Tariq Alhindi ([email protected])
Christopher Hidey ([email protected])
Tuhin Chakrabarty ([email protected])
Automatic Detection of propaganda is essential to build tools that can assist people to navigate the web with more awareness of deliberate or indeliberate messages of what they read.
50000 articles for task 1
21000 sentences for task 2
No preprocessing or feature-engineering done.
We experimented with deep learning approaches
Task1: BiLSTM with sizes 32, 64, 128 with and without max-pooling. We ended up using a BiLSTM of 128 with max-pooling as it got the highest F1 score on the train-dev set.
Task2: We experimented with ULMfit and BERT models for sentence classification and ended up using BERT.
F1 score on the propaganda class was used to compare between different experiments