Question-Classification

Classifying questions from UIUC's CogComp QC Dataset

Classifying Questions into Coarse (6 classes) and Fine (50 classes) classes.

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.

Results

• 88.2% accuracy on Coarse classes.
• 81.6% accuracy on Fine classes.

Variations in Features Used	Coarse Set Accuracy	Coarse:Fine Set Accuracy	Fine Set Accuracy
Named Entity Recognition + Lemmas + POS Tags + Syntactic Dependency + Shape	87.8	80.4	80.8
Named Entity Recognition + Lemmas + POS Tags + Syntactic Dependency	87.2	80.6	81.4
Named Entity Recognition + Lemmas + POS Tags	88.2	81.4	81.2
Named Entity Recognition + Lemmas	86.4	80.6	81.6
Lemmas	86.2	80.4	81.6

References

https://nlp.stanford.edu/courses/cs224n/2010/reports/olalerew.pdf