SLIDE 1
Inoculation by Fine-Tuning: A Method for Analyzing Challenge Datasets
Nelson F. Liu UWNLP NAACL 2019—June 4, 2019 Roy Schwartz Noah A. Smith
2
Training Dataset Model Architecture
Two Key Ingredients of NLP Systems
NLP System
Inoculation by Fine-Tuning: A Method for Analyzing Challenge - - PowerPoint PPT Presentation
Inoculation by Fine-Tuning: A Method for Analyzing Challenge - - PowerPoint PPT Presentation
Inoculation by Fine-Tuning: A Method for Analyzing Challenge Datasets Nelson F. Liu Roy Schwartz Noah A. Smith NAACL 2019June 4, 2019 UWNLP Two Key Ingredients of NLP Systems Training Model Dataset Architecture NLP System 2
SLIDE 2
SLIDE 3
3
Training Dataset Model Architecture
Why Might NLP Systems Fail?
NLP System
🤓
SLIDE 4
4
Training Dataset Model Architecture
Dataset Weaknesses
NLP System
🤓
SLIDE 5
5
Training Dataset Model Architecture
Model Weaknesses
NLP System
🤓
SLIDE 6
Challenge Datasets Break Models
6
SLIDE 7
7
Challenge Datasets Break Models
SLIDE 8
8
Challenge Datasets Break Models
SLIDE 9
NLP Systems Are Brittle
9
SLIDE 10
NLP Systems Are Brittle
10
SLIDE 11
Inoculation by Fine-Tuning
11
SLIDE 12
Inoculation by Fine-Tuning
12
SLIDE 13
Inoculation by Fine-Tuning
13
SLIDE 14
Inoculation
14
SLIDE 15
Inoculate Models to Better Understand Why They Fail
15
SLIDE 16
Three Clear Outcomes of Interest
16
Challenge Evaluation Outcome Inoculation
?
SLIDE 17
(1) Dataset Weakness
17
Challenge Evaluation Outcome Inoculation Dataset Weakness
SLIDE 18
(2) Model Weakness
18
Model Weakness Challenge Evaluation Outcome Inoculation
SLIDE 19
(3) Predictive Artifacts / Other
19
Predictive Artifacts / Other Challenge Evaluation Outcome Inoculation
SLIDE 20
Three Clear Outcomes of Interest
20
Dataset Weakness Model Weakness
Predictive Artifacts / Other
Challenge Evaluation Outcome Inoculation
SLIDE 21
Case Studies
- Inoculating natural language
inference (NLI) models
- Inoculating SQuAD reading
comprehension models
21
SLIDE 22
Natural Language Inference (NLI)
Premise: "I have done what you asked." Hypothesis: "I have disobeyed your orders."
22
Entailment Contradiction Neutral
[Dagan et al., 2004] Example from MultiNLI [Williams et al., 2018]
SLIDE 23
Two NLI Challenge Datasets
[Naik and Ravichander et al., 2018]
23
Premise: "I have done what you asked." Hypothesis: "I have disobeyed your orders."
SLIDE 24
Two NLI Challenge Datasets
Premise: "I have done what you asked." Hypothesis: "I have disobeyed your orders and true is true."
24
Premise: "I have done what you asked." Hypothesis: "I have disobeyed your orders."
Word Overlap Challenge Dataset
[Naik and Ravichander et al., 2018]
SLIDE 25
Two NLI Challenge Datasets
Premise: "I have done what you asked." Hypothesis: "I have disobeyed your orders and true is true." Premise: "I have done what you asked." Hypothesis: "I have disobeyed your ordets."
25
Premise: "I have done what you asked." Hypothesis: "I have disobeyed your orders."
Word Overlap Challenge Dataset Spelling Errors Challenge Dataset
[Naik and Ravichander et al., 2018]
SLIDE 26
Small Perturbations Break NLI Models
26
Word Overlap Spelling Errors
- 12.6%
(absolute)
- 4.8%
(absolute)
SLIDE 27
Inoculating NLI models
27
Word Overlap Spelling Errors
SLIDE 28
Inoculating NLI models
28
Word Overlap Spelling Errors
Dataset Weakness Model Weakness
SLIDE 29
More Examples in the Paper!
29
Dataset Weakness Dataset Weakness Model Weakness Model Weakness
Predictive Artifacts / Other
SLIDE 30
SQuAD
[Rajpurkar et al., 2016]
Example from Robin Jia
Question: "The number of new Huguenot colonists declined after what year?" Passage: "The largest portion of the Huguenots to settle in the Cape arrived between 1688 and 1689…but quite a few arrived as late as 1700; thereafter, the numbers declined…" Correct Answer: "1700"
30
SLIDE 31
Adversarial SQuAD
Question: "The number of new Huguenot colonists declined after what year?" Passage: "The largest portion of the Huguenots to settle in the Cape arrived between 1688 and 1689…but quite a few arrived as late as 1700; thereafter, the numbers declined. The number of
- ld Acadian colonists declined after the year of
1675." Correct Answer: "1700"
[Jia and Liang, 2017]
Example from Robin Jia
31
SLIDE 32
Small Perturbations Break SQuAD Models
32
- 24.5 F1
(absolute)
SLIDE 33
Inoculating SQuAD models
33
SLIDE 34
34
Predictive Artifacts / Other
Inoculating SQuAD models
SLIDE 35
Takeaways
35
- Inoculation by Fine-Tuning helps us understand
why our models fail.
- While all challenge datasets break our models,
they stress them in different ways.
Dataset Weakness Model Weakness
Predictive Artifacts / Other
- Potentially many situations where inoculation can help
clarify model results when transferring to other datasets.
SLIDE 36
Takeaways
36
Thank You! Questions?
- Inoculation by Fine-Tuning helps us understand
why our models fail.
- While all challenge datasets break our models,
they stress them in different ways.
Dataset Weakness Model Weakness
Predictive Artifacts / Other
- Potentially many situations where inoculation can help
clarify model results when transferring to other datasets.
SLIDE 37
Limitations of Inoculation by Fine-Tuning
- Requires a somewhat balanced label distribution in the
challenge dataset.
- Else, fine-tuned model will always predict majority label
- This method is not a silver bullet!
- First step toward disentangling failures of {original /
challenge} datasets and models.
37
SLIDE 38
38
SLIDE 39
Inoculating Multiple SQuAD Reading Comprehension Models
39
SLIDE 40
Inoculating Multiple NLI Models Against Word Overlap Adversary
40
SLIDE 41
Inoculating Multiple NLI Models Against Spelling Errors
41