Imagine scrolling through your phone and seeing a forwarded message in your native language—a claim about a politician, a medicine, or a community—that sounds alarming. You want to know if it is true, but the fact-checking tools you trust are built mainly for English. This is the reality for hundreds of millions of people around the world, and it is particularly acute in India, a country where social media has exploded far faster than the systems designed to police it. Fake news does not respect borders, and it certainly does not respect languages. While major tech companies have poured resources into automated detection for English-language content, the vast majority of the world’s languages remain badly underserved. A new study published in Discover Artificial Intelligence by Sushama Nandgaonkar and Sunil Mane of COEP Technological University in Pune directly confronts this gap. They developed a Hierarchical Attention Network, or HAN, that detects false news across English, Hindi, and Marathi with extraordinary accuracy—98.9 percent overall accuracy and an F1-score of 98.8 percent on a combined multilingual test set. It is a powerful reminder that the tools we build to defend truth must speak the languages of the people they are meant to protect.
The stakes could hardly be higher. India has the second-largest internet-using population on Earth, and the study cites striking numbers: more than 375 million social media users in 2019, rising to 518 million in 2020, with projections suggesting the total could reach 1.5 billion by 2040. Affordable mobile data has fueled this growth, and platforms like Facebook, Twitter, and WhatsApp now carry content in dozens of regional languages. Misinformation circulating in Hindi or Marathi can do more than mislead—it can inflame community tensions, distort political discourse, and shape life-or-death health decisions. During the COVID-19 pandemic, false claims about miracle cures, vaccine dangers, and conspiracy theories spread through WhatsApp forwards faster than official corrections could possibly catch up. The researchers carefully distinguish between related concepts that are often tangled together: disinformation is deliberately created and spread to deceive; rumors are unverified pieces of information circulated with the intent to mislead; and misinformation is simply false content shared without necessarily intending harm. The motives range from religious hatred and political advantage to financial gain and public confusion. Fact-checking websites, while valuable, can carry their own biases and can never scale to billions of posts. Human moderation cannot possibly review every suspicious message. This is why automated detection systems have become an essential line of defense—and why, as the study argues, those systems must operate across the languages people actually use, not just the ones favored by Silicon Valley algorithms.
At the heart of the new system lies an architecture that imitates the way human readers process a document. The model does not scan a headline in isolation; it reads the way we read, moving carefully from one sentence to the next while holding the overall context in mind. First, a Bidirectional Long Short-Term Memory network processes each sentence word by word, reading in both forward and backward directions so that every word is understood in relation to what came before and what comes after. Then, a custom attention layer assigns dynamic weights to individual words, essentially learning during training which words matter most for separating genuine news from false stories. Some words are emotionally charged; others are deceptively neutral. The model learns to weigh them accordingly. These weighted word representations are then combined into a single vector for each sentence. A second Bi-LSTM layer, with its own attention mechanism, models the relationships between sentences, producing a document-level representation that finally feeds into a classifier which makes the true-or-false decision. This hierarchical design is not merely an engineering convenience; it mirrors the reality that fake news articles often contain misleading patterns, emotional polarization, and contextual inconsistencies scattered across multiple sentences rather than concentrated in any single phrase. By attending to both words and sentences, the model captures local semantic clues and the broader narrative structure at the same time. And because attention weights are visible and interpretable, researchers can see exactly which parts of an article drove a classification decision—a significant advantage over opaque black-box models that offer no explanation for their verdicts.
A crucial part of this study is less about algorithms and more about the labor-intensive work of creating trustworthy data. When the researchers began, no publicly available fake news dataset existed for Marathi, one of the most widely spoken languages in western India. So they built one from scratch. The team collected 1,957 news articles from Marathi outlets including Loksatta, Lokmat, Maharashtra Times, and Zee News—707 labeled as fake and 1,250 labeled as true. Every article was manually reviewed, and labels were assigned based on source credibility, fact-checking reports, and consistency across multiple platforms. For Hindi, the researchers combined two existing datasets from Kaggle and GitHub, while English experiments used the widely adopted ISOT dataset. Yet raw datasets are rarely enough. The Marathi and Hindi data suffered from severe class imbalance, with far fewer fake examples than genuine ones, which would bias any model toward simply saying “true.” To address this, the team applied translation-based augmentation using IndicTrans, a neural translation framework, generating additional training samples from the under-represented classes. Crucially, they left the test set untouched to prevent data leakage and ensure honest results. The final augmented corpus contained 45,386 English samples, 10,481 Hindi samples, and 5,000 Marathi samples. The choice of word embeddings also proved important. FastText, developed by Facebook’s AI Research lab, represents each word as a bag of character n-grams, allowing it to capture morphological information and generate meaningful vectors even for words it has never seen, simply from their sub-word pieces. This is especially valuable for morphologically rich languages like Hindi and Marathi, where word forms can vary dramatically. The researchers compared FastText against GloVe and random initialization under identical conditions, and found that pretrained embeddings substantially improved performance. FastText also converged faster, completing training in about 383 seconds compared with 501 seconds for GloVe.
The experimental results are striking. The HAN model achieved 99.42 percent accuracy on English news, 99.04 percent on Hindi, and 80.90 percent on Marathi. It outperformed traditional machine learning baselines such as Logistic Regression, Linear Support Vector Machines, and Random Forest, as well as deep learning models like CNN and standalone Bi-LSTM. The comparison with multilingual transformers, however, is especially telling. Multilingual BERT, or mBERT, reached 98.72 percent overall accuracy, and DeBERTa-v3-base reached 98.66 percent—both slightly below the proposed model’s 98.90 percent. Yet DeBERTa’s performance collapsed to just 74.42 percent on Marathi, revealing how sensitive large pretrained transformers are to data imbalance and low-resource languages. Even more dramatic was the evaluation of LLaMA 3.1, an 8-billion-parameter instruction-tuned model, which was tested through prompting rather than fine-tuning. In a zero-shot setting, LLaMA managed only 67.98 percent accuracy; with six examples in context, it rose to 84.76 percent, but still produced 1,638 failed predictions out of 10,965 test instances. This matters because many people assume that massive generative models are inherently good at everything. The study shows that a smaller, carefully designed hierarchical attention network with robust sub-word embeddings and thoughtful data augmentation can beat much larger models on multilingual misinformation detection. Ablation experiments confirmed that both levels of attention contribute meaningfully. A Bi-LSTM-only model achieved an F1-score of 97.43 percent. Adding word-level attention alone raised it to 97.82 percent, while adding sentence-level attention alone lifted it to 98.17 percent. The complete hierarchical architecture reached 98.72 percent. Interestingly, sentence-level attention proved more influential than word-level attention in isolation, suggesting that contextual dependencies across sentences play a decisive role in identifying deceptive content. Paired t-tests across multiple random seeds showed all improvements were statistically significant, with p-values below 0.01.
No study is without limitations, and the authors are refreshingly honest about theirs. Marathi’s error rate of 19.10 percent—compared with just 0.58 percent for English and 0.96 percent for Hindi—reflects the enormous challenge of working with low-resource languages. The most troubling failures involved fake articles written in a style so similar to legitimate journalism that the model made incorrect predictions with high confidence, a sobering reminder that even the best system can be fooled by deliberate mimicry. Generalization to other Indian languages remains unexplored, and multimodal signals such as images, videos, and metadata have not yet been incorporated. Future work will expand the dataset to additional regional languages and integrate textual and visual information using transformer-based Indic language models. But the broader takeaway is hopeful. This research demonstrates that with careful attention mechanisms, sub-word embeddings, and thoughtful data augmentation, state-of-the-art misinformation detection can be brought to languages that have long been ignored. It is not merely an academic exercise. In a world where a single forwarded message can change an election, spark violence, or cost a life, building systems that understand every language is one of the most important technological missions of our time. The study does not solve the problem overnight, but it lights a path forward—one sentence, one word, one language at a time.