Source-linked AI summary
Generating Fact Checking Explanations
Pepa Atanasova, Jakob Grue Simonsen, Christina Lioma, Isabelle Augenstein
TL;DR
Automated fact checking systems have largely focused on predicting claim veracity, while generating justifications remains understudied. This paper models extractive explanation generation and veracity prediction separately and jointly, finding that joint optimisation improves fact-checking performance and explanation quality.
Problem
Generating justifications for veracity verdicts remains understudied, although existing fact-checking systems commonly predict veracity from metadata and evidence.
Method
The paper frames explanation generation as extractive summarisation from ruling comments and trains models for explanation generation, veracity prediction, or both jointly.
Results
Jointly optimising veracity prediction and explanation generation improves the fact-checking system and yields explanations with better coverage and overall quality.
Takeaways & Limitations
The joint objective produces explanations that better support the correct veracity label than explanations trained solely to mimic human justifications.
Takeaways & Limitations
Extractive explanation systems still have room for improvement when restricted to selecting sentences from ruling comments.
Abstract
from arXiv · showhide
Most existing work on automated fact checking is concerned with predicting the veracity of claims based on metadata, social network spread, language used in claims, and, more recently, evidence supporting or denying claims. A crucial piece of the puzzle that is still missing is to understand how to automate the most elaborate part of the process -- generating justifications for verdicts on claims. This paper provides the first study of how these explanations can be generated automatically based on available claim context, and how this task can be modelled jointly with veracity prediction. Our results indicate that optimising both objectives at the same time, rather than training them separately, improves the performance of a fact checking system. The results of a manual evaluation further suggest that the informativeness, coverage and overall quality of the generated explanations are also improved in the multi-task model.
1 Introduction
Automated fact checking has advanced in predicting claim veracity, but producing concise, human-like justifications remains understudied. This paper studies automatic veracity-explanation generation and joint learning with veracity prediction.
- Limited manual fact-checking resources create a need to automate the fact-checking process.
- Existing systems estimate claim veracity from metadata and evidence, but rarely optimise relevant evidence selection as a self-sufficient explanation.
- Automatic production of justifications for veracity predictions remains an understudied problem, despite short human explanations improving macro F1 by over 10%.
- The paper frames explanation generation as summarisation from ruling comments and explores jointly learning explanation generation with veracity prediction.
- Jointly optimising veracity prediction and explanation improves veracity classification and produces explanations with better coverage and overall quality than explanations trained solely to mimic human justifications.
2 Dataset
The LIAR-PLUS dataset pairs political statements with veracity justifications extracted from PolitiFact ruling comments. The resulting examples contain substantially longer ruling comments than justifications.
- LIAR-PLUS contains 12,836 statements with veracity justifications from PolitiFact ruling comments.
- Justifications are automatically extracted from locations indicated at the ends of ruling comments, while label-indicating sentences are removed.
- The final dataset contains 10,146 training, 1,278 validation, and 1,255 test examples after removing reports shorter than three sentences.
- Ruling comments average 39 sentences or 904 words, whereas justifications average four sentences or 89 words.
3 Method
The method uses DistilBERT-based models for extractive explanation, veracity prediction, and joint learning. The joint model shares information between task-specific representations and optimises both objectives together.
- The paper compares separate explanation-extraction and veracity-prediction models with a joint model optimising both tasks.
- All models use language-model-pretrained DistilBERT embeddings fine-tuned for the specific task.
- Generating Explanations: The explanation model selects four ruling-comment sentences whose predicted relevance forms the extractive explanation.
- Generating Explanations: Oracle explanation labels are created by greedily selecting the top k = 4 sentences according to ROUGE-2 F1 against the gold justification.
- Fact Checking: The fact-checking model predicts one of six veracity labels from the claim and ruling comments using a softmax output.
- Joint Training: The joint model uses cross-stitch layers to combine task-specific and cross-task representations before predicting explanations and veracity labels.
- Joint Training: Joint training minimises LMT = γ ∗H(pE, yE) + η ∗ H(pF, yF), a weighted combination of explanation and veracity losses.
4 Automatic Evaluation
The automatic evaluation compares separate and jointly trained models for veracity prediction and explanation generation. Joint training achieves the best veracity-prediction performance, while explanation results reveal a trade-off between ROUGE similarity and fact-checking relevance.
- Experimental Setup: The evaluation compares jointly trained and separately trained models for veracity prediction and extractive explanation generation.The experiments include baselines and upper bounds based on metadata, ruling comments, oracle sentences, and gold justifications.
- Veracity Prediction: Macro F1 remains below 0.5 even when veracity models receive gold explanations, underscoring the task’s difficulty.The authors attribute this difficulty to the small dataset and/or the challenge of the task for human annotators.
- Veracity Prediction: The jointly trained MT-Veracity@Rul model achieves the best macro F1 among the three systems.Training the explanation objective helps the veracity model identify ruling-comment regions close to the gold explanation.
- Explanation Generation: The separately trained Explain-Extractive model outperforms the jointly trained Explain-MT system on automatic explanation evaluation.The authors hypothesise that joint training may prioritize fact-checking knowledge that does not match the wording of oracle explanations.
- Explanation Generation: Joint training can select sentences more important for the fact check, even when those sentences are less similar to the gold justification.This may improve the explanation’s information about the predicted veracity label despite lower ROUGE scores in some examples.
- Explanation Generation: Extractive explanation systems still have room for improvement because the Oracle represents only a greedy approximation of the gold justification.The Oracle serves as an approximate upper bound for extraction from ruling comments.
5 Manual Evaluation
The manual evaluation compares gold, extractive, and jointly learned explanations for quality and informativeness. Explain-MT generally outperforms Explain-Extr on coverage and overall quality, although annotator agreement is low.
- Evaluation Reliability: Krippendorff’s α was 0.26 for Coverage, 0.18 for Non-redundancy, -0.1 for Non-contradiction, and 0.32 for Overall.The reported agreement threshold is 0.67 < α < 0.8, although values vary by domain.
- Explanation Quality: Annotators ranked gold justifications highest on every criterion except Non-contradiction.The disagreement on Non-contradiction indicates no consensus for that criterion.
- Explanation Quality: Explain-MT improved over Explain-Extr by 0.21 in Coverage and 0.13 in Overall ranking position.These manual results favor Explain-MT despite its lower automatic ROUGE F1 scores.
- Explanation Quality: Explain-Extr ranked higher in Non-redundancy, with only a marginal 0.04 improvement, and in Non-contradiction.Annotators disagreed on Non-contradiction.
- Explanation Informativeness: Gold explanations were most sufficient for identifying veracity and explaining the correct label, followed by Explain-MT.Binary-label agreement was highest for Explain-Extr, with Fleiss’ κ values of 0.269 for Just, 0.345 for Explain-MT, and 0.399 for Explain-Extr.
6 Related Work
The paper addresses an understudied problem in NLP and fact checking: generating human-readable veracity explanations. It positions joint explanation generation and veracity prediction as a contribution distinct from prior prediction and interpretability systems.
- Generating Explanations: Earlier natural-language-inference work compared explanation-first, prediction-first, and joint multi-task pipelines.That study found explanation-first improved explanation results but harmed classification accuracy.
- Generating Explanations: The paper presents the first study of automatically generating veracity explanations for fact-checking claims.It frames explanation generation as producing summaries close to human justifications.
- Generating Explanations: Jointly optimizing veracity explanation and prediction improves explanation coverage and overall quality.The paper also reports that generated explanations improve veracity prediction performance.
- Fact Checking Interpretability: Prior fact-checking interpretability systems highlight salient inputs, whereas this work directly optimizes human-readable textual explanations.The cited systems use user comments, salient news sentences, or statement-only attention mechanisms.
- Dataset Scope: FEVER and MultiFC are larger fact-checking datasets but lack veracity explanations, so they were not used to train joint models in this study.This limits the datasets available for the paper’s joint training setup.
7 Conclusions
The paper’s conclusions show that veracity explanation generation can be combined with veracity prediction, with multi-task training improving both system performance and manually assessed explanation quality. Future work includes generating explanations from Web-crawled evidence pages and exploring methods that may improve fluency or reduce redundancy.
- The paper presents the first study of generating veracity explanations.
- Jointly training veracity prediction and explanation generation improves the performance of the veracity system.
- Manual evaluation finds that multi-task training improves explanation coverage and overall quality.
- Future work should investigate generating explanations from evidence pages crawled from the Web.
- Other explanation-generation approaches may improve fluency or decrease redundancy in generated text.
A.1 Comparison of different sources of evidence
The ruling comments provide substantial unigram and sequence coverage of the target justification, while oracle selection mainly improves precision and yields limited bigram coverage.
- Over 70% of the justification’s words and long sequences can be found in the ruling comments.This is reflected in high ROUGE-1 and ROUGE-F recall.
- Ruling comments have small coverage for bigrams in the justification.
- Selecting ruling oracles from all ruling sentences mainly improves ROUGE-F1 precision.
A.2 Extractive Gold Oracle Examples
The extractive oracle examples illustrate varying degrees of match with the gold justification, including an example that preserves important information and precise numbers.
- The three oracle examples represent different degrees of matching the gold summary.
- The first oracle matches the gold summary with a ROUGE-L F1 score of 60.40.
- The first oracle contains all important information from the gold summary and includes precise rather than rounded numbers.
A.3 Manual 6-way Veracity Prediction from explanations
Manual evaluation used six-way veracity annotations to compare explanations and report annotator agreement, including results for the gold and generated explanation systems.
- Fleiss’ κ agreement was 0.20 for Just explanations, 0.230 for Explain-MT, and 0.333 for Explain-Extr.
- The manual evaluation used six veracity classes: true, false, half-true, barely-true, mostly-true, and pants-on-fire.
- Table 9 reports manual veracity-prediction results for Just, Explain-Extr, and Explain-MT explanations.