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

TL;DR

This work tackles the challenge of generating reliable Hindi news explanations by introducing DeFactoX, a two-stage framework that couples veracity prediction with explanation generation. It advances a curriculum-based training paradigm and a novel Hin-DPO objective that leverages Actuality (factual correctness) and Finesse (output stability) to align model explanations with human reasoning. A synthetic, ranking-based Hindi preference dataset grounds the alignment in human-like explanations, while experiments across multiple LLMs and PLMs demonstrate improved semantic quality and veracity alignment over strong baselines. The approach offers a scalable path to automated, trustworthy explanations for Hindi misinformation, with potential extension to other low-resource languages through multilingual transfer and human-in-the-loop feedback.

Abstract

In an era of rampant misinformation, generating reliable news explanations is vital, especially for under-represented languages like Hindi. Lacking robust automated tools, Hindi faces challenges in scaling misinformation detection. To bridge this gap, we propose a novel framework integrating Direct Preference Optimization (DPO) with curriculum learning to align machine-generated explanations with human reasoning. Fact-checked explanations from credible sources serve as preferred responses, while LLM outputs highlight system limitations and serve as non-preferred responses. To refine task-specific alignment, we introduce two key parameters -- Actuality and Finesse -- into the DPO loss function, enhancing explanation quality and consistency. Experiments with LLMs (Mistral, Llama, Gemma) and PLMs (mBART, mT5) confirm the framework's effectiveness in generating coherent, contextually relevant explanations. This scalable approach combats misinformation and extends automated explanation generation to low-resource languages.

Figures (5)

• Figure 1: Overview of DeFactoX framework. (Left) Dataset creation with human-written explanations as preferred responses and LLM-generated explanations as rejected responses. (Center) Curriculum-based dataset construction, where samples are ranked and bucketed into easy, moderate, and hard levels. (Right) Model training with Hin-DPO under curriculum learning, fine-tuning the model to generate human aligned explanations.
• Figure 2: Snippet of fake news explanation with explicit reasoning for its veracity.
• Figure 3: Example of a human-written true news explanation that is primarily informational, summarizing facts without explicitly confirming authenticity.
• Figure 4: Snippet of True news transformation.
• Figure 5: Snippet of Non-Preferred Response Generation.