computational linguistics

Tags: languages/linguistics

• two possible motivations for research
  1. technological: best performance on some applied task
  2. cognitive: single model with human-like performance across a broad range of abilities, plus biological/cognitive plausibility
• what does success look like?
  • most nlu takes a behavioral approach, a modelis assumed to learn/understand iff
    • can perform the task
    • on data that wasn’t used to build/train it
    • (possibly) makes errors and uses resources similarily to humans
  • note that this sets aside the question of “what is understanding”, “what/when does conciousness emerge”, etc
• what is the background for nlu need?

  • two main questions:

    1. how much does a system need?
    2. how much can it learn from data, how much built-in knowledge does it need?

  • one big trend is that learning from the data is actually better than building explicit knowledge

    Every time I fire a linguist, the performance of the speech recongizer goes up - apocryphal, Fred Jelinek, 1988

  • relates to are experts real?