Skip to content

Latest commit

 

History

History
79 lines (62 loc) · 2.54 KB

00-Introduction.md

File metadata and controls

79 lines (62 loc) · 2.54 KB
Raw

Introduction

Week 1

  • Linguistincs
  • Assignment (individual)

Week 2

  • Machine learning. Concepts: laplacian → ... → neural networks
  • Assignment (individual)

Week 3

  • Combine
  • Choice of projects as inspiration. Brief

Natural Language Processing:

  • People is good at using language, machines... meh
  • Languages not unique: not unique to humans (elephants, birds).

What is language:

  • Areas: broca / Wernicke in the head.
  • The whole brain is used to process language.
  • Knowledge, content, planning, imagination.
  • Velar closure: drink vs. breathing. We can open/close very quickly.
  • Animals are good with vowels.
  • Started many (thousands) of years ago, but what it's said is gone.

Written language:

  • Written language is approx 3000.
  • Not every language can be written.
  • Not "transiency" (transiency).
  • Written language extends cognition (you are "more intelligent").
  • Mapping speech to text.

Grand challenges:

  • Machine translation.
  • Sequence to sequence: I like your hat / J'aime otn chapeau.
  • Humans can teach languages to each other well.
  • Ground sound to actions: movements + objects + saying word.

Chinese room:

  • Person in a room: gets language A as in. They don't understand, but they have a translation given a production rule. Do you know then language B? Nope.
  • Symbol manipulation ≠ consciousness.

When to get translation?:

  • In two years, in five years...? Never?
  • "Every time I fire a linguistics, accuracy goes up"

Dialog:

  • We prefer to communicate with dialog.
  • Screen is black and white. Give instruction to computer: "PICK UP A BIG RED BLOCK". "OK".

Turing test:

  • Whether the thing you are talking with is a human or a computer.
  • Remove distracting factors: everything is mediated through a computer.
  • Conversation through typing.
  • Good questions for the test are the ones that involve context.

Semantic extraction:

  • "Frames" - Minsky. Scripts that we learn (go to the store, pick up a basket, pick up items, pay...).
  • Solves the problem of bureaucracy - interpreting legal documents.

Language generation:

  • Summarisation: reduce a document.
  • Descriptions: map + image + location + camera → describe.
  • Journalism: sport and stocks market. Very standarized format.
  • Timestamp + summary + prior context + ... + past event.

Information filtering:

  • What to include? There can be way too much.

Legal issues:

  • Who owns the content?
  • Who is responsible?

Question answering:

  • In: question. Out: answer.
  • Natural interrogaiton mode: "Where are you going?", "How tall is the Eiffel tower?"
  • When you talk is less natural to speak queries.