicate vectors.
  • use_binary_feats - whether to use the coreference binary features.
  • wd_entity_coref_file - a path to a file (provided in this repo) which contains the predictions of a WD entity coreference system on the ECB+. We used CoreNLP for that purpose.

Configuration file for testing (test_config.json):

Used with the script src/all_models/predict_model.py . The provided test_config.json file is configured to test the joint model for cross-document entity and event coreference.

The main attributes of this configuration files are:

  • test_path - path to the pickle file of the test set, created by the build_features script (and can be downloaded from https://drive.google.com/open?id=197jYq5lioefABWP11cr4hy4Ohh1HMPGK).
  • cd_event_model_path - path to the tested event model file.
  • cd_entity_model_path - path to the tested entity model file.
  • event_merge_threshold/entity_merge_threshold - merge threshold during testing, tuned on the dev set.
  • use_args_feats- whether to use argument/predicate vectors.
  • use_binary_feats - whether to use the coreference binary features.
  • wd_entity_coref_file - a path to a file (provided) which contains the predictions of a WD entity coreference system on the ECB+. We use CoreNLP for that purpose.
  • event_gold_file_path - path to the key (gold) event coreference file (for running the evaluation with the CoNLL scorer), provided in this repo.
  • entity_gold_file_path - path to the key (gold) entity coreference file (for running the evaluation with the CoNLL scorer), provided in this repo.
  • predicted_topics_path - path to a pickle file which contains the predicted topics, provided in this repo at data/external/document_clustering or can be obtained using the code in the folder src/doc_clustering.
  • wd_entity_coref_file - a path to a file (provided in this repo) which contains the predictions of a WD entity coreference system on the ECB+. We used CoreNLP for that purpose.