Ayush Sehgal, Shreyans Sethi, Gayatri Babel
This project serves as an NLP exploration into quantifying condescension in New York Times articles. This was made as part of the Annotation Project in the Info159: Introduction to Natural Language Processing course at UC Berkeley.
The project includes 1000 data points pulled from NY Times Archive. Articles were rated on their headline and description as found in the archives page. Only articles between 2011-2022 were used. This data split into train, dev, test sets can be found in the data folder.
In depth guidelines for an individual to rate each articles are provided in guidelines.pdf where we use a rating system ranging from 1, being not condescending to 5 being very condescending. We have also defined what we mean by condescending in the document.
All the work done to develop the classifier with feature engineering is present in the classifier.ipynb. The classifier as it stands today reaches a final test accuracy of 79%, which beats the baseline majority class classifier, which has an accuracy of 76%.
In analysis.pdf we report how the classifier did and reflect on its performance.
Finally, datasheet.md contains reflection on the experimentation process as well as on the process of choosing the dataset.