Course Overview

Why Natural Language Understanding? #

  • Perfect moment because the field is at or near peak
  • Recent resurgence of interest
  • Heavy industry use
  • Still many weaknesses in existing systems
  • Far from solved - still remain big breakthroughs to be found

Course info #

Course site: https://cs224u.stanford.edu

Code repo: https://github.com/cgpotts/cs224u

Topics #

  1. Vector-space models
  2. Sentiment analysis
  3. Contextual word representations
  4. Grounded language generation
  5. Relation extraction
  6. Natural Language Inference
  7. NLU and info. retrieval
  8. Adversarial testing
  9. Methods and metrics

Assignments #

  1. Word relatedness
  2. Cross-domain sentiment analysis
  3. Generating color descriptions in context

Each assignment culminates in a bakeoff - informal competition in which original models are entered

9/10 points for completing assignment

Bakeoff entries get full 1/10 credit if

  • they do well on the bakeoff metrics OR
  • they are creative/well-motivated, even if they do not do well on the metrics

Other entries get less than full credit

Final project/paper #

  1. Literature review
  2. Experiment protocol
  3. Final paper

Course Policies #

Grading #

  • Quizzes 5%
  • Homeworks and bakeoffs 40%
  • Literature review 10%
  • Experimental protocol 15%
  • Final paper 30%