Alignment-Based Composi2onal Seman2cs for Instruc2on Following - - PowerPoint PPT Presentation

alignment based composi2onal seman2cs for instruc2on
SMART_READER_LITE
LIVE PREVIEW

Alignment-Based Composi2onal Seman2cs for Instruc2on Following - - PowerPoint PPT Presentation

Alignment-Based Composi2onal Seman2cs for Instruc2on Following Jacob Andreas and Dan Klein UC Berkeley Following instruc2ons go down the green hall go all the way then face the chair Following instruc2ons go down the green hall


slide-1
SLIDE 1

Alignment-Based Composi2onal Seman2cs for Instruc2on Following

Jacob Andreas and Dan Klein

UC Berkeley

slide-2
SLIDE 2

Following instruc2ons go down the green hall go all the way 
 then face the chair

slide-3
SLIDE 3

Following instruc2ons go down the green hall go all the way 
 then face the chair

slide-4
SLIDE 4

Following instruc2ons: parsing go down the green hall go all the way 
 then face the chair

slide-5
SLIDE 5

Following instruc2ons: parsing go down the green hall go all the way 
 then face the chair

slide-6
SLIDE 6

Following instruc2ons: planning go down the green hall go all the way 
 then face the chair

slide-7
SLIDE 7

Following instruc2ons: planning go down the green hall go all the way 
 then face the chair

slide-8
SLIDE 8

Following instruc2ons Planning Parsing

[ go down the hall, go all the way ] green hall face the chair (no instruc2on)

slide-9
SLIDE 9

Previous work Planning Parsing

Branavan et al. (2009) Vogel and Jurafsky (2010) Chen and Mooney (2011) Artzi and ZeTlemoyer (2013) Kim and Mooney (2013) Tellex et al. (variously)

slide-10
SLIDE 10

Parsing and planning Sequence-to-sequence Structure-to-structure

go down the green hall turn le2 go down the green hall

slide-11
SLIDE 11

SEQUENCE-TO-SEQUENCE MODEL

slide-12
SLIDE 12

go down the yellow hall turn le2

Sequence-to-sequence

slide-13
SLIDE 13

go down the yellow hall turn le2

Sequence-to-sequence

Alignments

slide-14
SLIDE 14

go down the yellow hall turn le2

Sequence-to-sequence

1 2 3 Ac2ons Alignments

slide-15
SLIDE 15

go down the yellow hall turn le2

Sequence-to-sequence

1 2 3 Ac2ons Alignments

slide-16
SLIDE 16

go down the yellow hall turn le2

Sequence-to-sequence

1 2 3 Ac2ons Alignments

2

slide-17
SLIDE 17

go down the yellow hall turn le2

Sequence-to-sequence

1 2 3 Ac2ons Alignments

2

slide-18
SLIDE 18

go down the yellow hall turn le2

Sequence-to-sequence

1 2 3

2

slide-19
SLIDE 19

go down the yellow hall turn le2

2

Sequence-to-sequence

1 2 3

slide-20
SLIDE 20

go down the yellow hall turn le2

2

Sequence-to-sequence

1 2 3

slide-21
SLIDE 21

go down the yellow hall turn le2

2

Sequence-to-sequence

1 2 3

slide-22
SLIDE 22

Sequence-to-sequence

go down the yellow hall turn le2

2

1 2 3

slide-23
SLIDE 23

go down the yellow hall turn le2

Sequence-to-sequence

2

?

slide-24
SLIDE 24

STRUCTURE-TO-STRUCTURE MODEL

slide-25
SLIDE 25

Structure-to-structure

go down the yellow hall

slide-26
SLIDE 26

Structure-to-structure

go down hall the yellow

slide-27
SLIDE 27

Structure-to-structure

go down hall the yellow

slide-28
SLIDE 28

Structure-to-structure

go down the yellow hall

slide-29
SLIDE 29

Structure-to-structure

go down the yellow hall

slide-30
SLIDE 30

Structure-to-structure

go down hall the yellow

slide-31
SLIDE 31

JOINT MODEL

slide-32
SLIDE 32

2

go down the yellow hall turn le2

Joint model

slide-33
SLIDE 33

go down the yellow hall turn le2

Joint model

go down the yellow hall

slide-34
SLIDE 34

go down the yellow hall turn le2

Joint model

go down hall the yellow

slide-35
SLIDE 35

go down the yellow hall turn le2

Joint model

go down the yellow hall

slide-36
SLIDE 36

Joint model

ψ !

θ>⇣

go

φ( , )

down

+ φ( , )

=

+ · · ·

go down the yellow hall

slide-37
SLIDE 37

Joint model

ψ !

exp X

p(actions, alignment | text; θ) ∝

go down the yellow hall

slide-38
SLIDE 38

go down the yellow hall turn le2

Joint model

go down the yellow hall

slide-39
SLIDE 39

LEARNING / INFERENCE

slide-40
SLIDE 40

Learning: coordinate ascent

max

θ,alignments p(actions, alignments | text; θ)

max

θ

max

alignments

using L-BFGS using custom alignment dynamic program

slide-41
SLIDE 41

Decoding: coordinate ascent

max

alignments

using a planner using custom alignment dynamic program

max

actions

max

actions, alignments

p(actions, alignments | text; θ)

slide-42
SLIDE 42

EXPERIMENTS

slide-43
SLIDE 43

Three tasks

slide-44
SLIDE 44

Puzzle solving

clear the two long columns, and then the row

[Branavan+ 09]

slide-45
SLIDE 45

Puzzle solving

clear the two long columns, and then the row

[Branavan+ 09]

slide-46
SLIDE 46

Puzzle solving

30 60 90

70 63 54

No text

Branavan et al.

2009 This work

Plan execu2on (%)

slide-47
SLIDE 47

Puzzle solving

30 60 90

86 78

No text This work

Task comple2on (%)

slide-48
SLIDE 48

Maze naviga2on

you should see a grey floor to your right and a chair in front of you

[MacMahon+ 91]

slide-49
SLIDE 49

Maze naviga2on

35 45 55 65

65 63 60 57 57

Kim + Mooney

2012

Chen + Mooney

2012 This work

Kim + Mooney

2013

Artzi et al.

2014

Task comple2on (%)

slide-50
SLIDE 50

Map reading

right round the white water [...] but stay quite close ’cause you don’t

  • therwise you’re going to

be in that stone creek

[Anderson+ 91]

slide-51
SLIDE 51

Map reading

15 30 45 60

60 45 48

Vogel+ 2010 Andreas+ 2013 This work

Plan execu2on (F0)

slide-52
SLIDE 52

Abla2ons: maze naviga2on

35 45 55 65

60 52 59

Without grounding graphs

Task comple2on (%)

This work Without planning

slide-53
SLIDE 53

Abla2ons: map reading

15 30 45 60

60 52 45

Without grounding graphs This work

Plan execu2on (F0)

Without planning

slide-54
SLIDE 54

Conclusion

  • Composi2onal seman2cs (like a parser)
  • Sequence structure (like a planner)

Structured alignment/decoding gives us best aspects of:

slide-55
SLIDE 55

THANK YOU