Lecture 20 – Future of NLP + Deep Learning

#scaling #unsupervised learning #risks #social impact

We want to utilize the unlabeled data for low-resource languages and also because of rare labeled data;

Unsupervisation in Machine Translation

  • Pre-training: train two LMs with unlabeled data;
  • Back-translation: back translate the source and target sentences. Both models have perfect target reference translation;

Unsupervised Word Translation

Align two word embedding spaces by learning an asymmetrical transformation matrix.

  • Discriminator: given an embedding if it comes from space \(X\) or space \(Y\)(\(WX\))

Unsupervised NMT

  • De-noising auto-encoder: to capture input meaning
  • Back translation: the same model but different BOS(<Fr> and <En>)

Cross-lingual BERT

  • Language embeddings

Social Impact

  • decision making (bias)

Future Directions

  • Multitasking Training
  • More difficult tasks
  • Low-resource settings:
    • edge devices
    • low-resource data
    • meta-learning
  • Interpretability
  • Dialogue
  • Healthcare