Chains of Reasoning over Entities, Relations, and Text...

Chains of Reasoning over Entities, Relations, and Text using

Recurrent Neural Networks

Rajarshi Das, Arvind Neelakantan, David Belanger, Andrew McCallum

Presentation in Deep Learning Summer School,

Montreal Aug 4 2016

2

Relation Extraction

January 15, 2000Tech pioneer Bill Gates stepped down today as chief executive officer of Microsoft, the Seattle-headquartered software giant. His long-time friend, Steve Balmer, will take over as CEO of Microsoft.

Text docs

2

Relation Extraction


Text docs

January 15, 2000Tech pioneer Bill Gates stepped down today as chief executive officer of Microsoft, the Seattle-headquartered software giant. His long-time friend, Steve Balmer, will take over as CEO of Microsoft. Steve Balmer

Microsoft

Bill Gates

MicrosoftSteve Balmer

3

Relation Extraction


Text docs


Steve Balmer

Microsoft

Bill Gates

Microsoft

Steve Balmer

3

Relation Extraction


Text docs


Steve Balmer

Microsoft

Bill Gates

Microsoft

Steve BalmerCEO?

3

Relation Extraction


Text docs


Steve Balmer

Microsoft

Bill Gates

Microsoft

Steve BalmerCEO?

Friends?

Co-workers?

3

Relation Extraction


Text docs


Steve Balmer

Microsoft

Bill Gates

Microsoft

Steve BalmerCEO?

Friends?

Co-workers?

: CEO ?

: worked_together ?

3

Relation Extraction


Text docs


Steve Balmer

Microsoft

Bill Gates

Microsoft

Steve BalmerCEO?

Friends?

Co-workers?

: CEO ?

: worked_together ?

long-time friendtake over as

CEO

4

Relation Extraction


Text docs


Steve Balmer

Microsoft

Bill Gates

Microsoft

Steve BalmerCEO?

Friends?

Co-workers?

: CEO ?

: worked_together ?

long-time friend

take over as

CEO

Open IE textual relation types (Banko et al, 2007)

5

Universal Schema for Relation Extraction


Text docs


Steve Balmer

Microsoft

Bill Gates

Microsoft

Steve BalmerCEO?

Friends?

Co-workers?

: CEO ?

: worked_together ?

long-time friend

take over as

CEO


Riedel, Yao, McCallum, Marlin, 2013

Union of all relation types from all structured sources and text

5



Text docs


Steve Balmer

Microsoft

Bill Gates

Microsoft

Steve BalmerCEO?

Friends?

Co-workers?

: CEO ?

: worked_together ?

long-time friend

take over as

CEO



5



Text docs


Steve Balmer

Microsoft

Bill Gates

Microsoft

Steve BalmerCEO?

Friends?

Co-workers?

: CEO ?

: worked_together ?

long-time friend

take over as

CEO



Predict relation between entities based on mentions in text

Chains of Reasoning for Relation Extraction

6

Microsoft Seattle

isBasedIn??


headquartered


7

Microsoft Seattle

isBasedIn??


headquarteredBill

MicrosoftSeattle


7

Microsoft Seattle

isBasedIn??


headquartered

Gates will now focus on the charitable foundation he runs with his wife Melinda French Gates. Bill and Melinda were married in a ceremony in Hawaii, rather than her hometown of Dallas.

Bill

MicrosoftSeattle


7

Microsoft Seattle

isBasedIn??


headquartered

Gates will now focus on the charitable foundation he runs with his wife Melinda French Gates. Bill and Melinda were married in a ceremony in Hawaii, rather than her hometown of Dallas.

Bill

MicrosoftSeattle

Melinda

8

Microsoft Seattle

isBasedIn??

January 15, 2000Tech pioneer Bill Gates stepped down today as chief executive officer of Microsoft, the Seattle-headquartered software giant. His long-time friend, Steve Balmer, will take over as CEO of Microsoft

headquartered

Gates will now focus on the charitable foundation he runs with his wife Melinda French Gates. He will continue to serve as the chairman of Microsoft. Bill and Melinda were married in a ceremony in Hawaii, rather than her

Seattle

Melinda Bill Microsoft

headquartered

chairmanspouse


8

Microsoft Seattle

isBasedIn??


headquartered


Seattle


headquartered

chairmanspouse

lives in??


8

Microsoft Seattle

isBasedIn??


headquartered


Seattle


headquartered

chairmanspouse

lives in??

A Path between ‘Melinda’ and ‘Seattle’


8

Microsoft Seattle

isBasedIn??


headquartered


Seattle


headquartered

chairmanspouse

lives in??


Multi-hopInference


8

Microsoft Seattle

isBasedIn??


headquartered


Seattle


headquartered

chairmanspouse

lives in??


Multi-hopInference

Single-hopInference


8

Microsoft Seattle

isBasedIn??


headquartered


Seattle


headquartered

chairmanspouse

lives in??


Multi-hopInference

Single-hopInference

Spouse(A,B) & Chairman(B,C) & HQ-in(C,D) ➞ Lives-in(A,D)


8

Microsoft Seattle

isBasedIn??


headquartered


Seattle


headquartered

chairmanspouse

lives in??


Multi-hopInference

Single-hopInference


• Sherlock (Schoenmackers et al. 2010) • Path Ranking Algorithm(Lao et al, 2011, 2012),


8

Microsoft Seattle

isBasedIn??


headquartered


Seattle


headquartered

chairmanspouse

lives in??


Multi-hopInference

Single-hopInference



Symbolic; Doesn't generalize


8

Microsoft Seattle

isBasedIn??


headquartered


Seattle


headquartered

chairmanspouse

lives in??


Multi-hopInference

Single-hopInference



Spouse(A,B) & CEO(B,C) & HQ-in(C,D) ➞ Lives-in(A,D)Symbolic;

Doesn't generalize


8

Microsoft Seattle

isBasedIn??


headquartered


Seattle


headquartered

chairmanspouse

lives in??


Multi-hopInference

Single-hopInference



Spouse(A,B) & CEO(B,C) & HQ-in(C,D) ➞ Lives-in(A,D)Spouse(A,B) & COO(B,C) & HQ-in(C,D) ➞ Lives-in(A,D)

Symbolic; Doesn't generalize


9

SeattleMelinda Bill Microsoftheadquarteredchairmanspouse

RNNs for Chains of Reasoning

9


RNN


9


RNN RNN


9


RNN RNN RNN


9


RNN RNN RNN

Vector representation of the path


9


RNN RNN RNN


9


RNN RNN RNN

similarity metriclives in??


9


RNN RNN RNN


0.94


9


RNN RNN RNN


0.94

RNNs for Chains of ReasoningLogical Inference

inVector space!

9


RNN RNN RNN


0.94

Neelakantan et al’ 15

RNNs for Chains of ReasoningLogical Inference

inVector space!

10

Contributions

10

Contributions

1. Single RNN for chains of reasoning.

10

Contributions


2. Combine evidence from multiple paths between entity pairs.

10

Contributions



3. Entity Aware RNN for chains of reasoning

10

Contributions




%MAP 64.43% 73.26%

10

Contributions




%MAP 64.43% 73.26%13.7%

11


11


lives in??

11


lives in?? works in??

11


lives in?? works in?? shops in??

12

/ /lives in?? works in?? shops in?? /….


12


12



RNN RNN RNN

similarity metric

0.94

12



RNN RNN RNN

similarity metric

0.94

Shared representations for relations

{

Single Composition Matrix for all query

relations

12



RNN RNN RNN

similarity metric

0.94

Shared representations for relations

{

Single Composition Matrix for all query

relations

{All query relation vectors

lives in??

shops in??

works in??

Results

13

Results

Model %MAP

PRA 64.43

PRA + Bigram 64.93

13

Results

Model %MAP

PRA 64.43

PRA + Bigram 64.93

RNN-Path(Neelakantan’ 15) 65.23

13

Results

Model %MAP

PRA 64.43

PRA + Bigram 64.93


Single Model 68.77

13

Results

Model %MAP

PRA 64.43

PRA + Bigram 64.93

5%

p < 5e-6


Single Model 68.77

13

14

Contributions




Multiple Paths


15

Multiple Paths


Gates Foundation

CEOlocated in

15

Multiple Paths


Gates Foundation

CEOlocated in

Jane Doe John Doe

friendmarried to

born in

15

Multiple Paths


Gates Foundation

CEOlocated in

Jane Doe John Doe

friendmarried to

born in

16

Multiple Paths


Gates Foundation

CEOlocated in

Jane Doe John Doe

friendmarried to

born in

RNNencoder

16

Multiple Paths


Gates Foundation

CEOlocated in

Jane Doe John Doe

friendmarried to

born in

RNNencoder

lives in??

16

Multiple Paths


Gates Foundation

CEOlocated in

Jane Doe John Doe

friendmarried to

born in

RNNencoder

similarity metric

0.96 0.94 0.18

lives in??

17

Score Pooling

0.96 0.94 0.18

lives in??

similarity metric

Path 1 Path 2 Path 3

Score Pooling

0.96 0.94 0.18

lives in??

similarity metric


1. Max pool: max(↵1,↵2, . . . ,↵N )


Score Pooling

0.96 0.94 0.18

lives in??

similarity metric


1. Max pool: max(↵1,↵2, . . . ,↵N )


Max

Score Pooling

0.96 0.94 0.18

lives in??

similarity metric


1. Max pool: max(↵1,↵2, . . . ,↵N )


Max

0.96

Score Pooling

0.96 0.94 0.18

lives in??

similarity metric


1. Max pool: max(↵1,↵2, . . . ,↵N )


Max

Score Pooling

0.96 0.94 0.18

lives in??

similarity metric


1. Max pool: max(↵1,↵2, . . . ,↵N )


Max

2. Avg. pool: 1

N

NX

i=1

↵i

!

Avg.

Score Pooling

0.96 0.94 0.18

lives in??

similarity metric


1. Max pool: max(↵1,↵2, . . . ,↵N )


Max

2. Avg. pool: 1

N

NX

i=1

↵i

!

Avg.

0.69

Score Pooling

0.96 0.94 0.18

lives in??

similarity metric


1. Max pool: max(↵1,↵2, . . . ,↵N )


Max

2. Avg. pool: 1

N

NX

i=1

↵i

!

Avg.

Score Pooling

0.96 0.94 0.18

lives in??

similarity metric


1. Max pool: max(↵1,↵2, . . . ,↵N )


Max

2. Avg. pool: 1

N

NX

i=1

↵i

!

Avg.

3. Top(k): 1

k

kX

i=1

si

!

Top(2)

Score Pooling

0.96 0.94 0.18

lives in??

similarity metric


1. Max pool: max(↵1,↵2, . . . ,↵N )


Max

2. Avg. pool: 1

N

NX

i=1

↵i

!

Avg.

3. Top(k): 1

k

kX

i=1

si

!

Top(2)

0.95

Score Pooling

0.96 0.94 0.18

lives in??

similarity metric


4. LogSumExp: log(

NX

i=1

exp(↵i))

Score Pooling

0.96 0.94 0.18

lives in??

similarity metric


4. LogSumExp: log(

NX

i=1

exp(↵i))

LSE

Score Pooling

0.96 0.94 0.18

lives in??

similarity metric


4. LogSumExp: log(

NX

i=1

exp(↵i))

LSE

1.85

Score Pooling

0.96 0.94 0.18

lives in??

similarity metric


4. LogSumExp: log(

NX

i=1

exp(↵i))

LSE

1.85

@LSE

@↵i=

exp(↵i)Pi exp(↵i)

Score Pooling

0.96 0.94 0.18

lives in??

similarity metric


4. LogSumExp: log(

NX

i=1

exp(↵i))

LSE

1.85

@LSE

@↵i=

exp(↵i)Pi exp(↵i)

Gradient Steps #1040 0.5 1 1.5 2

Trai

ning

Los

s

0.4

0.5

0.6

0.7

0.8

0.9

1

1.1

1.2LogSumExpMaxMeanTop-k

Results

20

Results

Model %MAP Pooling

RNN-Path(Neelakantan ’15) 65.23 Max

20

Results

Model %MAP Pooling


RNN-Path 68.43 LogSumExp

20

Results

Model %MAP Pooling



Single Model 68.77 Max

Single Model 55.8 Avg.

Single Model 68.20 Top(K)

20

Results

Model %MAP Pooling




Single Model 55.8 Avg.

Single Model 68.20 Top(K)

Single Model 70.11 LogSumExp20

21

Contributions




JFK NYC NY

located in located in

airport_serves_location??

JFK NYC NY



0.9276~

JFK NYC NY



0.9276~

Yankee Stadium NYC NY



JFK NYC NY



0.9276~

Yankee Stadium NYC NY



~ 0.9276

SeattleMelinda Bill Microsoft headquarteredchairmanspouse

Entity Representation

SeattleMelinda Bill Microsoft



Entity Representation1. Learn separate representation for each entity.



Jane Doe



Jane Doe

rare occurrence; hard to learn good representations.



2. Represent entities by their annotated types 1. Learn separate representation for each entity.

Jane Doe




2. Represent entities by their annotated types 1. Learn separate representation for each entity.

Jane Doe

Annotated types in KB

1. CEO2. Philanthropist3. Duke University alumni4. American Citizen

1. American Citizen2. Small business owner

Jane DoeMelinda


Melinda


Jane Doe

1. CEO 2. Philanthropist 3. Duke University alumni 4. American Citizen

1. American Citizen 2. Small business owner

+ + +

+

=

=

Melinda


Jane Doe

1. CEO 2. Philanthropist 3. Duke University alumni 4. American Citizen

1. American Citizen 2. Small business owner

+ + +

+

=

=

Entity is represented as the sum of their annotated types

Entity Aware RNNs for Chains of Reasoning

RNN RNN RNN RNN

similarity metriccountryofHQ??

0.94

SeattleMelinda Bill Microsoftchairmanspouse headquartered dummy_rel

Results

Results

Model %MAP

Single Model 70.11

Results

Model %MAP

Single Model 70.11

Single Model + Entity 71.74

Results

Model %MAP

Single Model 70.11


Single Model + Types 73.26

Results

4.4%

p < 5e-6

Model %MAP

Single Model 70.11


Single Model + Types 73.26

Results

Results

Model %MAP Pooling

PRA 64.43 n/a

RNN-Path(Neelkantan et al’15) 65.23 Max

Results

Model %MAP Pooling

PRA 64.43 n/a



Results

Model %MAP Pooling

PRA 64.43 n/a



Single Model 70.11 LogSumExp

Results

Model %MAP Pooling

PRA 64.43 n/a




Single Model + Types 73.26 LogSumExp

Results

Model %MAP Pooling

PRA 64.43 n/a


13.7% Single Model 68.77 Max


Single Model + Types 73.26 LogSumExp

Predictive Clauses

29

Predictive Clauses

/people/person/place_of_birth(A,B) ’was born in’(A,X) & ‘commonly known as’(X,B)

29

Predictive Clauses

/people/person/place_of_birth(A,B) ’was born in’(A,X) & ‘commonly known as’(X,B)Textual relations

29

Predictive Clauses


Universal Schema

29

Predictive Clauses


29

Predictive Clauses


/book/written_work/original_language(A,B) /book/written_work_author(A,X) & ‘address’(X,Y)

& /people/person/nationality-1(Y,Z) & people/person/languages(Z,B)

29

Predictive Clauses




Entity Aware Model

29

Predictive Clauses




Sandy_Lake_Airport Sandy_Lake_First_Nation

/aviation/airport/serves?? Entity Aware Model

29

Predictive Clauses




Sandy_Lake_Airport/location/contains

Ontario‘(in northwestern)’-1

Sandy_Lake_First_Nation



29

Predictive Clauses






29

Predictive Clauses




Sandy_Lake_Airport/location/contains

Ontario‘(and at the)’

Toronto_Royal_Music_Hall

’including the’

Canada

/geography/lake/basin_countries

Big Trout Lakeand toSandy_La

ke_First_Nation



29

Conclusion

30

Conclusion

30

• Introduced a high-capacity single RNN model for relation extraction

Conclusion

30


• Combines evidences among multiple paths and sources of evidence

Conclusion

30



• Entity Aware!

Conclusion

30



• Entity Aware!

Thanks!

Chains of Reasoning over Entities, Relations, and Text...

Documents

Transcript of Chains of Reasoning over Entities, Relations, and Text...