Deep learning solutions for Natural Language Understanding using Keras, Tensorflow. Implemented neural models for a Chatbot virtual assistant with no feature engineering, greatly improved system performance with very limited data.
• Domain Classification: Designed Convolutional Neural Net with word embedding and beat the production SVM model.
• Utterance Slots Tagging: Implemented a Bidirectional-LSTM that automatically learns features for sequential tagging (23 tags) with 95.73% accuracy, which outperformed the production system: CRF with 55 human-designed features.