Given/new information and the discourse coherence problem Micha - - PowerPoint PPT Presentation
Given/new information and the discourse coherence problem Micha Elsner joint work with: Eugene Charniak Joseph Browne Given/new information Unfamiliar information: Sir Walter Elliot, of Kellynch Hall, in Somersetshire , was a man
2
– Sir Walter Elliot, of Kellynch Hall, in
– Sir Walter had improved it...
– He had been remarkably handsome in his
3
– Earlier sentences make later ones more
4
5
– Hard even for long documents.
6
P l a n A i r p l a n e C
d i t i
F l i g h
– Nothing to say about
7
8
– because you don't know who he is.
9
– because you don't know who he is.
10
– because you don't know who he is.
11
– because you don't know who he is.
12
– Sir, Professor (usually new); Prof., Inc. (usually old)
Vieira+Poesio '00 Ng+Cardie '02 Uryupina '03 ...
13
– Linguists (Hawkins '78, Gundel '93 and
– A question of rules.
– Relatives (Fox+Thompson '90) – Various modifiers (Fraurud '90, Vieira+Poesio
– A question of typicality.
14
– Don't resolve the new NPs. – Do resolve the old ones.
Joint decisions: Denis+Baldridge '07 Sequential: Poesio+al '05 Ng+Cardie '02
15
Sir Walter Elliot, of Kellynch Hall, in Somersetshire Walter Elliot Sir Walter Sir Walter Elliot Sir Walter he his himself
Sir Walter Elliot, of Kellynch Hall, in Somersetshire Walter Elliot Sir Walter Sir Walter Elliot Sir Walter he his himself
16
Sir Walter Elliot, of Kellynch Hall, in Somersetshire Walter Elliot Sir Walter Sir Walter Elliot Sir Walter he his himself
new
17
– (I'll talk more about this later).
– Works about 2/3 of the time (Poesio+Vieira). – Means we can't use the same head feature to
18
Sir Walter Elliot, of Kellynch Hall, in Somersetshire Walter Elliot Sir Walter Sir Walter Elliot Sir Walter he his himself
new
new
19
– Unique referents: The FBI, the Golden
– Our technique will mislabel these.
– singleton: no subsequent coreferent NPs – often look more like old than new
corpus study: Fraurud '90 classifiers: Bean+Riloff '91 Uryupina '03
20
– to construct a joint generative model. – Principled, but mixtures might improve?
21
– P(label) P(syntax | label) – Modifiers generated by Markov chains.
– As a classifier. – And as a coherence model.
22
– P(chain) = Π P( syntax , label) – Π P( label | syntax ) P(syntax)
– P(chain) ~ Π P( label | syntax )
23
24
25
26
27
P(words | he) ≈ 0 P(yacht | he) ≈ 0
28
– So we need P(text | antecedent)
– Ge+Hale+Charniak '98 – Accuracy 79.1% (on markables)
29
Probability that the antecedent is a given how far away a is, and how
30
Probability that the antecedent is a given how far away a is, and how
Probability of the pronoun gender given the antecedent.
31
Probability that the antecedent is a given how far away a is, and how
Probability of the pronoun gender given the antecedent. Probability of the pronoun number given the antecedent.
32
Probability that the antecedent is a given how far away a is, and how often it has
33
– Usually one structure has all the mass
a
34
– On its own, this model is not as strong as the
35
– Not all the cases are this hard. – But so far, no advantage.
36
– A generative model now exists:
– Model the decision to generate pronoun or
– Doesn't work! We don't know why...
37
– Open quote “ comes before close quote “ – The stuff inside should be relatively short. – We can model this...
– Based on discourse patterns. – Not just typography!
38
– Almost always “real” speech. – Unlikely in first sentence.
39
40
– Dr. Evil doesn't do this, which is part of the
41
– Works for things other than nouns. – “recombinant” DNA – The Fed appears to be “sterilizing” the
42
43
– Bush called Mr. Clymer a “major league
– “You” is a second person pronoun.
– Yeah, that's really “helpful”!
– Chaucer's “Book of the Duchess”
44
– Precision 92, recall 24 – Works well, but only on a few documents.
45
46
– The plane crashed. The pilot was injured. – Looks like a familiar (discourse-old) NP. – But really a new entity. – Similar to unique NPs (the FBI), but licensed
Poesio+Vieira+Teufel '97 Poesio+al '04
47
– Soricut+Marcu '06, related to Lapata '03
48
49
– Ex: words prefer to align to subjects. – Biases learned during EM (IBM model 2).
50