Towards Explainable Fact Checking
1 like336 views
Automatic fact checking is one of the more involved NLP tasks currently researched: not only does it require sentence understanding, but also an understanding of how claims relate to evidence documents and world knowledge. Moreover, there is still no common understanding in the automatic fact checking community of how the subtasks of fact checking — claim check-worthiness detection, evidence retrieval, veracity prediction — should be framed. This is partly owing to the complexity of the task, despite efforts to formalise the task of fact checking through the development of benchmark datasets. The first part of the talk will be on automatically generating textual explanations for fact checking, thereby exposing some of the reasoning processes these models follow. The second part of the talk will be on re-examining how claim check-worthiness is defined, and how check-worthy claims can be detected; followed by how to automatically generate claims which are hard to fact-check automatically.
![Generating Explanations from Ruling Comments
Claim:
We’ve tested more than every country
combined.
Ruling Comments:
Responding to weeks of criticism over his administration’s
COVID-19 response, President Donald Trump claimed at a
White House briefing that the United States has well
surpassed other countries in testing people for the virus.
"We’ve tested more than every country combined,"
Trump said April 27 […] We emailed the White House for
comment but never heard back, so we turned to the data.
Trump’s claim didn’t stand up to scrutiny.
In raw numbers, the United States has tested more
people than any other individual country — but
nowhere near more than "every country combined" or,
as he said in his tweet, more than "all major countries
combined.”[…] The United States has a far bigger
population than many of the "major countries" Trump often
mentions. So it could have run far more tests but still
have a much larger burden ahead than do nations like
Germany, France or Canada.[…]
Joint Model
Veracity Label
Justification/
Explanation](https://image.slidesharecdn.com/2020fevertalkpostedit-201027073033/85/Towards-Explainable-Fact-Checking-19-320.jpg)
Ad
More Related Content
What's hot (20)
Similar to Towards Explainable Fact Checking (20)
Ad
More from Isabelle Augenstein (20)
Ad
Recently uploaded (20)
Towards Explainable Fact Checking
- 1. FEVER Workshop 9 July 2020 Towards Explainable Fact Checking Isabelle Augenstein* [email protected] @IAugenstein http://isabelleaugenstein.github.io/ *partial slide credit: Pepa Atanasova, Dustin Wright
- 2. Overview of Today’s Talk • Introduction • Fact checking – where are we at? • Open research questions • Part 1: Generating explanations • A first case study • Generating explanations inspired by how humans fact-check • Part 2: Understanding claims • What even is a check-worthy claim? • Automatically generating confusing claims
- 3. Full Fact Checking Pipeline Claim Check- Worthiness Detection Evidence Document Retrieval Stance Detection / Textual Entailment Veracity Prediction “Immigrants are a drain on the economy” not check-worthy check-worthy “Immigrants are a drain on the economy” “Immigrants are a drain on the economy”, “EU immigrants have a positive impact on public finances” positive negative neutral true false not enough info“Immigrants are a drain on the economy”
- 4. How is fact checking framed in NLP? 4 Veracity Prediction true false neutral“Immigrants are a drain on the economy” • Most early work on detecting false information follows this setup • Deception detection (Mihalcea & Strapparava, 2009; Feng et al., 2012, Morris et al., 2012) • Clickbait detection (Chen et al., 2015; Potthast et al., 2016) • Fake news detection (Rubin et al., 2016; Volkova et al., 2017; Wang, 2017) • Option for early detection of fake news or rumours (”zero-shot setting”) • Prediction not evidence-based, but strictly based on linguistic cues 👍 👎
- 5. How is fact checking framed in NLP? 5 • Relatively well established tasks in NLP • Stance Detection (Anand et al., 2011; Mohammad et al., 2016) • Recognising Textual Entailment (Dagan et al., 2004; Bowman et al., 2015) • … more recently also studied in the context of fact checking • Fake News Challenge (Pomerleau & Rao, 2017) • Option for fact checking based on one evidence document • Central component of fact checking can be studied in isolation • Additional methods are needed for combining predictions of multiple evidence documents 👍 👎 Stance Detection / Textual Entailment “Immigrants are a drain on the economy”, “EU immigrants have a positive impact on public finances” positive negative neutral 👍
- 6. How is fact checking framed in NLP? 6 • Setup followed by more recent research, e.g. FEVER (Thorne et al., 2018) and MultiFC (Augenstein et al., 2019) • Includes most tasks of the fact checking pipeline • Inter-dependencies of components makes it difficult to understand how a prediction is derived • The difficulty of the task means models are relatively easily fooled 👍 👎 Evidence Document Retrieval Stance Detection / Textual Entailment Veracity Prediction “Immigrants are a drain on the economy” “Immigrants are a drain on the economy”, “EU immigrants have a positive impact on public finances” positive negative neutral true false not enough info“Immigrants are a drain on the economy” 👎
- 7. Research Challenges in Fact Checking 7 • Zero-shot or few-shot fact checking • How to fact-check claims with little available evidence? • Can we do better than the ”claim-only” setting? • Another option: prediction based on user responses on social media (e.g. Chen et al., 2019) • Combining multiple evidence documents for fact checking • Strength of evidence can differ • Learn this as part of the model (Augenstein et al. 2019) • Documents can play different roles, e.g. evidence, counter-evidence • Some early work on combining one piece of evidence with one of counter-evidence (Yin and Roth, 2018) • Documents can address different parts of the claim • Multi-hop fact checking (Nishida et al., 2019)
- 8. Research Challenges in Fact Checking 8 • Better understanding how a fact checking prediction is derived • Explainability metrics or tools can be used • Those do not necessarily highlight inter-dependencies between claim and evidence or provide an easy human-understandable explanation Ø Addressed in this talk • Claim check-worthiness detection • This is typically seen as a preprocessing task and rarely studied as part of the fact checking pipeline • Due to being relatively understudied, the problem is ill-defined Ø Addressed in this talk
- 9. Research Challenges in Fact Checking 9 • Fact-checking systems are easily fooled • ... even with completely non-sensical inputs • Studying what types of claims can fool models can help make them more robust (FEVER 2.0) • However, it is very challenging to fully automatially create claims that look ”natural” to humans Ø Addressed in this talk
- 11. Generating Fact Checking Explanations Pepa Atanasova, Jakob Grue Simonsen, Christina Lioma, Isabelle Augenstein ACL 2020 11
- 12. A Typical Fact Checking Pipeline Claim Check- Worthiness Detection Evidence Document Retrieval Stance Detection / Textual Entailment Veracity Prediction “Immigrants are a drain on the economy” not check-worthy check-worthy “Immigrants are a drain on the economy” “Immigrants are a drain on the economy”, “EU immigrants have a positive impact on public finances” positive negative neutral true false not enough info“Immigrants are a drain on the economy”
- 15. Automating Fact Checking Statement: “The last quarter, it was just announced, our gross domestic product was below zero. Who ever heard of this? Its never below zero.” Speaker: Donald Trump Context: presidential announcement speech Label: Pants on Fire Wang, William Yang. "“Liar, Liar Pants on Fire”: A New Benchmark Dataset for Fake News Detection." Proceedings ACL’2017.
- 16. Automating Fact Checking Claim: We’ve tested more than every country combined. Metadata: topic=healthcare, Coronavirus, speaker=Donald Trump, job=president, context=White House briefing, history={4, 10, 14, 21, 34, 14} Hybrid CNN Veracity Label 0,204 0,208 0,247 0,274 0 0,1 0,2 0,3 Validation Test Fact checking macro F1 score Majority Wang et. al Wang, William Yang. "“Liar, Liar Pants on Fire”: A New Benchmark Dataset for Fake News Detection." Proceedings ACL’2017.
- 17. Automating Fact Checking Alhindi, Tariq, Savvas Petridis, and Smaranda Muresan. "Where is your Evidence: Improving Fact-checking by Justification Modeling." Proceedings of FEVER’2018. Claim: We’ve tested more than every country combined. Justification: Trump claimed that the United States has "tested more than every country combined." There is no reasonable way to conclude that the American system has run more diagnostics than "all other major countries combined." Just by adding up a few other nations’ totals, you can quickly see Trump’s claim fall apart. Plus, focusing on the 5 million figure distracts from the real issue — by any meaningful metric of diagnosing and tracking, the United States is still well behind countries like Germany and Canada. The president’s claim is not only inaccurate but also ridiculous. We rate it Pants on Fire! Logistic Regression Veracity Label 0,204 0,208 0,247 0,274 0,370 0,370 0 0,2 0,4 Validation Test Fact checking macro F1 score Majority Wang et. al Alhindi et. Al
- 18. How about an explanation? Claim: We’ve tested more than every country combined. Justification: Trump claimed that the United States has "tested more than every country combined." There is no reasonable way to conclude that the American system has run more diagnostics than "all other major countries combined." Just by adding up a few other nations’ totals, you can quickly see Trump’s claim fall apart. Plus, focusing on the 5 million figure distracts from the real issue — by any meaningful metric of diagnosing and tracking, the United States is still well behind countries like Germany and Canada. The president’s claim is not only inaccurate but also ridiculous. We rate it Pants on Fire! Model Veracity Label Explanation
- 19. Generating Explanations from Ruling Comments Claim: We’ve tested more than every country combined. Ruling Comments: Responding to weeks of criticism over his administration’s COVID-19 response, President Donald Trump claimed at a White House briefing that the United States has well surpassed other countries in testing people for the virus. "We’ve tested more than every country combined," Trump said April 27 […] We emailed the White House for comment but never heard back, so we turned to the data. Trump’s claim didn’t stand up to scrutiny. In raw numbers, the United States has tested more people than any other individual country — but nowhere near more than "every country combined" or, as he said in his tweet, more than "all major countries combined.”[…] The United States has a far bigger population than many of the "major countries" Trump often mentions. So it could have run far more tests but still have a much larger burden ahead than do nations like Germany, France or Canada.[…] Joint Model Veracity Label Justification/ Explanation
- 20. Related Studies on Generating Explanations • Camburu et. al; Rajani et. al generate abstractive explanations • Short input text and explanations; • Large amount of annotated data. • Real world fact checking datasets are of limited size and the input consists of long documents • We take advantage of the LIAR-PLUS dataset: • Use the summary of the ruling comments as a gold explanation; • Formulate the problem as extractive summarization. • Camburu, Oana-Maria, Tim Rocktäschel, Thomas Lukasiewicz, and Phil Blunsom. "e-SNLI: Natural language inference with natural language explanations." In Advances in Neural Information Processing Systems,. 2018. • Rajani, Nazneen Fatema, Bryan McCann, Caiming Xiong, and Richard Socher. "Explain Yourself! Leveraging Language Models for Commonsense Reasoning." In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pp. 4932-4942. 2019.
- 21. Example of an Oracle’s Gold Summary Claim: “The president promised that if he spent money on a stimulus program that unemployment would go to 5.7 percent or 6 percent. Those were his words.” Label: Mostly-False Just: Bramnick said “the president promised that if he spent money on a stimulus program that unemployment would go to 5.7 percent or 6 percent. Those werehis words.” Two economic advisers estimated in a 2009 report that with the stimulus plan, the unemployment rate would peak near 8 percent before dropping to less than 6 percent by now. Those are critical details Bramnick’s statement ignores. To comment on this ruling, go to NJ.com. Oracle: “The president promised that if he spent money on a stimulus program that unemployment would go to 5.7 percent or 6 percent. Those were his words,”Bramnick said in a Sept. 7 interview on NJToday. But with the stimulus plan, the report projected the nation’s jobless rate would peak near 8 percent in 2009 before falling to about 5.5 percent by now. So the estimates in the report were wrong.
- 22. Which sentences should be selected for the explanation?
- 23. What is the veracity of the claim?
- 24. Joint Explanation and Veracity Prediction Ruder, Sebastian, et al. "Latent multi-task architecture learning." Proceedings of the AAAI’2019. Cross-stitch layer Multi-task objective
- 25. Fact Checking Results 0,208 0,274 0,323 0,300 0,313 0,370 0,443 0 0,05 0,1 0,15 0,2 0,25 0,3 0,35 0,4 0,45 0,5 Test Macro F1 score on the test split Majority Wang et. al MT Ruling Ruling Oracles Ruling Alhindi et. al Justification
- 26. 28,11 6,96 24,38 43,57 22,23 39,26 35,70 13,51 31,5831,13 12,90 30,93 0,00 5,00 10,00 15,00 20,00 25,00 30,00 35,00 40,00 45,00 50,00 ROUGE-1 ROUGE-2 ROUGE-L ROUGE scores on the test split Lead-4 Oracle Extractive Extractive-MT Explanation Results *ROUGE-N: Overlap of N-grams between the system and reference summaries. *ROUGE-L : Longest Common Subsequence (LCS) between the system and reference summaries.
- 27. Manual Evaluation • Explanation Quality • Coverage. The explanation contains important, salient information and does not miss any important points that contribute to the fact check. • Non-redundancy. The summary does not contain any information that is redundant/repeated/not relevant to the claim and the fact check. • Non-contradiction. The summary does not contain any pieces of information contradictory to the claim and the fact check. • Overall. Rank the explanations by their overall quality. • Explanation Informativeness. Provide a veracity label for a claim based on a veracity explanation coming from the justification, the Explain-MT, or the Explain- Extractive system.
- 28. Explanation Quality 1,48 1,48 1,45 1,58 1,89 1,75 1,40 2,03 1,68 1,79 1,48 1,90 - 0,50 1,00 1,50 2,00 2,50 Coverage Non-redundancy Non-contradicton Overall Mean Average Rank (MAR). Lower MAR is better! (higher rank) Justification Extractive Extractive MT
- 29. Explanation Informativeness - 0,100 0,200 0,300 0,400 0,500 Agree-Correct Agree-Not Correct Agree-Nost Sufficient Disagree Manual veracity labelling, given a particular explanation as percentages of the dis/agreeing annotator predictions. Justification Extractive Extractive-MT
- 30. Summary • First study on generating veracity explanations • Jointly training veracity prediction and explanation • improves the performance of the classification system • improves the coverage and overall performance of the generated explanations
- 31. Future Work • Can we generate better or even abstractive explanations given limited resources? • How to automatically evaluate the properties of the explanations? • Can explanations be extracted from evidence pages only (lots of irrelevant and multi-modal results)?
- 33. Claim Check-Worthiness Detection as Positive Unlabelled Learning Dustin Wright, Isabelle Augenstein Findings of EMNLP, 2020 33
- 34. A Typical Fact Checking Pipeline Claim Check- Worthiness Detection Evidence Document Retrieval Stance Detection / Textual Entailment Veracity Prediction “Immigrants are a drain on the economy” not check-worthy check-worthy “Immigrants are a drain on the economy” “Immigrants are a drain on the economy”, “EU immigrants have a positive impact on public finances” positive negative neutral true false not enough info“Immigrants are a drain on the economy”
- 35. What is Check-Worthiness? 35 - A check-worthy statement is ”an assertion about the world that is checkable” (Konstantinovskiy et al., 2018) - However, there is more to this: - The statement should be factual - Typically, only sentences unlikely to be believed without verification are marked as check-worthy - Coupled with claim importance, which is subjective - Domain interests skew what is selected as check-worthy (e.g. only certain political claims)
- 36. What is Check-Worthiness? 36 - Many different types of claims (Francis and FullFact, 2016) - numerical claims involving numerical properties about entities and comparisons among them - entity and event properties such as professional qualifications and event participants - position statements such as whether a political entity supported a certain policy - quote verification assessing whether the claim states precisely the source of a quote, its content, and the event at which it supposedly occurred.
- 37. Wikipedia Citation Needed Detection Twitter Rumour Detection Political Check-Worthiness Detection Given a statement from Wikipedia, determine if the statement needs a citation in order to be accepted as true 37 Determine if the content of a tweet was verified as true or false at the time of posting Given statements from a political speech or debate, rank them by how check-worthy they are What is Check-Worthiness?
- 39. Baseline Approach for Claim Check- Worthiness Detection: Binary Classification 39 𝑔 ∶ 𝑥 → (0, 1) Reviewers described the book as “magistral”, “encyclopaedic” and a “classic”. l = 1 As was pointed out above, Lenten traditions have developed over time. l = 0
- 40. A Unified Approach • Utilise existing published datasets on check- worthiness detection and related tasks • Use Positive Unlabelled (PU) learning to overcome issues with subjective annotation • Transfer learning from Wikipedia citation needed detection • Evaluate how well each dataset reflects a general definition of check-worthiness 40
- 41. Solution: Positive Unlabelled Learning 41 • Assumptions: • Not all ‘true’ instances are labelled as such • Labelled samples are selected at random from an underlying distribution • If a sample is labelled, then the label cannot be ‘false’ • Intuition: • Train another model, 𝑓, first, which estimates the confidence in a sample being ‘true’ • Automatically label unlabelled samples using 𝑓 to train model 𝑔
- 42. A Traditional Classifier 42 𝑔 ∶ 𝑥 → (0, 1) Reviewers described the book as “magistral”, “encyclopaedic” and a “classic”. l = 1 As was pointed out above, Lenten traditions have developed over time. l = 0 Positive-Negative Classifier All three of his albums appeared on Rolling Stone magazine's list of The 500 Greatest Albums of All Time in 2003 l = 0 True negative?
- 43. Positive Unlabelled Learning 43 𝑓 ∶ 𝑥 → (0, 1) s = 0 s = 0 s = 1 Positive-Unlabelled Classifier 1.0 0.0 0.05 0.95 0.8 0.2 𝑤 𝑥
- 44. Positive Unlabelled Learning 44 𝑓 s = 0 s = 0 s = 1 1.0 0.0 0.05 0.95 0.8 0.2 𝑔 l = 1 l = 0 l = 1 l = 0 l = 1 l = 0 𝑤 𝑥 Charles Elkan and Keith Noto. 2008. Learning classifiers from only positive and unlabeled data. In Proceedings of KKD.
- 45. Positive Unlabelled Learning + Conversion 45 𝑓 s = 0 s = 0 s = 1 1.0 0.0 0.05 0.95 0.8 0.2 𝑔 l = 1 l = 0 l = 1 l = 0 l = 1 l = 0 1.0 Use estimate of 𝑝(𝑦 = 1) 𝑤 𝑥
- 46. Cross-Domain Check-Worthiness Detection 46 𝑓0 𝑔0 𝑔1 𝑔2 All models: BERT base uncased pretrained
- 47. Research Questions • To what degree is check-worthiness detection a PU learning problem, and does this enable a unified approach to check-worthiness detection? • How well do statements in each dataset reflect the definition: ”makes an assertion about the world that is checkable?” 47
- 48. Results: Citation Needed Detection 48 • Dataset: • ’Citation needed’ detection (Redi et al., 2019) • 10k positive, 10k negative for training (featured articles) • 10k positive, 10k negative for testing (flagged statements) 0,68 0,7 0,72 0,74 0,76 0,78 0,8 0,82 0,84 Redi et al. 2019 BERT BERT + PU BERT + PUC Macro F1
- 49. Results: Twitter 49 • Dataset: • Pheme Rumour Detection (Zubiaga et al., 2017) • 5802 tweets from 5 events • 5-fold cross-validation, event hold-out 0,54 0,56 0,58 0,6 0,62 0,64 0,66 0,68 0,7 BiLSTM Zubiaga et al. 2017 BERT BERT + Wiki BERT + Wiki + PU BERT + Wiki + PUC Micro-F1
- 50. Results: Political Speeches 50 • Dataset: • Claims from political speeches (Hansen et al., 2018) • Consists of 4 political speeches from 2018 Clef CheckThat! Competition & 3 speeches from ClaimRank • 2602 statements overall • 7-fold cross-validation, speech hold-out 0,28 0,29 0,3 0,31 0,32 0,33 0,34 0,35 Hansen et al. 2019 BERT BERT + Wiki BERT + Wiki + PU BERT + Wiki + PUC Mean Average Precision
- 51. Error Analysis: How Check-Worthy are Check- Worthy Statements? 51 • Method: • Get top 100 positive predictions by PUC model on each dataset • Re-annotate those manually (2 annotators) • Compare against ’gold’ annotations 0 0,2 0,4 0,6 0,8 1 Wikipedia Twitter Politics
- 52. Dataset Analysis • Wikipedia and Twitter equally well reflect a general definition of check-worthiness • Political speeches have high false negative rate • Also documented in previous literature (Konstantinovskiy et al. 2018) • Clear examples: these two statements appear in a single document, only one was labelled as check-worthy • ”It’s why our unemployment rate is the lowest it’s been in so many decades” • ”New unemployment claims are near the lowest we’ve seen in almost half a century” 52
- 53. Key Takeaways • Check-worthiness annotation is highly subjective • Because of this, automatic check-worthiness detection is a difficult task • Approaching the problem as PU learning shows improvements • Transferring from Wikipedia improves on downstream domains when PU learning is used 53
- 54. Generating Label Cohesive and Well-Formed Adversarial Claims Pepa Atanasova*, Dustin Wright*, Isabelle Augenstein Proceedings of EMNLP, 2020 *equal contributions 54
- 55. A Typical Fact Checking Pipeline Claim Check- Worthiness Detection Evidence Document Retrieval Stance Detection / Textual Entailment Veracity Prediction “Immigrants are a drain on the economy” not check-worthy check-worthy “Immigrants are a drain on the economy” “Immigrants are a drain on the economy”, “EU immigrants have a positive impact on public finances” positive negative neutral true false not enough info“Immigrants are a drain on the economy”
- 56. Generating Adversarial Claims • Fact checking models can overfit to spurious patterns • Making the right predictions for the wrong reasons • This leads to vulnerabilities, which can be exploited by adversaries (e.g. agents spreading mis- and disinformation) • How can one reveal such vulnerabilities? • Generating explanations for fact checking models (first part of talk) • Generating adversarial claims (this work)
- 57. Previous Work • Universal adversarial attacks (Gao and Oates, 2019; Wallace et al, 2019) • Single perturbation changes that can be applied to many instances • Change the meaning of the input instances and thus produce label-incoherent claims • Are not per se semantically well-formed • Rule-based perturbations (Riberio et al., 2018) • Semantically well-formed, but require hand-crafting patterns
- 58. Previous Work • FEVER 2.0 shared task (Thorne et al., 2019) • Builders / breakers setup • Methods of submitted systems: • Producing claims requiring multi-hop reasoning (Niewinski et al., 2019) • Generating adversarial claims manually (Kim and Allan, 2019)
- 59. Goals of this Work • Generate claims fully automatically • Preserve the meaning of the source text • Produce semantically well-formed claims
- 60. Model RoBERTa-based FEVER model to predict FC label 1) HotFlip attack model to find triggers 2) STS auxiliary model to preserve FC label
- 61. Examples of Generated Claims 61 EMNLP 2020 Submission ***. Confidential Review Copy. DO NOT DISTRIBUTE. Evidence Triggers Generated Claim SUPPORTS Claims Since the 19th century, some Romani have also migrated to the Americas. don,already,more,during,home Romani have moved to the Americas during the 19th century. Cyprus is a major tourist destination in the Mediter- ranean . foreign,biggest,major,every,friends Cyprus is a major tourist des- tination. The first Nobel Prize in Chemistry was awarded in 1901 to Jacobus Henricus van’t Hoff, of the Nether- lands, “for his discovery of the laws of chemical dynam- ics and osmotic pressure in solutions.” later,already,quite,altern,whereas Henricus Van’t Hoff was al- ready awarded the Nobel Prize. REFUTES Claims California Attorney General phys,incarn,not,occasionally,something Kamala Harris did not defeat Since the 19th century, some Romani have also migrated to the Americas. don,already,more,during,home Romani have moved to the Americas during the 19th century. Cyprus is a major tourist destination in the Mediter- ranean . foreign,biggest,major,every,friends Cyprus is a major tourist des- tination. The first Nobel Prize in Chemistry was awarded in 1901 to Jacobus Henricus van’t Hoff, of the Nether- lands, “for his discovery of the laws of chemical dynam- ics and osmotic pressure in solutions.” later,already,quite,altern,whereas Henricus Van’t Hoff was al- ready awarded the Nobel Prize. REFUTES Claims California Attorney General Kamala Harris defeated Sanchez , 61.6% to 38.4%. phys,incarn,not,occasionally,something Kamala Harris did not defeat Sanchez, 61.6% to 38.4%. Uganda is in the African Great Lakes region. unless,endorsed,picks,pref,against Uganda is against the African Great Lakes region. Times Higher Education World University Rankings is an annual publication of university rankings by Times Higher Education (THE) magazine. interested,reward,visit,consumer,conclusion Times Higher Education World University Rankings is a consumer magazine. NOT ENOUGH INFO Claims ranean . The first Nobel Prize in Chemistry was awarded in 1901 to Jacobus Henricus van’t Hoff, of the Nether- lands, “for his discovery of the laws of chemical dynam- ics and osmotic pressure in solutions.” later,already,quite,altern,whereas Henricus Van’t Hoff was al- ready awarded the Nobel Prize. REFUTES Claims California Attorney General Kamala Harris defeated Sanchez , 61.6% to 38.4%. phys,incarn,not,occasionally,something Kamala Harris did not defeat Sanchez, 61.6% to 38.4%. Uganda is in the African Great Lakes region. unless,endorsed,picks,pref,against Uganda is against the African Great Lakes region. Times Higher Education World University Rankings is an annual publication of university rankings by Times Higher Education (THE) magazine. interested,reward,visit,consumer,conclusion Times Higher Education World University Rankings is a consumer magazine. NOT ENOUGH INFO Claims The KGB was a military ser- vice and was governed by army laws and regulations , similar to the Soviet Army or MVD Internal Troops. nowhere,only,none,no,nothing The KGB was only con- trolled by a military service. The first Nobel Prize in Chemistry was awarded in 1901 to Jacobus Henricus van’t Hoff, of the Nether- lands, “for his discovery of the laws of chemical dynam- ics and osmotic pressure in solutions.” later,already,quite,altern,whereas Henricus Van’t Hoff was al- ready awarded the Nobel Prize. REFUTES Claims California Attorney General Kamala Harris defeated Sanchez , 61.6% to 38.4%. phys,incarn,not,occasionally,something Kamala Harris did not defeat Sanchez, 61.6% to 38.4%. Uganda is in the African Great Lakes region. unless,endorsed,picks,pref,against Uganda is against the African Great Lakes region. Times Higher Education World University Rankings is an annual publication of university rankings by Times Higher Education (THE) magazine. interested,reward,visit,consumer,conclusion Times Higher Education World University Rankings is a consumer magazine. NOT ENOUGH INFO Claims The KGB was a military ser- vice and was governed by army laws and regulations , similar to the Soviet Army or MVD Internal Troops. nowhere,only,none,no,nothing The KGB was only con- trolled by a military service. The series revolves around says,said,take,say,is Take Me High is about
- 62. Results: Automatic Evaluation • How well can we generate universal adversarial triggers? • Reduction in F1 for both trigger generation methods • Trade-off between how potent the attack is (reduction in F1) vs. how semantically coherent the claim is (STS) • Most potent attack: S->R for FC Objective (F1: 0.060) • Adds negation words, which change meaning of claim 0 0,5 1 F1 F1 of FC model (lower = better) No Triggers FC Objective FC + STS Objective 4 4,5 5 5,5 STS Semantic Textual Similarity with original claim (higher = better) No Triggers FC Objective FC + STS Objective
- 63. Results: Manual Evaluation • How well can we generate universal adversarial triggers? • Confirms results of automatic evaluation • Macro F1 of generated claim w.r.t. original claim: 56.6 (FC Objective); 60.7 (FC + STS Objective) • STS Objective preserves meaning more often 0 5 Overall SUPPORTS REFUTES NEI Semantic Textual Similarity with original claim (higher = better) FC Objective FC + STS Objective 0 0,5 1 Overall SUPPORTS REFUTES NEI F1 of FC model (lower = better) FC Objective FC + STS Objective
- 64. Key Take-Aways • Novel extension to the HotFlip attack for universarial adversarial trigger generation (Ebrahimi et al., 2018) • Conditional language model, which takes trigger tokens and evidence, and generates a semantically coherent claim • Resulting model generates semantically coherent claims containing universal triggers, which preserve the label • Trade-off between how well-formed the claim is and how potent the attack is
- 65. Overall Take-Aways • Automatic fact checking is still a very involved NLP task • Different sub-tasks: 1) claim check-worthiness detection; 2) evidence retrieval; 3) stance detection / textual entailment, 4) veracity prediction • What is needed to automatically generate explanations for fact checking models? • Common understanding of how sub-tasks should be defined (Part 2.1 of this talk) • Lack thereof leads to inconsistent models and findings, limited transferability of learned representations
- 66. Overall Take-Aways • What is needed to automatically generate explanations for fact checking models? • More realistic benchmarks for FC explanation generation (Part 1 of this talk) • Generating explanations from ruling comments in the LIAR-PLUS dataset is somewhat useful for researching models, but not very realistic • The dataset is also relatively small (10k training instances) • A more realistic setup and benchmark would inspire research of e.g. multi- modal models, and models for better understanding evidence documents
- 67. Overall Take-Aways • What is needed to automatically generate explanations for fact checking models? • More research on fooling FC models (Part 2.2 of this talk) • Adversarial trigger generation can reveal inconsistencies and vulnerabilities in FC models • Those can explain what lexical patterns a FC model struggles with • Most research focuses on instance-level perturbations of claims, whereas universal triggers reveal more general vulnerabilities of the model
- 69. Thanks to my PhD students and colleagues! 69 CopeNLU https://copenlu.github.io/ Pepa Atanasova Dustin Wright Jakob Grue Simonsen Christina Lioma
- 70. Presented Papers Pepa Atanasova, Jakob Grue Simonsen, Christina Lioma, Isabelle Augenstein. Generating Fact Checking Explanations. In Proceedings of ACL 2020. https://arxiv.org/abs/2004.05773 Dustin Wright, Isabelle Augenstein. Claim Check-Worthiness Detection as Positive Unlabelled Learning. In Findings of EMNLP 2020. https://arxiv.org/abs/2003.02736 Pepa Atanasova*, Dustin Wright*, Isabelle Augenstein. Generating Label Cohesive and Well-Formed Adversarial Claims. In Proceedings of EMNLP, 2020. https://arxiv.org/abs/2009.08205 *equal contributions