dec 3
homework
- dow problem.jpg
- 15.1 : 2nd order markov as 1st order markov; add extra states as discussed earlier in class
(1,2) - (3) - (4,5)
2nd order Markov :
P(4|3,1) can be, say 1, i.e. 1 - 3 - 4
P(5|3,2) can be, say 1, i.e. 2 - 3 - 5
so that transition prob depends on previous two states
and what comes after 3 depends on 1 or 2.
(1) - (13) - (4)
(2) - (23) - (5)
1st order Markov :
modify original to two different "3" states
... which has the same effect.
- 15.2 : umbrella for long-term future times
- 22.9 : "natural" language via three different grammars
- 22.1, 22.7, 22.14 : discuss text "understanding" and "context"
parsing
Concept/questions for class discussion:
What is the "left" and "right" they're talking about in LR vs LL descriptions?
What is a "leftmost derivation" as opposed to a "rightmost derivation"?
(see "derivations and syntax trees" discussion in wikipedia's "context free grammar")
Which is easier, "generation" or "parsing" and what are they?
22.8 in class :
context-free grammar for a**n b**n ?
context-free for palindromes ?
context-sensitive for duplicates?
And how would you parse these ?
(see /private/22-8.jpg)
How do we attach "meaning" to parse tree once we have it?
START and text retrieval
"Show the military expenditures of the 2 most populous countries in OPEC."
probabilistic language methods (chap 23)
Another approach; use conditional probabilities of n-word strings
in huge document collections to
- deduce grammar rules
- conduct searches for "meaning" without "parsing" per se
- simple models can go a long way with big statistical databases behind 'em
