Question-Tagging

Classifying Questions into 5 categories (What, Who, When, Affirmation(Yes/No) and Unknown.

Approach

1. Text Exploration
2. Text Cleaning
3. Obtaing POS Tags, Identifying Named Entities, Lemmas, Syntactic Dependency Relations and Orthographic Features.
4. Using the obtained properties as features.
5. Using a Linear SVM model on the engineered features.
6. Predict Categories of Unseens Data.

Results

Used an 80:20 test/train split to obtain 96.296% accuracy on test set.

Tagged 3000 unseen questions from here

The Tagged.csv files contains the tagged results.

The tagging_model.pkl file contains the trained LinearSVM model.

We tried a combination of features, results are below

Variations in Features Used	Test Set Accuracy
Named Entities, Lemmas, POS Tags, Syntactic Dependency, Orthography	95.96
Named Entities, Lemmas, POS Tags	96.296

References

Classifying What-type Questions by Head Noun Tagging (http://www.aclweb.org/anthology/C08-1061)