Misinformation detection has predominantly been studied in monolingual or language-aligned settings, where claims and evidences are assumed to be written in the same language. However, this assumption does not reflect real-world information ecosystems, in which misinformation frequently spans multiple languages and verification requires cross-lingual reasoning over heterogeneous evidence sources. As a result, cross-lingual claim-evidence reasoning remains largely unexplored and has not been systematically studied, and existing methods are not explicitly designed to support such scenarios. To address this gap, we introduce a novel setting for multilingual misinformation detection with cross-lingual evidences and propose GraDuLa, a graph-based dual-branch language-agnostic modeling for misinformation detection. GraDuLa projects evidential information into a language-agnostic subspace, disentangling semantic representations from language-specific associations through dual-branch graph modeling. For evaluation, we construct and release a benchmark GraDuLaEval and compare our approach against strong baseline methods. We further plan to release the dual-branch graph representations as an additional resource for future research.

Conventional prompting and prompt-based finetuning offer only weak semantic structure, leaving models to derive task-specific reasoning from limited cues. Existing techniques, such as using demonstrations, chain-of-thought prompting, or retrieval augmentation, can improve performance, but often at the cost of longer, more redundant prompts, higher inference overhead, and reliance on external knowledge sources. We present ConceptWeave, a novel concept-centric prompt-based finetuning approach built on adaptive semantic infilling. Instead of depending on external resources, it extracts concise, task-relevant concepts directly from a small portion of downstream data and organizes them into a dynamically evolving hierarchical memory. During fine-tuning, ConceptWeave identifies informative spans in the input through semantic alignment and augments them with corresponding concept representations, enabling fine-grained, localized conditioning of the model. As training progresses, the concepts are continuously refined and structured into a hierarchy that supports efficient retrieval while preserving semantic richness and scalability. We validate ConceptWeave across a range of NLP tasks, including misinformation detection, question answering, sentiment analysis, and hate speech classification, where it consistently outperforms strong baselines and recent state-of-the-art methods.

Lexicon or dictionary generation across domains has the potential for societal impact, as it can potentially enhance information accessibility for a diverse user base while preserving language identity. Prior work in the field primarily focuses on bilingual lexical induction, which deals with word alignments using mapping-based or corpora-based approaches. However, these approaches do not cater to domain-specific lexicon generation that consists of domain-specific terminology. This task becomes particularly important in specialized medical, engineering, and other technical domains, owing to the highly infrequent usage of the terms and scarcity of data involving domain-specific terms especially for low-resource languages. We propose a new model to generate dictionary words for 6 Indian languages in the multi-domain setting. Our model consists of domain-specific and domain-generic layers that encode information, and these layers are invoked via a learnable routing technique. We also release a new benchmark dataset consisting of >75K translation pairs across 6 Indian languages spanning 8 diverse domains. We conduct both zero-shot and few-shot experiments across multiple domains to show the efficacy of our proposed model in generalizing to unseen domains and unseen languages. Additionally, we also perform a human post-hoc evaluation on unseen languages. The source code and dataset is present at https://github.com/Atulkmrsingh/lexgen.

We release Saamayik, a dataset of around 53,000 parallel English-Sanskrit sentences, written in contemporary prose. Sanskrit is a classical language still in sustenance and has a rich documented heritage. However, due to the limited availability of digitized content, it still remains a low-resource language. Existing Sanskrit corpora, whether monolingual or bilingual, have predominantly focused on poetry and offer limited coverage of contemporary written materials. Saamayik is curated from a diverse range of domains, including language instruction material, textual teaching pedagogy, and online tutorials, among others. It stands out as a unique resource that specifically caters to the contemporary usage of Sanskrit, with a primary emphasis on prose writing. Translation models trained on our dataset demonstrate statistically significant improvements when translating out-of-domain contemporary corpora, outperforming models trained on older classical-era poetry datasets. Finally, we also release benchmark models by adapting four multilingual pre-trained models, three of them have not been previously exposed to Sanskrit for translating between English and Sanskrit while one of them is multi-lingual pre-trained translation model including English and Sanskrit. The dataset and source code can be found at https://github.com/ayushbits/saamayik.