The document discusses language modeling approaches for assigning probabilities to sentences, including n-gram LMs and Charniak's parsing LM. It explains how Charniak's LM uses a parse tree to represent sentences as sequences of events like pre-terminals, terminals, and constituents. It also describes Charniak's method for determiner selection by calculating probabilities of possible determiners for each NP.
Unraveling Multimodality with Large Language Models.pdf
Language Modeling in Turner&Charniak (2007)
1. Language
Modeling in
Turner&Charniak
(2007)
Kilian Evang
Language Models
N-gram LMs
Language Modeling in Charniak’s LM
Determiner
Turner&Charniak (2007) Selection
Method
Results
Reasons for Success
References
Kilian Evang
2009-11-30
2. Language
Recap: Language Models Modeling in
Turner&Charniak
(2007)
Kilian Evang
Language Models
N-gram LMs
Charniak’s LM
◮ LMs assign probabilities to sentences Determiner
Selection
◮ a sentence is a complex event Method
Results
◮ LMs break it up into a sequence of “atomic” events Reasons for Success
References
◮ each “atomic” event conditioned on certain previous
events
◮ conditional probabilities approximated by counting and
smoothing
3. Language
N-gram LMs Modeling in
Turner&Charniak
(2007)
Kilian Evang
Language Models
n-gram LMs Charniak’s LM N-gram LMs
Charniak’s LM
sequence represents sentence Determiner
Selection
p(sent) = p(seq) Method
events are words, Results
Reasons for Success
end symbols References
conditioned on the n − 1
previous
events
4. Language
A Sentence – a Sequence of Events Modeling in
Turner&Charniak
(2007)
Kilian Evang
Language Models
N-gram LMs
Sentence Charniak’s LM
Determiner
Selection
Method
Results
Reasons for Success
put the ball in the box References
Event sequence
put, the, ball, in, the, box, ∆
5. Language
A Sentence – a Sequence of Events Modeling in
Turner&Charniak
(2007)
Kilian Evang
Language Models
N-gram LMs
Sentence Charniak’s LM
Determiner
Selection
Method
Results
Reasons for Success
put the ball in the box References
Conditional probability
p(wi = the|wi −2 = ball, wi −1 = in)
6. Language
N-gram LMs vs. Charniak’s Parsing LM Modeling in
Turner&Charniak
(2007)
Kilian Evang
n-gram LMs Charniak’s LM Language Models
N-gram LMs
sequence represents sentence parse tree Charniak’s LM
p(sent) = p(seq) p(seq) Determiner
Selection
seq Method
events are words, pre-terminals, Results
Reasons for Success
end symbols terminals, References
constituents,
end symbols
conditioned on the n − 1 certain previous
previous events, depending
events on type
7. Language
A Parse Tree – a Sequence of Events Modeling in
Turner&Charniak
(2007)
Kilian Evang
Parse tree
vp Language Models
N-gram LMs
Charniak’s LM
Determiner
np pp Selection
Method
Results
Reasons for Success
np References
verb det noun prep det noun
put the ball in the box
8. Language
A Parse Tree – a Sequence of Events Modeling in
Turner&Charniak
(2007)
Kilian Evang
Parse tree
vp Language Models
N-gram LMs
Charniak’s LM
Determiner
np pp Selection
Method
Results
Reasons for Success
np References
verb det noun prep det noun
put the ball in the box
Event sequence
verb, put, M, ∆, M, np, pp, ∆, noun, ball, M, det, ∆, M, ∆, the,
prep, in, M, ∆, M, np, ∆, noun, box, M, det, ∆, M, ∆, the
9. Language
Digression: Non-head Constituents Modeling in
Turner&Charniak
(2007)
Kilian Evang
Language Models
Tree fragment N-gram LMs
Charniak’s LM
l Determiner
Selection
Method
Results
Lm ... L1 t R1 ... Rn Reasons for Success
References
h
Event sequence fragment
M, L1 , . . ., Lm , ∆, M, R1 , . . ., Rn , ∆
10. Language
A Parse Tree – a Sequence of Events Modeling in
Turner&Charniak
(2007)
Kilian Evang
Parse tree
Language Models
vp N-gram LMs
Charniak’s LM
Determiner
np pp Selection
Method
Results
Reasons for Success
np References
verb det noun prep det noun
put the ball in the box
Conditional probability for a head pre-terminal
p(t = noun|l = np, m = vp, u = verb, i = put)
11. Language
A Parse Tree – a Sequence of Events Modeling in
Turner&Charniak
(2007)
Kilian Evang
Parse tree
Language Models
vp N-gram LMs
Charniak’s LM
Determiner
np pp Selection
Method
Results
Reasons for Success
np References
verb det noun prep det noun
put the ball in the box
Conditional probability for a head terminal
p(h = ball|t = noun, l = np, m = vp, u = verb, i = put)
12. Language
A Parse Tree – a Sequence of Events Modeling in
Turner&Charniak
(2007)
Parse tree Kilian Evang
vp Language Models
N-gram LMs
Charniak’s LM
Determiner
np pp Selection
Method
Results
Reasons for Success
np
References
verb det noun prep det noun
put the ball in the box
Conditional probability for a non-head constituent
p(Li = det|Li −1 = M, h = ball, t = noun, l = np, m =
vp, u = verb)
13. Language
Overview: Conditioning Modeling in
Turner&Charniak
(2007)
Kilian Evang
event type conditioned on
Language Models
head pre-terminal t constituent label l,
N-gram LMs
mother constituent label m, Charniak’s LM
mother constituent head pre-terminal u Determiner
Selection
mother consitutent head terminal i
Method
head terminal h head pre-terminal t, Results
Reasons for Success
constituent label l,
References
mother constituent label m,
mother constituent head pre-terminal u
mother consitutent head terminal i
non-head (part of) L1...i −1 (L1...m , R1...i −1 ),
constituent label Li (Ri ), head terminal h,
end symbol ∆ head pre-terminal t,
constituent label l,
mother constituent label m,
mother constituent head pre-terminal u
14. Language
Determiner Selection Modeling in
Turner&Charniak
(2007)
Kilian Evang
Language Models
N-gram LMs
Charniak’s LM
Determiner
◮ for each NP, Selection
Method
◮ for each possible determiner (the, a/an, null), Results
Reasons for Success
◮ determine probability of NP
References
◮ choose determiner resulting in highest probability
◮ note: sufficient to determine probabilities for events
that differ
15. Language
Determiner Selection – Example Modeling in
Turner&Charniak
(2007)
Kilian Evang
◮ “put [NP the ball] in the box”
Language Models
◮ p(L1 = det|m = vp, u = verb, l = np, t = noun, h = N-gram LMs
Charniak’s LM
ball) × p(L2 = ∆|m = vp, u = verb, l = np, t =
Determiner
noun, h = ball, L1 = det) × p(det → the|m = vp, u = Selection
verb, l = np, t = noun, h = ball, L1 = det) Method
Results
Reasons for Success
◮ “put [NP a/an ball] in the box”
References
◮ p(L1 = det|m = vp, u = verb, l = np, t = noun, h =
ball) × p(L2 = ∆|m = vp, u = verb, l = np, t =
noun, h = ball, L1 = det) × (p(det → a|m = vp, u =
verb, l = np, t = noun, h = ball, L1 = det) + p(det →
an|m = vp, u = verb, l = np, t = noun, h = ball, L1 =
det))
◮ “put [NP ball] in the box”
◮ p(L1 = ∆|m = vp, u = verb, l = np, t = noun, h = ball)
16. Language
Results Modeling in
Turner&Charniak
(2007)
Kilian Evang
Language Models
N-gram LMs
Charniak’s LM
Determiner
Selection
Method
Results
Reasons for Success
References
17. Language
Reasons for Success Modeling in
Turner&Charniak
(2007)
Kilian Evang
Language Models
N-gram LMs
Charniak’s LM
Determiner
◮ syntactic structure allows for long-distance conditioning, Selection
e.g. Method
Results
Reasons for Success
◮ he [VP gave [NP the sultan of Brunei] [NP a cactus]]
References
◮ constituent head enforces selectional preferences,
reflected in head-first strategy
◮ ...
18. Language
References Modeling in
Turner&Charniak
(2007)
Kilian Evang
Language Models
Eugene Charniak (2000) N-gram LMs
A Maximum-Entropy-Inspired Parser Charniak’s LM
Determiner
Proceedings of the First Meeting of the North American Selection
Chapter of the Association for Computational Linguistics Method
Results
Reasons for Success
Eugene Charniak (2001) References
Immediate-Head Parsing for Language Models
Proceedings of the 39th Annual Meeting of the
Association for Computational Linguistics
Jenine Turner & Eugene Charniak (2007)
Language Modeling for Determiner Selection
Proceedings of NAACL HLT 2007, Companion Volume