Transcript Additional comments on Cohen and Sarawagi paper

IE with Dictionaries
Cohen & Sarawagi
Announcements
• Current statistics:
–
–
–
–
days
with unscheduled student talks: 2
students with unscheduled student talks: 0
Projects are due: 4/28 (last day of class)
Additional requirement: draft (for comments)
no later than 4/21
Finding names you know about
• Problem: given dictionary of names, find them in
email text
– Important task beyond email (biology, link analysis,...)
– Exact match is unlikely to work perfectly, due to
nicknames (Will Cohen), abbreviations (William C) ,
misspellings (Willaim Chen), polysemous words (June,
Bill), etc
– In informal text it sometimes works very poorly
– Problem is similar to record linkage (aka data
cleaning, de-duping, merge-purge, ...) problem of
finding duplicate database records in heterogeneous
databases.
Finding names you know about
• Technical problem:
– Hard to combine state of the art similarity
metrics (as used in record linkage) with state of
the art NER system due to representational
mismatch:
• Opening up the box, modern NER systems don’t
really know anything about names....
IE as Sequential Word Classification
A trained IE system
models the relative
probability of
labeled sequences
of words.
person name
location name
background
To classify, find the most likely state sequence for the given words:
Yesterday Pedro Domingos spoke this example sentence.
Any words said to be generated by the designated “person name”
state extract as a person name:
Person name: Pedro Domingos
IE as Sequential Word Classification
Modern IE systems use a rich representation for words, and
clever probabilistic models of how labels interact in a
sequence, but do not explicitly represent the names extracted.
wt - 1
identity of word
ends in “-ski”
is capitalized
is part of a noun phrase
is “Wisniewski”
is in a list of city names
is under node X in WordNet
part of
ends in
is in bold font
noun phrase
“-ski”
is indented
O t 1
is in hyperlink anchor
last person name was female
next two words are “and Associates”
wt
wt+1
…
…
Ot
O t +1
Semi-Markov models for IE
with Sunita Sarawagi, IIT Bombay
• Train on sequences of
labeled segments, not
labeled words.
S=(start,end,label)
• Build probability
model of segment
sequences, not word
sequences
• Define features f of
segments
• (Approximately)
optimize feature
weights on training
data
f(S) = words xt...xu, length, previous
words, case information, ..., distance to
known name
m
maximize:
 log Pr(S
i 1
i
| xi )
Details: Semi-Markov model
Segments vs tagging
t
1
x
y
Fred
Person
2
please
other
3
stop
other
4
by
other
5
6
my
office
Loc
Loc
7
this
8
afternoon
other
Time
t4=u4=7
t5=u5=8
f(xt,yt)
t,u
x
y
t1=u1=1
Fred
Person
t2=2, u2=4
please
stop
other
t3=5,u3=6
by
my
office
Loc
f(xj,yj)
this
other
afternoon
Time
Details: Semi-Markov model
Conditional Semi-Markov models
CMM:
CSMM:
A training algorithm for CSMM’s (1)
Review: Collins’ perceptron training algorithm
Correct tags
Viterbi tags
A training algorithm for CSMM’s (2)
Variant of Collins’ perceptron training algorithm:
voted
perceptron
learner for
TTRANS
like Viterbi
A training algorithm for CSMM’s (3)
Variant of Collins’ perceptron training algorithm:
voted
perceptron
learner for
TTRANS
like Viterbi
A training algorithm for CSMM’s (3)
Variant of Collins’ perceptron training algorithm:
voted
perceptron
learner for
TSEGTRANS
like Viterbi
Viterbi for HMMs
Viterbi for SMM
Sample CSMM features
Experimental results
• Baseline algorithms:
– HMM-VP/1: tags are “in entity”, “other”
– HMM-VP/4: tags are “begin entity”, “end entity”,
“continue entity”, “unique”, “other”
– SMM-VP: all features f(w) have versions for “f(w) true for
some w in segment that is first (last, any) word of segment”
– dictionaries: like Borthwick
• HMM-VP/1: fD(w)=“word w is in D”
• HMM-VP/4: fD,begin(w)=“word w begins entity in D”,
etc, etc
• Dictionary lookup
Datasets used
Used small training sets (10% of available) in experiments.
Results
Results: varying history
Results: changing the dictionary
Results: vs CRF
Results: vs CRF