Explanation generation¶

Explanation generation (or justification generation) refers to automatically producing human-readable text that justifies or explains a system's prediction or decision. In the context of misinformation and fact-checking, this means generating explanations for why a claim is judged true, false, or unverifiable based on available evidence.

Problem formulation¶

Given a claim and supporting/refuting evidence, generate an explanation that:

Covers key evidence: Includes the salient facts needed to understand the verdict
Is non-redundant: Avoids repeating the same point
Is non-contradictory: Does not contain conflicting statements
Supports the verdict: Logically justifies the assigned truth label

Explanations may be extractive (selecting sentences from existing sources) or abstractive (generating novel text).

Approaches¶

Extractive explanation¶

Select and order existing sentences from fact-checking rulings, evidence articles, or ruling comments that best justify a verdict. Typically uses: - Sentence ranking or scoring - ROUGE or semantic similarity metrics - Oracle selection during training

Abstractive explanation¶

Generate novel natural language text using sequence-to-sequence models: - Transformer-based architectures (DistilBERT, BART, T5) - Conditioned on claim text and evidence - Optimized with ROUGE, BERTScore, or learned loss functions

Joint multi-task learning¶

Optimize explanation generation jointly with veracity prediction: - Improves both tasks by sharing representations - Veracity prediction benefits from learning what aspects of evidence matter - Explanation generation benefits from veracity signal (what aspects are predictive)

Challenges¶

Automatic evaluation mismatch: ROUGE metrics correlate poorly with human judgments of explanation quality
Faithfulness: Generated explanations may hallucinate facts not present in source evidence
Length-quality tradeoff: Longer explanations cover more but become verbose; short explanations are concise but may omit key points
Domain specificity: Effective explanations depend on domain (legal verdicts, scientific claims, political statements require different reasoning)
Annotation effort: Creating gold-standard explanations requires expert review and careful guidelines

Fact-checking and corrections — fact-checking systems typically output a verdict; explanation generation adds the "why"
Veracity prediction — explanation generation is often paired with veracity prediction
Interpretability — explanation generation is a form of model interpretability for decision support
textual entailment — related task of determining whether evidence supports a claim

Key papers¶

Atanasova et al. (2020) — Generating Fact Checking Explanations — first study on automatically generating veracity explanations for fact-checking; jointly optimizes explanation generation with veracity prediction; shows improvements in explanation quality (coverage, non-contradiction) and veracity accuracy over extractive baselines