Joint Processing of Attribution Components with Contextual Embeddings

The github for the Attribute Relation Detection project for the course Natural Language Processing Technology at the Vrije Universiteit Amsterdam (2021). Students: Myrthe Buckens, Bas Diender, Marcell Fekete and Adrielli Lopes Rego

Requirements

The needed requirements can be found in requirements and installed by running pip install requirements from your terminal.

Scripts

This folder contains the following subfolders with the code files to run the project code:

data exploration

This subfolder contains code for data exploration on file and annotation level.

• annotation_level_data_collection.py: code to explore data on annotation level.
• file_level_data_collection.py: code to explore data on file level.
• specific_data_exploration.py: code to generate statistics and provide insights in the data for this project for annotation and file level insights combined.

encoding

This subfolder contains the code to generate the DistilBERT encodings for the tokens in the articles.

• bert.py: code to generate DistilBERT representations for the sentences and tokens in the file given as input.
• extract_encodings.py: code that runs the bert.py scripts and saves the output to data/encodings.

models

This subfolder contains the code for running and training the models used in this project.

• baseline.py: code to run a baseline model based on syntactic features.
• lstm.py: code to train the lstm network, after training, the model is saved.
• final_model_predictions.py: code that loads the trained and saved model and makes predictions on the test data.

processing

This subfolder contains code for processing the data in the right format for the models.

• preprocessing.py: code to clean and prepare the data set for the model input.
• training_instances.py: code to generate files per sentence as training instances for the lstm model and saves the output to data/instances.
• postprocessing.py: code to add the BIO-scheme after model predictions.

evaluation

This subfolder contains code for evaluating the final output of the models.

• evaluation.attribution.v5.pl: Evaluation script in pearl provided by Roser Morante Vallejo to compare the output of the models on partial and exact match.
• results_to_conll.py: code to convert the output files with results to conll files.

Data

This folder contains data and empty folders for output for data.

• cue_list.csv: csv file containing a list with cues for the baseline model.

models

This subfolder holds the saved and trained models, output from the lstm.py script.

• lstm_classifier_one_sentence_instances.json: json with the saved model configurations for the trained lstm network.

encodings

The following subfolders need to be created on your local device for saving the encodings output from BERT; these are too big to upload to github.

• dev-conll-foreval
• test-conll-foreval
• train-conll-foreval

instances

The following subfolders need to be created on your local device for saving the single sentence instances; these are too big to upload to github.

• dev-conll-foreval
• test-conll-foreval
• train-conll-foreval