Projects

Research Projects

• Zero-shot image classification with text descriptions
  • Improve on using BART as text embedding with pretraining objectives

• Amazon Alexa Prize Taskbot

• Comparing supervised and unsupervised neural machine translation with macro averaging
  • Even if a supervised NMT model and an unsupervised NMT model have the same BLEU, the SNMT model often translates with fewer untranslated words and less truncations.
• Transformer with Hard-Coded Attention [pdf] [code]
  • Modified multi-headed attention to hard-coded Gaussian attention
  • Reduced memory and inference time speed without much BLEU drop
• RNN Machine Translation Faster Decoder
  • Generate a fixed span of words at a time to speed up decoder
  • Tried to add syntax to improve performance. Proved that it will slow down the model
• Key phrase extraction
  • Literature review in short text summarization and key phrase extraction
  • Used rule-based methods, statistics based methods, and deep learning models (CSDDM) in Python to successfully extract key phrases for customer QA data