Edit Distances

Edit Distances

William W. Cohen

Midterm progress reports

• Talk for 5min per team– You probably want to have one person speak

• Talk about– The problem & dataset– The baseline results– What you plan to do next

• Send Brendan 3-4 slides in PDF by Mon night

Plan for this week• Edit distances

– Distance(s,t) = cost of best edit sequence that transforms st

– Found via….

• Learning edit distances– Probabilistic generative model: pair HMMs

– Learning now requires EM• Detour: EM for plain ‘ol HMMS

– EM for pair HMMs

• Why EM works• Discriminative learning for pair HMMs

Motivation

• Common problem: classify a pair of strings (s,t) as “these denote the same entity [or similar entities]”– Examples:

• (“Carnegie-Mellon University”, “Carnegie Mellon Univ.”)

• (“Noah Smith, CMU”, “Noah A. Smith, Carnegie Mellon”)

• Applications:– Co-reference in NLP

– Linking entities in two databases

– Removing duplicates in a database

– Finding related genes

– “Distant learning”: training NER from dictionaries

Levenshtein distance - example

• distance(“William Cohen”, “Willliam Cohon”)

W I L L I A M _ C O H E N

W I L L L I A M _ C O H O N

C C C C I C C C C C C C S C

0 0 0 0 1 1 1 1 1 1 1 1 2 2

s

t

op

cost

alignment

gap

Computing Levenshtein distance - 2

D(i,j) = score of best alignment from s1..si to t1..tj

= minD(i-1,j-1) + d(si,tj) //subst/copyD(i-1,j)+1 //insertD(i,j-1)+1 //delete

(simplify by letting d(c,d)=0 if c=d, 1 else)

also let D(i,0)=i (for i inserts) and D(0,j)=j

Computing Levenshtein distance – 4

D(i,j) = minD(i-1,j-1) + d(si,tj) //subst/copyD(i-1,j)+1 //insertD(i,j-1)+1 //delete

C O H E N

M 1 2 3 4 5

C 1 2 3 4 5

C 2 3 3 4 5

O 3 2 3 4 5

H 4 3 2 3 4

N 5 4 3 3 3

A trace indicates where the min value came from, and can be used to find edit operations and/or a best alignment (may be more than 1)

Extensions

• Add parameters for differential costs for delete, substitute, … operations– Eg “gap cost” G, substitution costs dxy(x,y)

• Allow s to match a substring of t (Smith-Waterman)

• Model cost of length-n insertion as A + Bn instead of Gn– “Affine distance”

– Need to remember if a gap is open in s, t, or neither

Forward-backward for HMMs

0.1]][1[

) is statestart Pr(]][0[

iT

ii

: stateeach for i

:1for ...Tt : stateeach for i

)(]][1[)(]][[ state

ijjtxiitj

t :1for T...t

: stateeach for i

j

t jtxjjiit state

]][1[)()(]][[

Txxx ... :string a and ,:Given 21

All paths to st=i and all emissions up to and including tAll paths after st=i and all emissions after t

EM for HMMs

: stateeach for i

]][1[)()(]][[1

)(

1

t

t

new

jtxjjiitZ

ji

)...,();...,(;... string ;:Given 111 TTT xxxxxx

i

iTZ state

]][[ where :pair stateeach for i,j

1):?0 ( ]][[]][[1

)( axititZ

ai tt

new

pass thru state i at t and emit a at t

pass thru states i,j at t,t+1 …and con’t to end

Pair HMM Example

1

e Pr(e)

<a,a> 0.10

<e,e> 0.10

<h,h> 0.10

<e,-> 0.05

<h,t> 0.05

<-,h> 0.01

... ..

Sample run: zT = <h,t>,<e,e><e,e><h,h>,<e,->,<e,e>Strings x,y produced by zT: x=heehee, y=teehe

Notice that x,y is also produced by z4 + <e,e>,<e,-> and many other edit strings

Distances based on pair HMMs

),,(:),,(:

)Pr()Pr()|,Pr(VTnnVTnn yxzEDITz i

i

yxzEDITz

nVT zzyx

)|,Pr(log),(stochastic TTTT yxyxd

)|Pr(maxarglog),(),,(:

viterbi n

yxzEDITz

TT zyxdTTnn

Pair HMM Inference

),()1,(),(),1(),()1,1(

),Pr(),(

vtvt

vt

yvtxvtyxvt

yxvt

t=1 t=2 t=2 ... t=T

v=1 ...

v=2 ...

...

v=K ...

h

e

h a h

α(3,2)

Pair HMM Inference

),()1,(),(),1(),()1,1(

),Pr(),(

vtvt

vt

yvtxvtyxvt

yxvt

t=1 t=2 t=2 ... t=T

v=1 ...

v=2 ...

...

v=K ...

h

e

h a h

α(3,2)

Pair HMM Inference: Forward-Backward

)1,(),(),1(),()1,1(),(

),Pr(),(),Pr(),(),Pr(),(

),Pr(),(

),Pr(),())','(),(Pr(

1111

''1

jiyjixjiyx

yxyyxxyxyx

yxji

yxjijirjir

jiji

Vj

Tij

Vj

Tii

Vj

Tiji

Vj

Ti

Vj

Tikk

t=1 t=2 t=2 ... t=T

v=1 ...

v=2 ...

...

v=K ...

EM to learn edit distances

• Is this really like edit distances? Not really:– Sim(x,x) ≠1– Generally sim(x,x) gets smaller with longer x– Edit distance is based on single best sequence;

Pr(x,y) is based on weighted cost of all successful edit sequences

• Will learning work?– Unlike linear models no guarantee of global

convergence: you might not find a good model even if it exists

Back to R&Y paper...

• They consider “coarse” and “detailed” models, as well as mixtures of both.

• Coarse model is like a back-off model – merge edit operations into equivalence classes (e.g. based on equivalence classes for chars).

• Test by learning distance for K-NN with an additional latent variable

K-NN with latent prototypes

test example y (a string of phonemes)

possible prototypes x (known word pronounciation )

x1 x2 x3 xm

words from dictionary

y

w1 w2 wK

learned phonetic distance

K-NN with latent prototypes

x1 x2 x3 xm

y

w1 w2 wK

learned phonetic distance

Method needs (x,y) pairs to train a distance – to handle this, an additional level of E/M is used to pick the “latent prototype” to pair with each y

Plan for this week• Edit distances

– Distance(s,t) = cost of best edit sequence that transforms st– Found via….

• Learning edit distances: Ristad and Yianolis– Probabilistic generative model: pair HMMs– Learning now requires EM

• Detour: EM for plain ‘ol HMMS

– EM for pair HMMs

• Why EM works• Discriminative learning for pair HMMs

• Mixturess: z is hidden mixture component …• HMMs: z is hidden state sequence string• Pair HMMs: z is hidden sequence of pairs (x1,y1),… given (x,y)• Latent-variable topic models (e.g., LDA): z is assignment of words to topics• ….

X = data

θ = modelz = something you can’t observeProblem:

Algorithm: Iteratively improve θ1 θ2 …

Θn=

“complete data likelihood”

EM:

Jensen’s inequality…

Jensen’s inequality

x3

and f convex

Jensen’s inequality

x3

ln

X = data

θ = model

z = something you can’t observe

Let’s think about moving from θn (our current parameter vector) to some new θ (the next one, hopefully better)

We want to optimize L(θ)- L(θn ) …. using something like…

z zX,

zX,X|z

)|(

)|(ln),(

nn P

PP

Comments

• Nice because we often know how to– Do learning in the model (if hidden variables are

known)– Do inference in the model (to get hidden

variables)– And that’s all we need to do….

• Convergence: local, not global

• Generalized EM: E but don’t M, just improve

Key ideas

• Pair of strings (x,y) associated with a label: {match,nonmatch}

• Classification done by a pair HMM with two non-initial states: {match, non-match} w/o transitions between them

• Model scores alignments – emissions sequences – as match/nonmatch.

Key ideas

Edit sequence is featurized:

Score the alignment sequence:

Marginalize over all alignments to score match v nonmatch:

Key ideas

To learn, combine EM and CRF learning:• compute expectations over (hidden) alignments• use LBFGS to maximize (or at least improve )the parameters, λ• repeat……

Initialize the model with a “reasonable” set of parameters:• hand-tuned parameters for matching strings• copy match parameters to non-match state and shrink them to zero.

Results

We will come back to this family of methods in a couple of weeks (discriminatively trained latent-variable models).