fair95/controversy_detection

IC_Project

This is the repository for the MSc Computing project

Description of the files

• trainer.py: contains our Trainer class
• loss.py: contains custom loss functions that we used
• cleaning.py: script to clean the Refinitiv raw data
• data_analysis.py: contains a class named Preprocessor to process data for the baseline model
• embedding_viewer: contains a class named BertEmbeddingView for model interpretation and visualisation
• data.py: contains the PyTorch dataset class to feed the inputs
• model.py: contains the Model class we used.
• my_tokenizers: contains Word-level tokeniser for the LSTM model
• xgb_train.py: contains baseline XGBoost model
• metrics.py: contains the evluation metrics used in the project
• FGM.py: Adverserial attack training techniques
• utils.py: contains useful functions for data analsis and manipulation
• main.py: the main script that runs the experiment
• transfer.py: contains the transfer learning experiment
• models: directory contains saved models
• emb_vis: directory contains visualisations for model interpretability
• result: directory contains results of the test set

Required packages

transformers 4.16.2
spacy 3.2.0
torch 1.10.0
numpy 1.19.5
pandas 1.1.5
scikit-learn 0.24.2
tqdm 4.62.3
xgboost 1.4.2
captum 0.5.0 upsetplot 0.6.1 beautifulsoup4 4.6.3 wordcloud 1.8.2.2 seaborn 0.11.2 matplotlib 3.2.2

How to run the classification

• To replicate the experiment, please first place the data in csv format in the current folder with a column story that contains the news stories. Then you can have as many controversial topics as you need, each with an individual column where 1 represent the presence of the controversy.

• Run cleaning.py by changing the file name to your placed csv file. The resulted cleaned file is named as cleaned_2.csv, of course you can change the name as you wish.

• Please modify config varaible in the main.py to test different hyperparameters/tricks, or leave it as it is for the best model found in the project.

• To train deep learning model, run python3 main.py. To train baseline model, run python3 xgb_train.py

How to run interpretation

• Have your trained model ready, e.g. models/ProsusAI_finbert_head_3e-06_10_512_False_None_saved_model.pt in the models folder

• Run python3 embedding_viewer.py

How to run transfer learning

• Have your trained model ready in models folder

• Have target domain data source available, e.g. 'twitter_data.csv'

• Run python3 transfer.py

Important notes

• Please make sure you have all required libraries installed

• MyBertModel is the base class the the BERT model implementation. Use MyBertModel_1 for default classification head and MyBertModel_2 for the customised classification head.

• The model is saved in models directory.