TREC Question Classification - PyTorch

Description

Classification of questions into 5 classes. The five classes are as follows :

• HUM for questions about humans
• ENTY for questions about entities
• DESC for questions asking you for a description
• NUM for questions where the answer is numerical
• LOC for questions where the answer is a location
• ABBR for questions asking about abbreviations

Concepts

1. RNN
2. LSTM
3. Word Embeddings
4. Multilayer and Bi-Directional RNNs

RESULTS

Trained for 5 epochs on Kaggle.

Test Accuracy of ~91%

Model is confused between Entity questions and questions about humans.

Dataset

TREC-6

This project is in PyTorch

Requirements:

• PyTorch
• Torchtext (Used torchtext.legacy)
• Spacy

Usage

• You can pretrain the model
• You can use the predict_classes function to try your own questions.

Acknowledgements

Thanks to the author of this for this amazing repo and its explanation and help.