Transcript prob-lex-parsing

Probabilistic and Lexicalized Parsing
CS 4705
Probabilistic CFGs: PCFGs
• Weighted CFGs
– Attach weights to rules of CFG
– Compute weights of derivations
– Use weights to choose preferred parses
• Utility: Pruning and ordering the search space, disambiguate,
Language Model for ASR
• Parsing with weighted grammars: find the parse T’ which
maximizes the weights of the derivations in the parse
tree for all the possible parses of S
• T’(S) = argmaxT∈τ(S) W(T,S)
• Probabilistic CFGs are one form of weighted CFGs
Rule Probability
• Attach probabilities to grammar rules
• Expansions for a given non-terminal sum to 1
R1: VP  V
.55
R2: VP  V NP
.40
R3: VP  V NP NP
.05
• Estimate probabilities from annotated corpora
– E.g. Penn Treebank
– P(R1)=counts(R1)/counts(VP)
Derivation Probability
• For a derivation T= {R1…Rn}:
– Probability of the derivation:
n
P (T )   P ( Ri )
i 1
• Product of probabilities of rules expanded in tree
– Most likely probable parse:
– Probability of a sentence:
T *  arg max P(T )
T
P( S )   P(T | S )
T
• Sum over all possible derivations for the sentence
• Note the independence assumption: Parse
probability does not change based on where the
rule is expanded.
One Approach: CYK Parser
• Bottom-up parsing via dynamic programming
– Assign probabilities to constituents as they
are completed and placed in a table
– Use the maximum probability for each
constituent type going up the tree to S
• The Intuition:
– We know probabilities for constituents lower in
the tree, so as we construct higher level
constituents we don’t need to recompute
these
CYK (Cocke-Younger-Kasami) Parser
• Bottom-up parser with top-down filtering
• Uses dynamic programming to store intermediate results
(cf. Earley algorithm for top-down case)
• Input: PCFG in Chomsky Normal Form
– Rules of form Aw or ABC; no ε
• Chart: array [i,j,A] to hold probability that non-terminal A
spans input i-j
– Start State(s): (i,i+1,A) for each Awi+1
– End State: (1,n,S) where n is the input size
– Next State Rules: (i,k,B) (k,j,C)  (i,j,A) if ABC
• Maintain back-pointers to recover the parse
Structural Ambiguity
•
•
•
•
•
• NP  John | Mary | Denver
• V -> called
• P -> from
S  NP VP
VP  V NP
NP  NP PP
VP  VP PP
PP  P NP
John called Mary from Denver
S
S
VP
NP
NP
PP
VP
V
John
called
VP
NP
P
NP
NP
Mary from Denver
John
V
NP
called
Mary
PP
P
NP
from Denver
Example
John
called
Mary
from
Denver
Base Case: Aw
NP
P
NP
V
NP
John
called
Mary
from
Denver
Recursive Cases: ABC
NP
P
NP
X
V
NP
called
John
Mary
from
Denver
NP
P
VP
NP
X
V
Mary
NP
called
John
from
Denver
NP
X
P
VP
NP
from
X
V
Mary
NP
called
John
Denver
PP
NP
X
P
Denver
VP
NP
from
X
V
Mary
NP
called
John
S
NP
John
PP
NP
X
P
Denver
VP
NP
from
V
Mary
called
PP
NP
Denver
X
X
P
S
VP
NP
from
X
V
Mary
NP
called
John
NP
X
S
VP
NP
X
V
Mary
NP
called
John
PP
NP
P
Denver
from
NP
PP
NP
Denver
X
X
X
P
S
VP
NP
from
X
V
Mary
NP
called
John
VP
NP
PP
NP
X
X
X
P
Denver
S
VP
NP
from
X
V
Mary
NP
called
John
VP
NP
PP
NP
X
X
X
P
Denver
S
VP
NP
from
X
V
Mary
NP
called
John
NP
PP
NP
X
VP1
VP2
X
X
P
Denver
S
VP
NP
from
X
V
Mary
NP
called
John
S
NP
PP
NP
X
VP1
VP2
X
X
P
Denver
S
VP
NP
from
X
V
Mary
NP
called
John
S
VP
NP
PP
NP
X
X
X
P
Denver
S
VP
NP
from
X
V
Mary
NP
called
John
Problems with PCFGs
• Probability model just based on rules in the derivation.
• Lexical insensitivity:
– Doesn’t use words in any real way
– But structural disambiguation is lexically driven
• PP attachment often depends on the verb, its object, and the
preposition
• I ate pickles with a fork.
• I ate pickles with relish.
• Context insensitivity of the derivation
– Doesn’t take into account where in the derivation a rule is used
• Pronouns more often subjects than objects
• She hates Mary.
• Mary hates her.
• Solution: Lexicalization
– Add lexical information to each rule
– I.e. Condition the rule probabilities on the actual words
An example: Phrasal Heads
• Phrasal heads can ‘take the place of’ whole
phrases, defining most important characteristics
of the phrase
• Phrases generally identified by their heads
– Head of an NP is a noun, of a VP is the main verb, of a
PP is preposition
• Each PFCG rule’s LHS shares a lexical item with
a non-terminal in its RHS
Increase in Size of Rule Set in Lexicalized
CFG
• If R is the number of binary branching rules in
CFG and ∑ is the lexicon, O(2*|∑|*|R|)
• For unary rules: O(|∑|*|R|)
Example (correct parse)
Attribute grammar
Example (less preferred)
Computing Lexicalized Rule Probabilities
• We started with rule probabilities as before
– VP  V NP PP
P(rule|VP)
• E.g., count of this rule divided by the number of
VPs in a treebank
• Now we want lexicalized probabilities
– VP(dumped)  V(dumped) NP(sacks)
PP(into)
• i.e., P(rule|VP ^ dumped is the verb ^ sacks is the
head of the NP ^ into is the head of the PP)
– Not likely to have significant counts in any
treebank
Exploit the Data You Have
• So, exploit the independence assumption and
collect the statistics you can…
• Focus on capturing
– Verb subcategorization
• Particular verbs have affinities for particular VPs
– Objects’ affinity for their predicates
• Mostly their mothers and grandmothers
• Some objects fit better with some predicates than
others
Verb Subcategorization
•
Condition particular VP rules on their heads
– E.g. for a rule r VP -> V NP PP
•
P(r|VP) becomes P(r ^ V=dumped | VP ^
dumped)
– How do you get the probability?
•
•
How many times was rule r used with dumped,
divided by the number of VPs that dumped
appears in, in total
How predictive of r is the verb dumped?
– Captures affinity between VP heads (verbs)
and VP rules
Example (correct parse)
Example (less preferred)
Affinity of Phrasal Heads for Other Heads:
PP Attachment
• Verbs with preps vs. Nouns with preps
• E.g. dumped with into vs. sacks with into
– How often is dumped the head of a VP which
includes a PP daughter with into as its head
relative to other PP heads or… what’s
P(into|PP,dumped is mother VP’s head))
– Vs…how often is sacks the head of an NP
with a PP daughter whose head is into
relative to other PP heads or…
P(into|PP,sacks is mother’s head))
But Other Relationships do Not Involve
Heads (Hindle & Rooth ’91)
• Affinity of gusto for eat is greater than for
spaghetti; and affinity of marinara for
spaghetti is greater than for ate
Vp (ate)
Vp(ate)
Pp(with)
np
v
Ate spaghetti with gusto
Vp(ate)
Np(spag)
np Pp(with)
v
Ate spaghetti with marinara
Log-linear models for Parsing
• Why restrict to the conditioning to the elements
of a rule?
– Use even larger context…word sequence,
word types, sub-tree context etc.
• Compute P(y|x); where fi(x,y) tests properties of
context and li is weight of feature
e i i
P( y | x) 
li * f i ( x , y )

e
l * f ( x, y )

yY
• Use as scores in CKY algorithm to find best
parse
Supertagging: Almost parsing
Poachers now control the underground trade
N
S
VP
N
N
Adv
VP
S
NP
NP
V
NP
control e
S
N
Adv
poachers
now
S
V
e
NP
VP
NP
NP VP
poachers
now
Det
NP
:
:
Adj
trade
N
NP VP
NP
control
N
N
trade
S
NP
S
N
underground
S
V
:
:
e
Adj
control
Adv
N
N
N
V
VP
NP
underground
the
NP VP
V
S
S
NP
NP VP
NP VP
now
poachers
S
S
S
NP VP
e V
e
NP
NP VP
V
NP
e
N
Adj
underground
:
trade
Summary
• Parsing context-free grammars
– Top-down and Bottom-up parsers
– Mixed approaches (CKY, Earley parsers)
• Preferences over parses using probabilities
– Parsing with PCFG and PCKY algorithms
• Enriching the probability model
– Lexicalization
– Log-linear models for parsing
– Super-tagging