voto.it J BO Ad bunnies START END monsters eat tasty t2 - - PDF document

SLIDE 1

Noun

Ooo Ooo

voto.it

Eo EEEEE J

Ad

BO

START

monsters

eat

tasty

bunnies

END

t2

Labels

States Observation I

Can

• nly

be

in

• ne

state

at

time

I

Observation 2

By

construction

we

have

assumed

That

We

• nly

care

about features

relating

adjacent

words

and

labels

Observation 3

Any path

Through

This

trellis

corresponds

to

a

unique

labeling

• f

X

Consider

E

the

edge from

Verb

Adj

b w

eat

and

Tasty

weight

E

w

tasty adj verb adj

SLIDE 2

How

does

this

allow

us to

compute

argmax

Naively

if

we

just

considered all unique

Paths

it

wouldn't

But

this

structure

permits

Dynamic

Programming

for

efficient

Computation

Viterbi

Algorithm

Define

Qe k

best

possible

• utput

up

to

and

including

l for

label

K

Max

w

e

X ay

K

Ot denotes

concatenation

in L 1

Consider

l

2 Assume

we

have

computed

Qs

up

to

this point

The

word

at

1

2

is

eat

Now

we want

to

derive

03 Adj

How

To

Calculate

this

score

Max

• ver

possible

previous

States

03 Adj

Max

Q2 Nan

t

WTasty Adj whom Adj

Q2verb

t

WTasty Adj

t Verb

Adj

Q2 Adj

t

WTasty Adj

t

WAdj

Adj

SLIDE 3

We

can

generalize

this

for

a

recursive

definition

CT

49

denote

the

label

at

position

1 1

that

achenes

The

Max

Init

• K

the

4

Tk

feature

vector for position

Best score

l

given

that

we are

at

coming from

Now

k

f

Qe

mna.xEQE.k.tw

e zCxlk k

K and

H K

going

to

k

Where

to

Transition

from

4

1 k

Same

but Erymax

At

The

end

can

find

best

Seg

argmax

K K

last

State

K

and

trace

backwards

SLIDE 4

OO

O u

• O

2

z o

features

• 1

I

• Pixel

Valve face

terms

1 1

11

O

O

• I

20

• Others

Mayber

Coordinate

into

• Ii

Marty features

Transition

to

face

noT face

from

neighbors

Yci 2 j Il

Yei Dj YiLj 1

Dynamic

Program

Start

from

• proceed