From Fragments to Facts: A Curriculum-Driven DPO Approach for Generating Hindi News Veracity Explanations

1. Problem Statement

The paper addresses the critical challenge of misinformation detection and explanation generation in Hindi, a low-resource language with over 600 million speakers. While high-resource languages like English have advanced automated fact-checking tools, Hindi lacks robust, scalable solutions. Existing Large Language Models (LLMs) struggle with:

• Factual Accuracy: Generating explanations that are factually grounded rather than hallucinated.
• Contextual Relevance: Providing coherent reasoning specific to Hindi news nuances.
• Alignment: Producing explanations that match human reasoning patterns, particularly the distinction between "informative" true news explanations and "argumentative" fake news explanations.

Current methods often fail to scale due to a lack of high-quality, fact-checked datasets and the inability of standard models to distinguish subtle differences between reliable and unreliable explanations.

2. Methodology: The DeFactoX Framework

The authors propose DeFactoX, a unified framework that integrates Curriculum Learning with an enhanced Direct Preference Optimization (DPO) approach. The methodology consists of four key stages:

A. Synthetic Preference Dataset Creation

To overcome data scarcity, the authors constructed a synthetic ranking-based dataset:

• Preferred Responses ($y_w$): Human-written explanations from credible fact-checking websites (e.g., OneIndia, Vishvas News). The authors noted that human explanations for true news were often purely informational, while fake news explanations were explicit and argumentative. They standardized true news explanations using prompts to ensure they explicitly affirmed veracity, creating a uniform structure.
• Rejected Responses ($y_l$): Explanations generated by state-of-the-art LLMs (GPT-4o-mini, Mistral-7B, Gemini-1.5-Flash) using simple prompts without fine-tuning. These were intentionally allowed to exhibit common LLM weaknesses (superficial reasoning, bias, lack of depth) to serve as negative training examples.
• Ratio: A balanced 1:3 ratio of preferred to rejected explanations.

B. Curriculum Learning Strategy

The framework employs a curriculum learning approach to progressively train the model:

• Scoring Function ($f_s$): Rejected explanations are scored based on alignment with ground truth using a weighted combination of BERTScore, ROUGE-L, and METEOR.
• Ranking: Explanations are bucketed into three difficulty levels:
  • Rank-0 (Easy): Clearly distinct from human reasoning.
  • Rank-1 (Moderate): Partially aligned.
  • Rank-2 (Hard): Subtly different, requiring fine-grained distinction.
• Training Flow: The model is trained sequentially from Rank-0 to Rank-2, allowing it to learn basic distinctions before tackling complex nuances.

C. Hin-DPO: Enhanced Loss Function

The core innovation is Hin-DPO, a modified DPO loss function that incorporates two novel parameters to address factual reliability and stability:

1. Actuality ($s$): A score quantifying factual correctness.
   • Calculation: Uses GPT-4o-mini with web search tools to extract factual statements from the explanation and verify them against external knowledge.
   • Role: Acts as a reward signal. The loss function amplifies the probability of preferred responses with high Actuality scores ($1 + s_w$) and penalizes rejected responses with low scores.
2. Finesse ($v$): A score quantifying hallucination via instability.
   • Calculation: Measures the variance of the model's output probability distributions across multiple stochastic generations (5 runs) of the same input. High variance indicates uncertainty and potential hallucination.
   • Role: Acts as a scaling factor in the denominator. Explanations with low variance (high stability) receive higher preference weights, while unstable outputs are down-weighted.

The Loss Function:
$$\mathcal{L}_{Hin-DPO} = -\mathbb{E}_{(x, y_w, y_l) \sim \mathcal{D}} \left[ \log \sigma \left( \frac{\beta}{v + \epsilon} \left[ (1 + s_w) \log \frac{\pi_\theta(y_w|x)}{\pi_{ref}(y_w|x)} - \max(0.01, s_l) \log \frac{\pi_\theta(y_l|x)}{\pi_{ref}(y_l|x)} \right] \right) \right]$$

3. Key Contributions

1. DeFactoX Framework: The first framework specifically designed for automated, veracity-driven explanation generation in Hindi.
2. Hin-DPO Loss Function: Introduces Actuality and Finesse parameters to DPO, explicitly optimizing for factual accuracy and generation stability (reducing hallucinations).
3. Curriculum-Driven Training: Demonstrates the efficacy of progressively training models on explanation difficulty levels to improve generalization.
4. Synthetic Preference Dataset: Created a high-quality, ranking-based dataset where human fact-checked explanations serve as gold standards, and LLM outputs serve as negative examples, bridging the data gap for Hindi.

4. Experimental Results

The framework was evaluated on five models: mBART, mT5, Gemma-2-9B, Llama-3.1-8B, and Mistral-7B.

• Automatic Metrics: Hin-DPO consistently outperformed baselines (Base, Base+SFT, standard DPO) across ROUGE-1/2/L, METEOR, and BERTScore.
  • Example: On Llama-3.1-8B, Hin-DPO achieved a BERTScore of 84.73 compared to 80.98 for standard DPO.
  • Veracity Prediction: Hin-DPO also improved classification accuracy, reaching 81.2% on Llama-3.1-8B.
• Human Evaluation: 800 explanations were rated by human annotators on a 0–5 scale (Factual Accuracy, Rationale Alignment, Completeness, Expression).
  • Hin-DPO achieved the highest scores (4.23 for Llama-3.1-8B and 4.12 for Gemma-2-9B), significantly outperforming standard DPO and SFT baselines.
• Ablation Study: Removing Curriculum Learning or individual components (Actuality/Finesse) resulted in performance drops, confirming that all components are necessary for optimal results.

5. Significance and Impact

• Bridging the Language Gap: This work provides a scalable methodology for low-resource languages, demonstrating that synthetic data combined with advanced alignment techniques can overcome data scarcity.
• Trustworthy AI: By explicitly penalizing hallucinations (via Finesse) and rewarding factual correctness (via Actuality), the framework generates explanations that are not just fluent but trustworthy.
• Practical Application: The system can be deployed by media houses and fact-checkers to detect misleading narratives and provide end-users with contextually accurate, human-aligned explanations, thereby mitigating the societal impact of misinformation in Hindi-speaking regions.
• Future Adaptability: The data augmentation and curriculum-driven design offer a blueprint for adapting similar systems to other underrepresented languages.

Limitations

• Data Dependency: Reliance on the availability of high-quality fact-checked Hindi data limits domain coverage.
• Computational Cost: Calculating the Finesse score requires multiple generations per input, increasing inference time.
• Model Size: Experiments were limited to models under 10B parameters due to computational constraints.
• External Bias: The Actuality score relies on external LLMs (GPT-4o), which may introduce their own biases or inaccuracies.