Lecture #25 (16 April 2002)

Natural Language Processing

Overall Reading
Brookshear: pp. 473-475
Decker/Hirshfield: pp. 305-309

Outline:

  • Natural Languages
  • Goals
  • Comprehension
  • Translation
  • Generating new content
  • Coping with Ambiguity
  • Syntactic ambiguity
  • Semantic ambiguity
  • The need for contextual information
  • The need for rules of conversation
  • The need for real-world, topical knowledge

    • Natural Languages

    We have developed many "high-level" languages for programming computers. But these languages are generally very structured and avoid ambiguity, so that they can be precisely interpretted.

    Humans have long developed languages for communicating with each other. We will refer to these as "natural" languages.

    Goals

    We consider whether computers can process natural languages. We will look at three distinct levels of ability:
  • Understanding language
  • Translating languages
  • Generating new content

    • Let's walk through the issues for each individual ability.

  • Understanding language

    What are the minimum requirements for understanding language?

    Must understand grammer (syntactic analysis)
    Must understand vocabulary (semantic analysis)
    But we actually use far more knowledge (contextual analysis)

  • Translating languages
    To really do this well, you need to have good comprehension of first language, and then also know how to express the content in the second language.

    Certainly cannot just translate words:
    "The spirit is willing, but the flesh is weak."
    [translated to Russion equivalent of]
    "The vodka is acceptable, but the meat has spoiled."

    Far beyond this, expression can be effected by the nuances, symbolism and historical context of the choice of words and expressions.

  • Generating new content
    If serving as a translator, at least it is known what sentiment to express. How might a machine generate original content (dare I say 'thoughts')?
  • How do we create poetry? (Lab 9.2 [DH])
  • Literature experts claim to be able to recognize authentic samples of Shakespeare's writing versus fakes. Can a machine recognize such a "style" and if so, can it create new stories in the style of Shakespeare.
  • Partaking in conversation.
    Example of a psychoanalyst (ELIZA) (DeeDee)

    • Coping with Ambiguity

    Let's go back to the first step, namely comprehending language. We are used to coping with many potential ambiguities in natural languages. The source of ambiguity may rise from a combination of factors:

  • Syntactic ambiguity

    Examples:

  • "John met Bill before he went to the store."

    Who was the person who later went to the store?

  • "They are racing horses."

    Is the word 'racing' a verb or an adjective? Is the sentance describing what kind of horses they are, or is it telling what some people are doing?

  • "Stampeding cattle can be dangerous."

  • Semantic ambiguity

    Even if you know the part of speech, certain words have several meanings.

    Examples:

  • "Ron lies asleep in his bed."

    'lies' is a verb. But does it mean to recline or to deceive?

  • "Cinderella had a ball."

  • The need for contextual information

    Examples:

  • "The clams are ready to eat."

    Do you interpret this differently if you hear it in a restaurant versus in an aquarium?

  • "Do you know what time it is?"

    Does this have a different meaning if said by a stranger on the street than from your boss, after you walk into a meeting 20 minutes late?

  • "The bat slipped from his hand."

    Whose hand? Are we talking about a baseball player or a zoo keeper?

  • "Cinderella had a ball."

    Did she have a round object, a good time, or a formal dance? Perhaps reading more of the story would help us in understanding.

  • The need for rules of conversation

    Examples:

  • "Do you know what time it is?"

    Let's assume we already know that this is asked by a stranger on the street. Do you think the stranger is expecting you to answer, "Yes, I do."

  • "Do you know that your tire is flat?"

    Almost certainly not. This is not really a question. It is informative.

  • "I'd like to ask everyone to hold a hand up for two seconds."

    Was I simply expression my feelings or was there a question/request in this sentance?

  • The need for real-world, topical knowledge

    Examples:

  • "Norman Rockwell painted people.

    Fortunately, I happen to know that he placed his paint on canvas (as opposed to skin).

  • "Sally was fed up. She got up angrily from ther table at the restaurant and left just enough to cover the check. The waitress sneered at her as she walked out."

    So why was the waitress upset? Because it is customary to leave a tip in addition to paying the check. Of course we need to have knowledge about this real-world custom to be able to understand. No part of the text is going to make this explicit.

  • "Ron lies asleep in his bed."

    Well, it seems unlikely that someone can tell a fib while sleeping, so I probably assume that this meant Ron was reclined.

    • comp150 Class Page
    mhg@cs.luc.edu
    Last modified: 11 April 2002