PaperlayerView exact revision on arXiv

Source-linked AI summary

Entailed Opinion Matters: Improving the Fact-Checking Performance of Language Models by Relying on their Entailment Ability

Gaurav Kumar, Ayush Garg, Debajyoti Mazumder, Aditya Kishore, Babu kumar, Jasabanta Patro

arXiv:2505.15050v5cs.CL

TL;DR

Fact-checking systems remain short of real-world deployment accuracy, in part because contradictory evidence confuses language models. The paper uses GLMs to classify evidence and generate justifications for training ELMs, with TBE-3 outperforming baselines on two English datasets and showing reported gains in multilingual and multimodal settings.

  • Problem

    Contradictory evidence can confuse language models, and existing automated fact-checking systems have not reached accuracy considered sufficient for real-world deployment.

    Source: §1 · p. 1

  • Method

    GLMs classify evidence as supporting or refuting and generate entailed justifications, which are used to fine-tune ELMs for veracity prediction.

    Source: §1 · p. 2

  • Results

    TBE-3 exceeded baseline macro-F1 by up to 28.57% on LIAR-RAW and 44.26% on RAW-FC, while inference-based methods performed consistently poorly.

    Source: §6 · p. 8

  • Takeaways & Limitations

    The authors conclude that TBE-3 was robust across multimodal and multilingual settings and emerged as a scalable, resource-efficient paradigm for real-world fact-checking.

    Source: §6 · p. 8

    The displayed limitation examples cover two LIAR claims, providing case-level rather than frequency evidence.

    Source: F Limitations · p. 20

Abstract

from arXiv · show

Automated fact-checking has been a challenging task for the research community. Prior work has explored various strategies, such as end-to-end training, retrieval-augmented generation, and prompt engineering, to build robust fact-checking systems. However, their accuracy has not been high enough for real-world deployment. We, on the other hand, propose a new learning paradigm, where evidence classification and entailed justifications made by generative language models (GLMs) are used to train encoder-only language models (ELMs). We conducted a rigorous set of experiments, comparing our approach with recent works along with various prompting and fine-tuning strategies. Additionally, we performed ablation studies, error analysis, quality analysis of model explanations, and a domain generalisation study to provide a comprehensive understanding of our approach.

1 Introduction:

The introduction frames fact-checking as difficult because contradictory evidence confuses language models and existing systems are not accurate enough for real-world deployment. It proposes using GLM-generated entailment classifications and justifications to train ELM veracity predictors, and reports gains across English, multilingual, and multimodal evaluations.

1 Introduction:

  • The proposed pipeline uses GLMs to classify evidence as supporting or refuting, consolidate each group into justifications, and train ELMs to predict veracity.Source: Figure 1 · p. 1, §1 · p. 2
  • TBE-3 exceeded baseline macro-F1 by up to 28.57% on LIAR-RAW and 44.26% on RAW-FC; TBE-2 improved RAW-FC by up to 16.39%.The introduction also reports up to 8.20% improvement on RAW-FC for TBE-1.Source: §1 · p. 2
  • Removing either entailed justification led to lower MF performance, indicating that both supporting and refuting justifications matter.Source: §1 · p. 2
  • Llama-generated explanations received the highest subjective scores, correlating with the highest veracity prediction performance; attention was more scattered on borderline samples.Attention focused more on appropriate keywords and factual elements for extreme labels such as true or false.Source: §1 · p. 2
  • TBE-3 improved MF1 by up to 19.70% in multilingual and 54.10% in multimodal domains.The study used X-Fact and Ru22fact for multilingual evaluation, and Factify-2, MOCHEG, and VERITE for multimodal evaluation.Source: §1 · p. 2

2 Related Works:

Related work covers changing modeling approaches and efforts to handle conflicting evidence. The paper identifies explicit entailment-based evidence discrimination as a gap in prior fact-checking pipelines.

2 Related Works:

  • Prior work spans manual feature engineering, fine-tuned ELMs, and GLMs, while language models can struggle with bias and hallucination when evidence is noisy or contradictory.Source: §2 · p. 2
  • Yue et al. (2024) generated supporting and opposing arguments but did not explicitly divide contradictory evidence; L-Defense (Wang et al., 2024a) filtered or weighted evidence.The authors note that attention or relevance scoring may miss useful information when evidence conflicts or only top-k items are retained.Source: §2 · p. 3
  • Prior entailment work used claim-evidence matching or similarity scoring, whereas this paper proposes GLM entailment to classify each evidence item as supporting or refuting before generating justifications.Source: §2 · p. 3

3 Dataset Details:

The study uses monolingual English datasets for its training- and inference-based experiments, and multilingual and multimodal datasets to assess domain generalisability.

3 Dataset Details:

  • Table 1 reports dataset statistics and distinguishes monolingual evaluation datasets from those used in the domain generalisation study.Source: Table 1 · p. 3
  • LIAR-RAW and RAW-FC are the monolingual datasets used to compare training-based and inference-based experiments.Source: §3 · p. 3
  • X-FACT and RU22Fact are the multilingual datasets, while Factify-2, MOCHEG, and VERITE are the multimodal datasets used for domain generalisability.Source: §3 · p. 3

4 Experiments:

The experiments compare training-based and inference-based approaches across raw evidence, overall claim-evidence understanding, and entailment-based justifications. They also compare against previously proposed baselines.

4 Experiments:

  • The study includes three training-based and four inference-based experiments, fine-tunes ELMs and adapter-equipped GLMs, and prompts GLMs for inference-based prediction.GLMs could not be fine-tuned directly without adapters because of their parameter size and computational constraints.Source: §4 · p. 3

4.1.1 TBE-1: Training based on raw-evidences:

  • TBE-1 fine-tunes GLMs with LoRA or LoRA+ using claims and raw evidence, or only the claim when no evidence sentence is available.ELMs were excluded because many input instances exceeded their maximum supported length.Source: §4.1.1 · p. 4

4.1.2 TBE-2: Training based on overall understanding:

  • TBE-2 prompts GLMs to produce claim-evidence understanding, then fine-tunes ELMs and adapter-equipped GLMs to predict veracity.Source: §4.1.2 · p. 4

4.1.3 TBE-3: Training based on entailment understanding:

  • TBE-3 has GLMs classify evidence as supporting or refuting, generate justifications for each group, and use those justifications with the claims to fine-tune veracity predictors.When a claim lacks supporting or refuting evidence, GLMs generate justifications from embedded knowledge.Source: §4.1.3 · p. 4

4.2 Inference Based Experiments (IBEs):

  • The four inference-based experiments range from direct zero-shot prediction to prompting with overall understanding, CoT reasoning, or entailment-based justifications.For claims without associated evidence, the claim text alone was used as input.Source: §4.2 · p. 4

4.3 Baselines:

  • The baseline comparison includes HiSS, FactLLaMa, RAFTS, and L-Defense; their evidence sources differ, and Table 2 reports previously reported LIAR-RAW and RAW-FC performance.HiSS and FactLLaMa retrieve external evidence, RAFTS uses dataset evidence, and L-Defense uses relevant evidence without additional retrieval.Source: §4.3 · pp. 4–5, Table 2 · p. 5

5 Results and Discussion:

TBE-3, which trains on GLM-generated entailed justifications, produced the strongest and most consistent baseline gains on LIAR-RAW and RAW-FC. Results also varied across inference methods, justification components, evidence volume, and multilingual and multimodal settings.

5.1 Observations for veracity prediction:

  • No LIAR-RAW IBE score surpassed the best baseline, whereas Llama in IBE-2 surpassed the best baseline on RAW-FC.Source: §5.1 · p. 5
  • RAW-FC TBE-1 reached MF1 0.66 (∼8.20% ↑), and TBE-2 reached ∼16.39% ↑; LIAR-RAW showed no baseline-beating TBE-2 result and clustered at MF1 0.22–0.30 in TBE-1.The best RAW-FC TBE-1 result was Qwen+LoRA; on LIAR-RAW, neither LoRA nor LoRA+ reliably helped.Source: §5.1 · pp. 5–6
  • 28.57% on LIAR-RAW and 44.26% on RAW-FC were the reported maximum TBE-3 MF1 gains over baselines.For RAW-FC, RoBERTa with Llama-based justifications achieved MF1 0.88, ∼44.26% ↑.Source: §5.1 · p. 6

5.2 Observations from ablation study:

  • Removing supporting justifications reduced MF1 more than removing refuting ones (31.48% vs 9.26% on LIAR-RAW; 12.50% vs 9.09% on RAW-FC), while removing both caused the largest drops.With both justification types removed, MF1 decreased by 55.56% on LIAR-RAW and 47.73% on RAW-FC.Source: §5.2 · p. 6
  • LIAR-RAW peaked at MF1 0.61 with 6–20 evidence pieces, then fell 29.50% to 0.43 above 50 pieces; RAW-FC peaked at MF1 0.94 with 11–20 and fell ∼7.45% to 0.87 above 50.The authors describe LIAR-RAW as sensitive to evidence overload and RAW-FC as comparatively robust to larger evidence sets.Source: §5.2 · p. 7

5.3 Observations from domain generalisation study:

  • In multimodal settings, QLoRA-tuned GLMs achieved the best results, with MF1 gains up to 36.8% on Factify-2, 54.1% on MOCHEG, and 36.2% on VERITE.ELMs also improved MF1 by up to 33.3%, 37.8%, and 14.9% on those datasets, respectively.Source: §5.3 · p. 7, Table 8 · p. 8
  • Multilingual ELM gains were limited, especially on X-Fact, while QLoRA-tuned GLMs improved MF1 by 9.5% on X-Fact and 19.7% on RU22Fact.The authors attribute the X-Fact difficulty mainly to its 25 languages and seven-class veracity labels.Source: §5.3 · p. 7

6 Conclusion:

The conclusion emphasizes the gains from GLM-entailed justifications in training, alongside evidence that explanation quality, justification components, and domain coverage matter. It also notes that attention can be scattered for borderline veracity labels.

6 Conclusion:

  • 28.57% on LIAR-RAW and 44.26% on RAW-FC were the maximum baseline MF1 gains from training with GLM-entailed justifications.Claim-evidence understanding was second-best, increasing RAW-FC MF1 by up to 16.39%; inference-based methods consistently performed poorly.Source: §6 · p. 8
  • Llama-generated explanations received higher subjective ratings for informativeness, readability, and logical structure, aligning with strong XLNet and RoBERTa TBE-3 performance.Source: §6 · p. 8
  • Removing either supporting or refuting justifications adversely affected performance in the ablation study.Source: §6 · p. 8
  • ELMs attended more accurately to factual words and support-or-contradiction signals, but attention appeared scattered on borderline-veracity samples.Source: §6 · p. 8
  • The authors report TBE-3 as robust across multimodal and multilingual settings and characterize it as scalable and resource-efficient for real-world fact-checking.Source: §6 · p. 8

A Additional dataset details

The appendix describes the monolingual, multilingual, and multimodal datasets used for evaluation, along with evidence preparation and dataset-specific evaluation choices. It also reports limits on reproducing two baselines.

A Additional dataset details

  • LIAR-RAW uses six veracity labels and RAW-FC uses three; both are open-sourced under Apache 2.0 and use the splits provided by Yang et al. (2022).The authors selected them because they have been widely used in prior work.Source: A Additional dataset details · p. 12
  • For X-FACT and RU22Fact, the authors crawled claim-associated webpages, segmented their content, and retrieved a top claim-matching chunk from each document.X-FACT covers 25 languages, while RU22Fact concerns Russia–Ukraine conflict claims in four languages.Source: A Additional dataset details · p. 12, Table 10 · p. 12
  • Factify-2, MOCHEG, and VERITE provide image–text pairs; VERITE is designed to reduce unimodal bias through joint visual-textual reasoning.The datasets differ in label granularity and reasoning complexity.Source: A Additional dataset details · p. 12, Table 10 · p. 12
  • Because Factify-2’s test set was unavailable, the authors sampled training instances for validation and used the official validation set as test; VERITE retained its original three-way split.Source: A Additional dataset details · p. 12
  • The authors could not reproduce RU22Fact or MOCHEG baselines because official code was unavailable for the former and MOCHEG required approximately 37 GPU hours per epoch.They reproduced Factify-2 with a reduced maximum sequence length of 128 tokens to fit GPU memory.Source: A Additional dataset details · p. 12

B Experimental set-up:

The experimental setup uses macro-F1 to evaluate model performance and combines lexical, semantic, and subjective assessments of generated explanations. It also reports a hyperparameter search, a fixed GLM temperature, and the hardware used.

B.1 Evaluation metrics:

  • Macro-F1 (MF1) was used to evaluate the models.Source: §B.1 · p. 12
  • Explanation quality was assessed through lexical overlap, semantic matching, and subjective ratings by decoder-only GLMs across five dimensions.The evaluation used ROUGE-1, ROUGE-2, ROUGE-L, BLEU, and BERT score, alongside ratings of informativeness, logicality, objectivity, readability, and accuracy.Source: §B.1 · pp. 12–13

B.2 Hyperparameter details:

  • After an extensive hyperparameter search, GLM temperature was fixed at 0.001, and experiments ran on an NVIDIA A100 80GB GPU.Source: §B.2 · p. 13

C Observations from model explainability:

Explanation evaluations varied by metric, while Llama explanations received the most balanced subjective ratings and their use coincided with the highest veracity-prediction MF1 for RoBERTa and XLNet.

C Observations from model explainability:

D Linguistic insights:

Attention patterns differed by claim veracity: models focused on supporting evidence for true claims, conflicting cues for half-true claims, and refutations for false claims.

D Linguistic insights:

  • For true claims, attention concentrated on affirmative and evidential phrases in supporting justifications, while refuting justifications received minimal focus.Examples include “supported by multiple sources” and “credible evidence.”Source: D Linguistic insights · p. 14
  • For half-true claims, attention spread across conflicting linguistic signals, reflecting sensitivity to mixed evidence and ambiguity.Examples include “accurate” versus “inconsistencies undermine.”Source: D Linguistic insights · p. 14
  • For false claims, models emphasized refuting terms and often ignored weak or unsupported supporting justifications.Examples include “false,” “disputes the claim,” and “outdated story.”Source: D Linguistic insights · p. 14
  • Overall, models prioritized salient factual signals, but dispersed attention across conflicting narratives remained a challenge for borderline cases such as half-true claims.Source: D Linguistic insights · p. 14

E Human Evaluation:

Human annotators evaluated Llama explanations from both datasets on five dimensions; agreement ranged from poor to moderate on LIAR-RAW and was lower overall on RAW-FC.

E Human Evaluation:

  • Three independent annotators rated Llama explanations from 40 randomly selected LIAR-RAW and RAW-FC samples on five dimensions using a 1–5 Likert scale.The dimensions were informativeness, accuracy, readability, objectivity, and logicality.Source: E Human Evaluation · p. 14
  • On LIAR-RAW, inter-annotator agreement ranged from poor to moderate (κ ≈0.12–0.46, α up to 0.54).Agreement was especially consistent for readability and logicality.Source: E Human Evaluation · p. 14
  • On RAW-FC, agreement was lower overall (κ ≈0.076–0.21, α from 0.06 to 0.20), reflecting greater subjectivity in judgments.Source: E Human Evaluation · p. 14

F Limitations:

The experiments were limited to open-source models and a closed-domain fact-checking setup, with quantized models used for domain-generalization experiments.

F Limitations:

  • Experiments used open-source language models for reproducibility and resource efficiency, limiting the evaluated model scope.The authors identify commercial GLMs as a possible direction for future generalization checks.Source: F Limitations · p. 14
  • Domain-generalization experiments used quantized models to expedite evaluation, leaving full-precision models and full GLM fine-tuning for future study.Source: F Limitations · p. 14
  • The study assumed closed-domain fact-checking; generalization to open-domain settings with evidence retrieved externally remains for future work.The authors also point to larger fact-checking datasets as future evaluation targets.Source: F Limitations · p. 14

Related papers

Loading 2505.15050v5…