Understanding Dataset Design Choices for Mul5-hop Reasoning Jifan - - PowerPoint PPT Presentation
Understanding Dataset Design Choices for Mul5-hop Reasoning Jifan - - PowerPoint PPT Presentation
Understanding Dataset Design Choices for Mul5-hop Reasoning Jifan Chen and Greg Durre; The University of Texas at Aus5n Mul5-hop Reasoning Ques%on : The Oberoi family is part of a hotel company that has a head office in what city? The Oberoi
Mul5-hop Reasoning
The Oberoi family is an Indian family that is famous for its involvement in hotels, namely through The Oberoi Group Ques%on: The Oberoi family is part of a hotel company that has a head office in what city? The Oberoi Group is a hotel company with its head office in Delhi.
…
…
Doc 1 Doc 2
Example picked from HotpotQA (Yang 2018)
Mul5-hop Reasoning
The Oberoi family is an Indian family that is famous for its involvement in hotels, namely through The Oberoi Group Ques%on: The Oberoi family is part of a hotel company that has a head office in what city? The Oberoi Group is a hotel company with its head office in Delhi.
…
…
Doc 1 Doc 2
Example picked from HotpotQA (Yang 2018)
Mul5-hop Reasoning
The Oberoi family is an Indian family that is famous for its involvement in hotels, namely through The Oberoi Group Ques%on: The Oberoi family is part of a hotel company that has a head office in what city? The Oberoi Group is a hotel company with its head office in Delhi.
…
…
Doc 1 Doc 2 Same en5ty
Example picked from HotpotQA (Yang 2018)
Mul5-hop Reasoning
The Oberoi family is an Indian family that is famous for its involvement in hotels, namely through The Oberoi Group Ques%on: The Oberoi family is part of a hotel company that has a head office in what city? The Oberoi Group is a hotel company with its head office in Delhi.
…
…
Doc 1 Doc 2 Same en5ty
Example picked from HotpotQA (Yang 2018)
Mul5-hop Reasoning
The Oberoi family is an Indian family that is famous for its involvement in hotels, namely through The Oberoi Group Ques%on: The Oberoi family is part of a hotel company that has a head office in what city? The Oberoi Group is a hotel company with its head office in Delhi.
…
…
Doc 1 Doc 2 Same en5ty Same en5ty
Example picked from HotpotQA (Yang 2018)
Mul5-hop Reasoning
The Oberoi family is an Indian family that is famous for its involvement in hotels, namely through The Oberoi Group Ques%on: The Oberoi family is part of a hotel company that has a head office in what city? The Oberoi Group is a hotel company with its head office in Delhi.
…
…
Doc 1 Doc 2 Same en5ty Same en5ty
Example picked from HotpotQA (Yang 2018)
Mul5-hop Reasoning
The Oberoi family is an Indian family that is famous for its involvement in hotels, namely through The Oberoi Group Ques%on: The Oberoi family is part of a hotel company that has a head office in what city? The Oberoi Group is a hotel company with its head office in Delhi.
…
…
Doc 1 Doc 2 Same en5ty Same en5ty
Example picked from HotpotQA (Yang 2018)
This is an idealized version of mul5-hop reasoning. Do models need to do this to do well on this task?
Mul5-hop Reasoning
The Oberoi family is an Indian family that is famous for its involvement in hotels, namely through The Oberoi Group Ques%on: The Oberoi family is part of a hotel company that has a head office in what city? The Oberoi Group is a hotel company with its head office in Delhi.
…
…
Doc 1 Doc 2
Example picked from HotpotQA (Yang 2018)
Mul5-hop Reasoning
The Oberoi family is an Indian family that is famous for its involvement in hotels, namely through The Oberoi Group Ques%on: The Oberoi family is part of a hotel company that has a head office in what city? The Oberoi Group is a hotel company with its head office in Delhi.
…
…
Doc 1 Doc 2
Example picked from HotpotQA (Yang 2018)
Ignore the bridging en5ty and directly jump to the answer!
High lexical overlap
Mul5-hop Reasoning
Ques%on: What government posi5on was held by the woman who portrayed Corliss Archer in the film Kiss and Tell ?
Example picked from HotpotQA (Yang 2018)
Mul5-hop Reasoning
Meet Corliss Archer is an American television sitcom that aired on CBS … Ques%on: What government posi5on was held by the woman who portrayed Corliss Archer in the film Kiss and Tell ? Shirley Temple Black was an American actress, businesswoman, and singer … Kiss and Tell is a comedy film in which 17-year-old Shirley Temple acts as Corliss Archer .
… …
Example picked from HotpotQA (Yang 2018)
As an adult, she served as Chief of Protocol of the United States
Doc 1 Doc 2 Doc 3
Mul5-hop Reasoning
Meet Corliss Archer is an American television sitcom that aired on CBS … Ques%on: What government posi5on was held by the woman who portrayed Corliss Archer in the film Kiss and Tell ? Shirley Temple Black was an American actress, businesswoman, and singer … Kiss and Tell is a comedy film in which 17-year-old Shirley Temple acts as Corliss Archer .
… …
Example picked from HotpotQA (Yang 2018)
As an adult, she served as Chief of Protocol of the United States
Doc 1 Doc 2 Doc 3
Mul5-hop Reasoning
Meet Corliss Archer is an American television sitcom that aired on CBS … Ques%on: What government posi5on was held by the woman who portrayed Corliss Archer in the film Kiss and Tell ? Shirley Temple Black was an American actress, businesswoman, and singer … Kiss and Tell is a comedy film in which 17-year-old Shirley Temple acts as Corliss Archer .
… …
Example picked from HotpotQA (Yang 2018)
As an adult, she served as Chief of Protocol of the United States
Doc 1 Doc 2
Same en5ty
Doc 3
Mul5-hop Reasoning
Meet Corliss Archer is an American television sitcom that aired on CBS … Ques%on: What government posi5on was held by the woman who portrayed Corliss Archer in the film Kiss and Tell ? Shirley Temple Black was an American actress, businesswoman, and singer … Kiss and Tell is a comedy film in which 17-year-old Shirley Temple acts as Corliss Archer .
… …
Example picked from HotpotQA (Yang 2018)
As an adult, she served as Chief of Protocol of the United States
Doc 1 Doc 2
Same en5ty
Doc 3
Mul5-hop Reasoning
Meet Corliss Archer is an American television sitcom that aired on CBS … Ques%on: What government posi5on was held by the woman who portrayed Corliss Archer in the film Kiss and Tell ? Shirley Temple Black was an American actress, businesswoman, and singer … Kiss and Tell is a comedy film in which 17-year-old Shirley Temple acts as Corliss Archer .
… …
Example picked from HotpotQA (Yang 2018)
As an adult, she served as Chief of Protocol of the United States
Doc 1 Doc 2
Same en5ty
Doc 3
Same en5ty
Mul5-hop Reasoning
Meet Corliss Archer is an American television sitcom that aired on CBS … Ques%on: What government posi5on was held by the woman who portrayed Corliss Archer in the film Kiss and Tell ? Shirley Temple Black was an American actress, businesswoman, and singer … Kiss and Tell is a comedy film in which 17-year-old Shirley Temple acts as Corliss Archer .
… …
Example picked from HotpotQA (Yang 2018)
As an adult, she served as Chief of Protocol of the United States
Doc 1 Doc 2
Same en5ty
Doc 3
Same en5ty
Mul5-hop Reasoning
Meet Corliss Archer is an American television sitcom that aired on CBS … Ques%on: What government posi5on was held by the woman who portrayed Corliss Archer in the film Kiss and Tell ? Shirley Temple Black was an American actress, businesswoman, and singer … Kiss and Tell is a comedy film in which 17-year-old Shirley Temple acts as Corliss Archer .
… …
Example picked from HotpotQA (Yang 2018)
As an adult, she served as Chief of Protocol of the United States
Doc 1 Doc 2
Same en5ty
Doc 3
Same en5ty
No simple lexical overlap.
Mul5-hop Reasoning
Meet Corliss Archer is an American television sitcom that aired on CBS …
Ques%on: What government posi5on was held by the woman who portrayed Corliss Archer in the film Kiss and Tell ?
Shirley Temple Black was an American actress, businesswoman, and singer … Kiss and Tell is a comedy film in which 17-year-old Shirley Temple acts as Corliss Archer .
… …
As an adult, she served as Chief of Protocol of the United States
Doc 1 Doc 2 Doc 3
Example picked from HotpotQA (Yang 2018)
Mul5-hop Reasoning
Meet Corliss Archer is an American television sitcom that aired on CBS …
Ques%on: What government posi5on was held by the woman who portrayed Corliss Archer in the film Kiss and Tell ?
Shirley Temple Black was an American actress, businesswoman, and singer … Kiss and Tell is a comedy film in which 17-year-old Shirley Temple acts as Corliss Archer .
… …
As an adult, she served as Chief of Protocol of the United States
Doc 1 Doc 2 Doc 3
Example picked from HotpotQA (Yang 2018)
Mul5-hop Reasoning
Meet Corliss Archer is an American television sitcom that aired on CBS …
Ques%on: What government posi5on was held by the woman who portrayed Corliss Archer in the film Kiss and Tell ?
Shirley Temple Black was an American actress, businesswoman, and singer … Kiss and Tell is a comedy film in which 17-year-old Shirley Temple acts as Corliss Archer .
… …
As an adult, she served as Chief of Protocol of the United States
Doc 1 Doc 2 Doc 3
Only one government posi5on appears in the context!
Example picked from HotpotQA (Yang 2018)
WikiHop (Welbl et al., 2018) HotpotQA (Yang et al., 2018)
Mul5-hop Reasoning Datasets
How much mul5-hop reasoning is really needed in the recently proposed mul5-hop oriented datasets?
WikiHop (Welbl et al., 2018) HotpotQA (Yang et al., 2018)
Mul5-hop Reasoning Datasets
How much mul5-hop reasoning is really needed in the recently proposed mul5-hop oriented datasets?
- Mul5ple-choice
- Span-based
WikiHop (Welbl et al., 2018) HotpotQA (Yang et al., 2018)
Mul5-hop Reasoning Datasets
How much mul5-hop reasoning is really needed in the recently proposed mul5-hop oriented datasets?
- Mul5ple-choice
- Final answer is given as supervision
- Span-based
- Answer and suppor5ng facts
Research Ques5ons
Research Ques5ons
Can a model iden5fy the answer with only a set of candidates?
Government posi5on Chief of Protocol, actress, singer
Research Ques5ons
Can a model iden5fy the answer with only a set of candidates? Can a model iden5fy where the answer is in a single hop?
Government posi5on Chief of Protocol, actress, singer Oberoi Family Delhi
Research Ques5ons
Can a model iden5fy the answer with only a set of candidates? Can a model iden5fy where the answer is in a single hop? Does the type of supervision ma;er?
Government posi5on Chief of Protocol, actress, singer Span-based vs. Mul5ple-choice Oberoi Family Delhi
Research Ques5ons
Can a model iden5fy the answer with only a set of candidates? Can a model iden5fy where the answer is in a single hop? Does the type of supervision ma;er?
Government posi5on Chief of Protocol, actress, singer Span-based vs. Mul5ple-choice Oberoi Family Delhi
Finding the answer directly
Ques%on: What government posi5on was held by the woman who portrayed Corliss Archer in the film Kiss and Tell ?
Kaushik and Lipton (2018) Shirley Temple Black was an American actress, businesswoman, and singer …
…
As an adult, she served as Chief of Protocol of the United States
Doc 1
Kiss and Tell is a comedy film in which 17-year-old Shirley Temple acts as Corliss Archer .
…
Doc 2
Meet Corliss Archer is an American television sitcom that aired on CBS …
Doc 3
Chief of Protocol businesswoman … actress
No Context Baseline
Dot Product
Ques%on: What government posi5on was held by the woman who portrayed Corliss Archer in the film Kiss and Tell ?
Chief of Protocol businesswoman actress Context Encoder Context Encoder
. . .
Chief of Protocol 0.7 businesswoman 0.2 actress 0.1
Accuracy
30 50 70
59.3 67.4 66.4 64.8 42.9 38.8
NoContext En5ty- GCN CFC BAG Majority- candidate BiDAF state-of-the-art weak baselines
Results on WikiHop
NoContext En5ty- GCN CFC BAG Majority- candidate BiDAF
Accuracy
30 50 70
59.3 67.4 66.4 64.8 42.9 38.8
NoContext En5ty- GCN CFC BAG Majority- candidate BiDAF state-of-the-art weak baselines
Results on WikiHop
NoContext En5ty- GCN CFC BAG Majority- candidate BiDAF
Accuracy
30 50 70
59.3 67.4 66.4 64.8 42.9 38.8
NoContext En5ty- GCN CFC BAG Majority- candidate BiDAF state-of-the-art weak baselines
Results on WikiHop
NoContext En5ty- GCN CFC BAG Majority- candidate BiDAF
Accuracy
30 50 70
59.3 67.4 66.4 64.8 42.9 38.8
NoContext En5ty- GCN CFC BAG Majority- candidate BiDAF state-of-the-art weak baselines
Results on WikiHop
NoContext En5ty- GCN CFC BAG Majority- candidate BiDAF
Accuracy
30 50 70
59.3 67.4 66.4 64.8 42.9 38.8
NoContext En5ty- GCN CFC BAG Majority- candidate BiDAF
More than half of ques5ons can be answered without even using the context!
state-of-the-art weak baselines
Results on WikiHop
NoContext En5ty- GCN CFC BAG Majority- candidate BiDAF
Accuracy
30 50 70
59.3 67.4 66.4 64.8 42.9 38.8
NoContext En5ty- GCN CFC BAG Majority- candidate BiDAF
- SOTA models trained on this may
learn ques5on-answer correspondences, not mul5-hop reasoning as adver5sed More than half of ques5ons can be answered without even using the context!
state-of-the-art weak baselines
Results on WikiHop
NoContext En5ty- GCN CFC BAG Majority- candidate BiDAF
Research Ques5ons
Can a model iden5fy the answer with only a set of candidates? Can a model iden5fy where the answer is in a single hop? Does the type of supervision ma;er?
Government posi5on Chief of Protocol, actress, singer Span-based vs. Mul5ple-choice Oberoi Family Dehli
Finding the answer directly
Ques%on: What government posi5on was held by the woman who portrayed Corliss Archer in the film Kiss and Tell ?
Example picked from HotpotQA (Yang 2018)
Shirley Temple Black was an American actress, businesswoman, and singer …
…
As an adult, she served as Chief of Protocol of the United States
Doc 1
Kiss and Tell is a comedy film in which 17-year-old Shirley Temple acts as Corliss Archer .
…
Doc 2
Meet Corliss Archer is an American television sitcom that aired on CBS …
Doc 3
A: Chief of Protocol B: businesswoman C: Actress ….
Finding the answer directly
Ques%on: What government posi5on was held by the woman who portrayed Corliss Archer in the film Kiss and Tell ?
Example picked from HotpotQA (Yang 2018)
Shirley Temple Black was an American actress, businesswoman, and singer …
…
As an adult, she served as Chief of Protocol of the United States
Doc 1
Kiss and Tell is a comedy film in which 17-year-old Shirley Temple acts as Corliss Archer .
…
Doc 2
Meet Corliss Archer is an American television sitcom that aired on CBS …
Doc 3
A: Chief of Protocol B: businesswoman C: Actress ….
There is no candidate set on HotpotQA
Sentence Factored Model
Find the answer by comparing each sentence with the ques5on separately! Ques%on: The Oberoi family is part of a hotel company that has a head
- ffice in what city?
The Oberoi Group is a hotel company with its head office in Delhi.
Doc 2
Future Fibre Technologies a fiber technologies company …
Doc 3
The Oberoi family is an Indian family that is …
Doc 1
Sentence Factored Model
The Oberoi family … what city? The Oberoi Group … in Delhi. Future Fibre Technologies is a fibre… The Oberoi family … Answer predic5on: Delhi
BiDAF BiDAF BiDAF
Sentence Factored Model
The Oberoi family … what city? The Oberoi Group … in Delhi. Future Fibre Technologies is a fibre… The Oberoi family … Answer predic5on: Delhi
- Sofmax over all sentences is the only cross-sentence interac5on
BiDAF BiDAF BiDAF
BiDAF++ QFE GRN DFGN Sentence Factored
F1 35 70
50.8 69.7 69.0 68.1 58.7
Results on HotpotQA
BiDAF++ QFE GRN DFGN Sentence Factored
BiDAF++ QFE GRN DFGN Sentence Factored
F1 35 70
50.8 69.7 69.0 68.1 58.7
Results on HotpotQA
BiDAF++ QFE GRN DFGN Sentence Factored
BiDAF++ QFE GRN DFGN Sentence Factored
F1 35 70
50.8 69.7 69.0 68.1 58.7
Results on HotpotQA
BiDAF++ QFE GRN DFGN Sentence Factored
BiDAF++ QFE GRN DFGN Sentence Factored
F1 35 70
50.8 69.7 69.0 68.1 58.7
Results on HotpotQA
BiDAF++ QFE GRN DFGN Sentence Factored
BiDAF++ QFE GRN DFGN Sentence Factored
F1 35 70
50.8 69.7 69.0 68.1 58.7
A simple single sentence reasoning model could solve more than half ques5ons on HotpotQA.
Results on HotpotQA
BiDAF++ QFE GRN DFGN Sentence Factored
Sentence Factored Model
The Oberoi family … what city? The Oberoi Group … in Delhi. Future Fibre Technologies … The Oberoi family … Containing answer: 1
BiDAF BiDAF BiDAF
Sentence Factored Model
The Oberoi family … what city? The Oberoi Group … in Delhi. Future Fibre Technologies … The Oberoi family … Containing answer: 1
- Predict which sentence contains the answer
BiDAF BiDAF BiDAF
Sentence Factored Model
The Oberoi family … what city? The Oberoi Group … in Delhi. Future Fibre Technologies … The Oberoi family … Containing answer: 1
- Predict which sentence contains the answer
BiDAF BiDAF BiDAF
- A simpler way to test how much mul5-hop reasoning is needed
Accuracy 45 90
5.4 57.2 6.5 66.1 22.1 88
SQuAD HotpotQA
Answer Sentence Predic5on Results
WikiHop SQuAD HotpotQA WikiHop
Random Sentence-factored
Accuracy 45 90
5.4 57.2 6.5 66.1 22.1 88
SQuAD HotpotQA
Answer Sentence Predic5on Results
WikiHop SQuAD HotpotQA WikiHop
Random Sentence-factored
- Answers can be found for
many ques5ons without a mul5-hop process.
Research Ques5ons
Can a model iden5fy the answer with only a set of candidates? Can a model iden5fy where the answer is in a single hop? Does the type of supervision ma;er?
Government posi5on Chief of Protocol, actress, singer Span-based vs. Mul5ple-choice Oberoi Family Delhi
Making HotpotQA Mul5ple-Choice
The Oberoi family is an Indian family that is famous for its involvement in hotels, namely through The Oberoi Group Ques%on: The Oberoi family is part of a hotel company that has a head
- ffice in what city?
The Oberoi Group is a hotel company with its head office in Delhi.
…
…
Doc 1 Doc 2
Future Fibre Technologies a fiber technologies company with its US head office in Mountain View, Middle East head office in Dubai …
Doc 3
Making HotpotQA Mul5ple-Choice
The Oberoi family is an Indian family that is famous for its involvement in hotels, namely through The Oberoi Group Ques%on: The Oberoi family is part of a hotel company that has a head
- ffice in what city?
The Oberoi Group is a hotel company with its head office in Delhi.
…
…
Doc 1 Doc 2
Future Fibre Technologies a fiber technologies company with its US head office in Mountain View, Middle East head office in Dubai …
Doc 3
- Randomly choose en55es in
the documents as candidates. The Oberoi family, The Oberoi Group, Delhi, Future Fibre Technologies, US, Mountain View, Middle East, Dubai
Results of HotpotQA Mul5ple-Choice
40 60 80 10 20 30 40 50
NoContext BiDAF++
Number of candidates Accuracy
Results of HotpotQA Mul5ple-Choice
40 60 80 10 20 30 40 50
NoContext BiDAF++
Number of candidates Accuracy
2.0
Results of HotpotQA Mul5ple-Choice
40 60 80 10 20 30 40 50
NoContext BiDAF++
Number of candidates Accuracy
2.0 2.9
Results of HotpotQA Mul5ple-Choice
40 60 80 10 20 30 40 50
NoContext BiDAF++
Number of candidates Accuracy
2.0 2.9 3.8
Results of HotpotQA Mul5ple-Choice
40 60 80 10 20 30 40 50
NoContext BiDAF++
Number of candidates Accuracy
2.0 2.9 3.8 2.5
Results of HotpotQA Mul5ple-Choice
40 60 80 10 20 30 40 50
NoContext BiDAF++
Number of candidates Accuracy
2.0 2.9 3.8 2.5 2.5
Results of HotpotQA Mul5ple-Choice
40 60 80 10 20 30 40 50
NoContext BiDAF++
More candidates does make the predic5on harder, but BiDAF++ does not learn more than the NoContext baseline
Number of candidates Accuracy
2.0 2.9 3.8 2.5 2.5
Conclusions
Can a model iden5fy the answer with only a set of candidates? Can a model iden5fy where the answer is in a single hop? Does the type of supervision ma;er?
Conclusions
Can a model iden5fy the answer with only a set of candidates? Can a model iden5fy where the answer is in a single hop? Does the type of supervision ma;er?
- Models can achieve high performance by “chea5ng” through
extra candidate informa5on.
Conclusions
Can a model iden5fy the answer with only a set of candidates? Can a model iden5fy where the answer is in a single hop? Does the type of supervision ma;er?
- Models can achieve high performance by “chea5ng” through
extra candidate informa5on.
- More than half of the ques5ons in WikiHop and HotpotQA can
be answered in a single hop.
Conclusions
Can a model iden5fy the answer with only a set of candidates? Can a model iden5fy where the answer is in a single hop? Does the type of supervision ma;er?
- Models can achieve high performance by “chea5ng” through
extra candidate informa5on.
- More than half of the ques5ons in WikiHop and HotpotQA can
be answered in a single hop.
- Span-based datasets may be less hackable.
Conclusions
Can a model iden5fy the answer with only a set of candidates? Can a model iden5fy where the answer is in a single hop? Does the type of supervision ma;er?
- Models can achieve high performance by “chea5ng” through
extra candidate informa5on.
- More than half of the ques5ons in WikiHop and HotpotQA can
be answered in a single hop.
- Span-based datasets may be less hackable.
*No results for WikiHop-Masked or HotpotQA fullwiki sepngs
Takeaways
When construc5ng mul5-hop based dataset:
Takeaways
- If using mul5ple-choice supervision, check the no-context baseline to see
whether there are strong correla5ons between ques5on and candidates. When construc5ng mul5-hop based dataset:
Takeaways
- If using mul5ple-choice supervision, check the no-context baseline to see
whether there are strong correla5ons between ques5on and candidates.
- Do an adversarial test using a sentence-factored model.
When construc5ng mul5-hop based dataset:
Takeaways
- If using mul5ple-choice supervision, check the no-context baseline to see
whether there are strong correla5ons between ques5on and candidates.
- Do an adversarial test using a sentence-factored model.