An AMR Aligner Tuned by Transition-based Parseryjliu.net/cv/res/2018-09-07-amr.compressed.pdf · 2020-02-18 · An AMR Aligner Tuned by Transition-based Parser YijiaLiu*,WanxiangChe,BoZheng,BingQin,TingLiu

An AMR Aligner Tuned by Transition-based Parser

Yijia Liu*, Wanxiang Che, Bo Zheng, Bing Qin, Ting Liu

Research Center for Social Computing and Information Retrieval

Harbin Institute of Technology

Abstract Meaning Representation(AMR) Parsing

North Korean froze its nuclear actions in

exchange for two nuclear reactors

exchange-01

country

ARG0 freeze-01

ARG1

recieve-01

ARG3

name

name

ARG0

act-02

ARG1

reactor

ARG1

"North"

op1

"Korea"

op2

nucleus~1

mod

2

quant

nucleus~2

mod

AMR Parsing: Concept Identification



exchange-01

country

ARG0 freeze-01

ARG1

recieve-01

ARG3

name

name

ARG0

act-02

ARG1

reactor

ARG1

"North"

op1

"Korea"

op2

nucleus~1

mod

2

quant

nucleus~2

mod

AMR Parsing: Relation Identificationexchange-01

country

ARG0 freeze-01

ARG1

recieve-01

ARG3

name

name

ARG0

act-02

ARG1

reactor

ARG1

"North"

op1

"Korea"

op2

nucleus~1

mod

2

quant

nucleus~2

mod

exchange-01

country

ARG0 freeze-01

ARG1

recieve-01

ARG3

name

name

ARG0

act-02

ARG1

reactor

ARG1

"North"

op1

"Korea"

op2

nucleus~1

mod

2

quant

nucleus~2

mod

AMR Parsing: Alignment Challenge



exchange-01

country

ARG0 freeze-01

ARG1

recieve-01

ARG3

name

name

ARG0

act-02

ARG1

reactor

ARG1

"North"

op1

"Korea"

op2

nucleus~1

mod

2

quant

nucleus~2

mod

Association between a concept and

a span of words is abstracted away

Aligner for Training an AMR Parser:Guessing Alignment



exchange-01

country

ARG0 freeze-01

ARG1

recieve-01

ARG3

name

name

ARG0

act-02

ARG1

reactor

ARG1

"North"

op1

"Korea"

op2

nucleus~1

mod

2

quant

nucleus~2

mod

its nuclear actions

e

nucleus~1

mod

country

name

name

act-02

"North"

op1

"Korea"

op2

nucleus

m

North Korean froze its

two nuclear reactors

quant

nucleus~2

mod

Aligner

JAMR Aligner (Flanigan et al. 2014)

A Set of Rules Greedy Decoding

• Manually define rule order

• Apply these rules one by one

• Alignments are mutuallyexcluded, a span once alignedwill never be aligned again.

Problems of JAMR

its nuclear actions

e

nucleus~1

mod

two nuclear reactors

quant

nucleus~2

mod

Limited semantic

resources

Ambiguities in

matching results

Parser training does

not feed back to

alignment

its nuclear actions

Our Contributions

A novel aligner A transition-based parser

Our Contributions


Our Aligner

Limited semantic

resources

Recall more alignments with

• Glove embedding (Koreanand Korea)

• morphological links(example and exemplify)

Ambiguities in

matching results


not feed back to

alignment

Our Aligner

Limited semantic

resources




Ambiguities in

matching results


not feed back to

alignment

Produce multiple alignments

Our Aligner

Limited semantic

resources:




Ambiguities in

matching results


not feed back to

alignment

Use an oracle transition-based parse to determine the quality

of an alignment

oracle transition-based parser

Produce multiple alignments

Oracle Transition-based Parser

Oracle Dependency

Parser

Input sentence

Gold dep. tree

Gold transition

sequence

Our Oracle Parser

Input sentence

Alignment

Gold AMR graph

Best transition

sequenceBest AMR

Oracle Transition-based Parser

• State: ! = ($ s&, (, b& b*|,, -)

• Actions:• extension of the list-baseddependency parser

• CONFIRM, ENTITY, and NEW forconcept identification• Oracle parser use the alignmentto derive these actions

buffer:

North Korean froze its nuclear actions in exchange for two nuclear reactors

stack: deque:

ENTITY(country)

SHIFT

A Run on a Successful Alignment

Input sentence: North Korean froze its nuclear actions in exchange for two nuclear reactors

Alignment: North Korean – country entity: ”North Korean”; froze – freeze; nuclear1 – nucleus~1; actions –

act; exchange – exchange; for – receive; two – 2; nuclear2 – nucleus~2; reactors – reactor

buffer:

froze its nuclear actions in exchange for two nuclear reactors

stack:

country

name

name

act-02

"North"

op1

"Korea"

op2

nucleus

m

deque:


SHIFT

CONCEPT(freeze)

LEFT(ARG0)




buffer:

its nuclear actions in exchange for two nuclear reactors

stack:

country

name

name

act-02

"North"

op1

"Korea"

op2

nucleus

m

deque:


DROP




buffer:

nuclear actions in exchange for two nuclear reactors

stack:

country

name

name

act-02

"North"

op1

"Korea"

op2

nucleus

m

deque:


SHIFT

CONCEPT(nucleus~1)




buffer:

actions in exchange for two nuclear reactors

stack:

country

name

name

act-02

"North"

op1

"Korea"

op2

nucleus

m

deque:


CONCEPT(act)

LEFT(mod)




buffer:

in exchange for two nuclear reactors

stack:

country

name

name

act-02

"North"

op1

"Korea"

op2

nucleus

m

deque:


REDUCE




buffer:


stack:

country

name

name

act-02

"North"

op1

"Korea"

op2

nucleus

m

deque:





RIGHT(ARG1)

buffer:


stack:

country

name

name

act-02

"North"

op1

"Korea"

op2

nucleus

m

deque:


…

Ends up with 32 actions




buffer:


stack:

country

name

name

act-02

"North"

op1

"Korea"

op2

nucleus

m

deque:

A Run on an Unsuccessful Alignment

SHIFT

CONCEPT(nucleus~2)




buffer:


stack:

country

name

name

act-02

"North"

op1

"Korea"

op2

nucleus

m

deque:


CACHE




buffer:


stack:

country

name

name

act-02

"North"

op1

"Korea"

op2

nucleus

m

deque:





CONCEPT(act)

RIGHT(ARG1)

buffer:


stack:

country

name

name

act-02

"North"

op1

"Korea"

op2

nucleus

m

deque:





SHIFT

buffer:


stack:

country

name

name

act-02

"North"

op1

"Korea"

op2

nucleus

m

deque:





Keep CACHEing

Ends up with 44 actions

buffer:


stack:

country

name

name

act-02

"North"

op1

"Korea"

op2

nucleus

m

deque:

A Run on another Unsuccessful Alignment

DROP


Alignment: North Korean – country entity: ”North Korean”; froze – freeze; actions – act; exchange –

exchange; for – receive; two – 2; nuclear2 – nucleus~2; reactors – reactor

buffer:


stack:

country

name

name

act-02

"North"

op1

"Korea"

op2

nucleus

m

deque:

A Run on another Unsuccessful Alignment


Alignment: North Korean – country entity: ”North Korean”; froze – freeze; actions – act; exchange –

exchange; for – receive; two – 2; nuclear2 – nucleus~2; reactors – reactor

Ends up with a graph with Smatch of

0.91

Aligner Experiments (intrinsic):Alignment Evaluation

90.6

95.2

85

87

89

91

93

95

97

99

Alignment F1

JAMR Ours

Aligner Experiments (extrinsic):Two Open-sourced AMR Parsers

65.9

67.6

63

64

65

66

67

68

69

70

JAMR (Flanigan14)

Baseline +our alignment

64.6

65.1

63

64

65

66

67

68

69

70

CAMR (Wang15)

Baseline +our alignment

1.7

0.5

On LDC2014T12

Aligner Experiments: Ablations

73.172.7

67.6

71.3

65

66

67

68

69

70

71

72

73

74

75

Smatch

JAMR parser

+ours -semantic matching -tuning +JAMR aligner

Our Contributions


Transition-based AMR Parseron Our Transition System

• Use StackLSTM [Dyer et al. 2015] to represent state

• Call a unique subroutine for span-to-concept classification

Parser Experiments

60

61

62

63

64

65

66

67

68

69

70

BA17 Ours Zhou16 Pust15 OursX3 Wang17 OursX10

Parser Experiments

60

61

62

63

64

65

66

67

68

69

70

BA17 Ours Zhou16 Pust15 OursX3 Wang17 OursX10

Comparison to Ballesteros andAl-Onaizan (2017)

60

61

62

63

64

65

66

67

BA17+JAMR alignment

Ours+JAMR alignment

Ours+Our Alignment

●

●

●

●

●

●

●

●●

●●

●

●

●●●

●

●

●●

●

●●

●●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

● ●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●●

● ●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●●

●

●

●

●

●

●

●

●●

●

●

●●

●

●

● ●

●

●

●

●

●

●

● ●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

● ●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●●

●

●

●

●●

●

●

●●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●●

●●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

● ●

●

●

●

●●

● ●

●

●

● ●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

● ●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

● ●

●

●

●

●●

●

●

●

●

●

●●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

● ●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

● ●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

● ●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

● ●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●●

●●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

● ●

●●

●

●

●

●

●

●

●

●

●●

●●

●

●

● ●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

● ●

●

●

●

●

●

●●

●

●

●

●

●

● ●

●

●

●●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

● ●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

0

50

100

150

200

250

0 10 20 30 40 50

# of tokens

# o

f a

ctio

ns

● BA17

Ours

Parser Experiments:Speed Comparison

Model Tokens/s

JAMR 7

CAMR 24

Wang17 (Extension of CAMR) <24

Our (X10) 43

Conclusion

• A new AMR aligner that improves two open-sourced parsers

• A transition-based parser that is accurate and fast

• Code and generated alignments (LDC2014T12 andLDC2017T10) are available at:https://github.com/Oneplus/tamr

https://github.com/Oneplus/tamr

Thanks and Q/A

Documents

An AMR Aligner Tuned by Transition-based Parseryjliu.net/cv/res/2018-09-07-amr.compressed.pdf · 2020-02-18 · An AMR Aligner Tuned by Transition-based Parser YijiaLiu*,WanxiangChe,BoZheng,BingQin,TingLiu