Greek, Urdu, Bangla, Korean, Sinhala, Catalan, and SEA Languages: Unlocking LLM Potential in Low-Resource Contexts

Latest 11 papers on low-resource languages: Sep. 13, 2026

The world of AI and Machine Learning is rapidly evolving, yet a significant challenge persists: bringing the power of Large Language Models (LLMs) to the vast array of low-resource languages. These languages, spoken by billions, often lack the digital data infrastructure and dedicated research focus enjoyed by high-resource languages like English. This disparity leads to an AI divide, where cutting-edge applications remain inaccessible. Fortunately, recent research is pushing the boundaries, offering novel approaches to empower LLMs for diverse linguistic and cultural contexts. Let’s dive into some of the latest breakthroughs.

The Big Idea(s) & Core Innovations

The central theme across these papers is the innovative adaptation and evaluation of LLMs to overcome the unique challenges posed by low-resource languages. Researchers are tackling issues from accurate speech transcription in singing to cultural nuances in story generation and robust text simplification.

For instance, the paper “Automatic Lyric Transcription for Greek Songs: Scaling and Task Composition Effects in Whisper Adaptation” by Maria Frangiadaki and her colleagues from the Institute for Language and Speech Processing, Athena R.C., Greece, introduces the first systematic benchmark for Greek Automatic Lyric Transcription (ALT). Their key insight is that while multitask learning (transcription + translation) helps smaller Whisper models, larger models benefit more from focused, transcription-only training and a two-stage speech-to-singing adaptation. This approach, which also leverages vocal source separation, drastically improves performance on singing voices, a notoriously difficult domain due to melismatic stretching and rhythmic irregularities.

Meanwhile, the complexities of cultural and linguistic alignment are starkly highlighted in “Multilingual in Name Only? Cultural and Linguistic Weaknesses of LLMs in Urdu” by Farah Adeeba et al. from the University of Konstanz, Germany, among others. Their work reveals that despite surface-level fluency, LLMs often fail on Urdu-specific grammatical features (like ergative case) and commit glaring cultural anachronisms (e.g., Muslim characters cremated). This demonstrates that English-mediated representations lead to fundamental misunderstandings that simple prompting cannot fix, and that cultural knowledge is critical for genuine multilingual competence.

Bridging the gap for Bangla dialects, Md Mahir Jawad and his team from BRAC University introduce “5-DIALECTS-BN: Unmasking the Impact of Transliteration on Bangla Dialectal LLMs”. A crucial finding here is that Romanized transliteration severely degrades modern LLM performance, an irreversible information loss that contrasts with earlier assumptions. Their research champions LoRA fine-tuning with minimal examples (just 160 per dialect), showing that targeted adaptation far surpasses general models and even sophisticated prompting techniques.

For Korean consultation services, “Leveraging Fine-grained Error Correction in Korean Speech Recognition for Consultation Services” by Yonghyun Jun et al. from Chung-Ang University addresses the challenge of error-sparse ASR transcripts. They propose DCSC, a two-stage detector-corrector framework that uses token-level detection and dialogue context to achieve state-of-the-art error correction, significantly improving accuracy where general LLMs fall short.

The historical evolution of language is explored in “Dynamics of meaning: Towards the Evaluation of Diachronic Semantic Change in Sinhala” by Nevidu Jayatilleke and Nisansa de Silva from the University of Moratuwa, Sri Lanka. Their work on Sinhala introduces a novel Bidirectional Semantic Impact Pruning approach, demonstrating that semantic drift is driven by selective contextual innovations rather than uniform gradual shifts. This provides a preliminary framework for diachronic analysis in low-resource settings, showing how fine-tuned FastText embeddings outperform Word2Vec.

Catalan text simplification gets a boost from “Reinforcement Learning for improving Large Language Models’ Catalan text simplification capabilities” by Arnau Ayguadé Domingo et al. from Universitat Pompeu Fabra. They successfully apply Reinforcement Learning (GRPO) with a novel reward function, showing that post-training on English datasets can cross-lingually transfer simplification abilities to Catalan, an elegant solution for data scarcity.

The overarching challenge of multilingual reliability is tackled by Ekata Mitra and Ameeta Agrawal from PortNLP, Portland State University, USA in “Choosing the Right Language Mode at Inference Time for Multilingual Reliability”. They introduce RAAI, a training-free adaptive inference framework. Their key insight is that English context helps understanding, but bilingual redundancy can cause interference; adaptively selecting language assistance based on reliability signals yields substantial accuracy gains (25-37.7%) for low-resource languages.

Finally, the critical area of fact-checking in social media for low-resource languages is addressed in “Multilingual Models for Check-Worthy Social Media Posts Detection” by Sebastian Kula from Kempelen Institute of Intelligent Technologies. His research highlights that fine-tuned multilingual models like XLM-RoBERTa are more efficient for cross-lingual claim detection than translation, and surprisingly, sentence length acts as a strong indicator of verifiable claims.

For typological feature prediction, “Typological Feature Prediction with Large Language Models: An In-Context Learning Approach” by Qianwen Wang et al. from the University of Toronto, Canada, shows that LLMs, when provided with phylogenetic and geographic neighbor evidence, significantly outperform traditional baselines, effectively narrowing the performance gap for low-resource languages.

And in understanding Bangla idioms, Mousumi Akter et al. from Technical University Dortmund in “To What Extent Do Large Language Models Understand Bangla Idioms?” introduce the first large-scale benchmark for Bangla idioms. Their findings indicate that no single LLM excels across all idiom tasks, and embedding-based metrics are far better at judging idiomatic paraphrasing than traditional lexical overlap metrics.

Under the Hood: Models, Datasets, & Benchmarks

This collection of papers highlights the vital role of new datasets and targeted model adaptations in advancing low-resource language NLP. Key resources and innovations include:

• GAD-ALT Corpus: The first systematic benchmark for Greek Automatic Lyric Transcription, a curated dataset of 17,458 aligned lyric segments (19.65 hours). Code and models are available on GitHub and Hugging Face.
• Urdu-Stories Corpus: A dataset of 93 LLM-generated Urdu stories with 13-label linguistic, semantic, and cultural annotations to expose LLM weaknesses. Code for evaluation scripts and the annotation framework is available.
• 5-DIALECTS-BN: The first multi-annotation benchmark for Bangla dialects, featuring 6,000 manually verified entries across five regional varieties, including Romanized transliteration. An annotation tool is available at bangla-dialect-annotator.vercel.app.
• DasanCallDial: The first large-scale Korean dialogue-level ASR error correction dataset, comprising 1,974 dialogues and 115,460 utterances from real call centers. The DCSC framework, based on pkoT5, achieves state-of-the-art performance.
• SiDiaC-v.2.5: An enhanced Sinhala Diachronic Corpus with lemmatization and POS tagging, available on Hugging Face and GitHub.
• SEA-SpeechBench: The first comprehensive benchmark for speech understanding across 11 Southeast Asian languages, with over 97,000 samples and 597 hours of audio data, including novel temporal reasoning tasks. Resources available at zwenyu.github.io/SEA-SpeechBench and code at GitHub.
• IberianLLM-7B-Instruct with GRPO: Post-trained on ASSET (English) and applied to Catalan text simplification using a novel SARI + copy/length penalty reward function. Code is available on GitHub and Hugging Face.
• RAAI Framework: A training-free test-time framework for adaptive language mode selection, demonstrating significant gains on Belebele and MMLU-ProX-Lite datasets.
• XLM-RoBERTa for Multilingual Claim Detection: Fine-tuned models for simultaneous detection of verifiable factual and harmful claims across multiple low-resource languages. Code is available on GitHub.
• Bangla Idiom Benchmark: The first large-scale dataset of Bangla idioms with over 10,000 entries and a synthetic MCQ corpus. The dataset will be made available upon acceptance, as referenced in the paper “To What Extent Do Large Language Models Understand Bangla Idioms?”

Impact & The Road Ahead

These advancements represent crucial steps toward achieving truly inclusive AI. The impact is far-reaching, from empowering fact-checkers to combat disinformation in diverse linguistic communities to making advanced speech services accessible in regions where they are currently lacking. The development of fine-grained error correction for ASR, robust dialectal understanding, and even historical linguistic analysis for low-resource languages opens up new avenues for research and application.

However, significant challenges remain. The findings regarding cultural misalignment in Urdu and the transliteration penalty in Bangla underscore the need for deeper, culturally-aware linguistic understanding, rather than just surface-level fluency. The struggle of LLMs with complex tasks like idiom understanding and temporal reasoning in speech highlights areas ripe for further innovation.

The future of low-resource language AI lies in continued collaboration between linguists and AI researchers, the creation of more diverse and robust datasets, and the development of models that are not just multilingual but truly plurilingual – capable of nuanced understanding across distinct linguistic and cultural frameworks. The progress shown here is exhilarating, proving that dedicated research can indeed unlock the full potential of AI for every language.

Please leave this field empty

Hi there 👋

Get a roundup of the latest AI paper digests in a quick, clean weekly email.

Check your inbox or spam folder to confirm your subscription.