Chris Riesbeck's guest lecture

---

10/19/04, Notes by Praveen Paritosh.
N.B. This will make sense in the context of the discussion, and is not meant to be complete summary of the class.

Different fields that study language:

* Linguistics
    Structural Linguistics: what kind of grammar rules can you have?
    Anthropology

* Computational Linguistics (Text processing)
      Search Engines
      Concordance
      Online dictionaries
      Grammar checker

*Artificial Intelligence
      Whats an accomplishment?
           To build programs that can read and learn
           NL interfaces. [Text and Speech]

* Computer Science

---

The computational linguistics/AI bridge (or divide)

Machine translation is something that both fields addressed, albeit in a substantially different manner. Meaning, understanding, and representation were important concerns to AI researchers, and not so much to CL researchers. That came from a practical reason: representations are hard, and a philosophical reason: Chomsky's forceful argument that meaning is too messy to be an object of linguistic study.

Computational Linguistics
                   |                          NLP: Japanese -> Dictionary -> English : No representations, easier/practical. The goal of Linguistics should be uncovering the structure of language, e.g., how and why "Colorless green ideas sleep furiously" makes sense thought it doesnt mean anything. "We will have to put an encyclopedia in our theories in linguistics". So Linguistics can study language structure without studying meaning. Its messy and not the goal of study.
    (Machine Translation)
                    |                          NLU: Japanese -> Meaning -> English
Artificial Intelligence

---

"Classic picture" of NLU

Text - John kissed Mary,
   |
Parser - Parse trees - S - NP John  S- VP - V - Kissed  S- VP - NP - Mary
   |
Semantic Interpretation - NP "John" - male human named "John"
  |  - (very few projects get to here)
Pragmatics interpreter - actual instances (Episodic memory)

The knowledge migration up this hierarchy?

I'm over there. (Riesbeck): need to know the Bank script to understand this.
George thinks vanilla. (Lakoff)
The cue-ball kissed the eight-ball. (Riesbeck?)
...
context
 
DMAP "short-circuits the above picture and goes directly from text to episodic memory.

---

Last updated by paritosh, 10/19/04, 3pm.