This repository contains code for SEMEVAL 2020 TASK 11 "DETECTION OF PROPAGANDA TECHNIQUES IN NEWS ARTICLES." Our system was ranked 5th for the Technique Classification subtask. The paper containing the system description can be found here and has been accepted for publication at SemEval2020. The source code is available in the src directory.
Note that this code uses version 2 of transformers library (2.8.0)
$ python src/train_SI.py
The trained models will be stored in model_dir. Note that currently we save model if number of trainining articles are greater than 150 (This can be changed here)
$ python get_outputs.py
This outputs the predicted spans for article in input.txt. The predicted spans are shown in bold.
If the model is not present in model_dir, it downloads our BERT_BIOE model avalaible here and stores it in model_dir. To make prediction for any custom model, change the model path in src/predict_SI.py
To get outputs for dev-set (or any other folder)
$ python get_outputs.py --dev-output True
$ python src/train_SI.py
The trained models will be stored in model_dir.
$ python src/predict_TC.py
This outputs the techniques of given spans.
Interactive iPython notebooks for both tasks and related experiments can be found in Colab Notebooks directory.