Tag
Orange Research presents 0.6B–14B open-weight Qwen models adapted to Wolof, a language spoken by 23 million people, using instruction tuning, preference alignment, and model merging. The models are competitive with much larger commercial systems on Wolof tasks and are being piloted in customer-care deployments, with data and weights publicly released.
ELARSpeech introduces a 30,202-hour speech dataset covering 675 underrepresented languages curated from the Endangered Languages Archive, along with trained ASR/TTS models and a full data pipeline to advance long-tail multilingual speech technologies.
该论文对八个闭源和开源大模型在英文与孟加拉语XL-Sum测试集上的零样本抽象摘要能力进行了标准化评测，结合词法、语义指标、事实性验证和LLM-as-a-Judge评估，发现英文表现优异并不自动迁移到低资源语言，且不同评估方法产生不同的模型排名。
文章探讨了大型AI模型在小语种和地方方言上的表现不佳问题：训练数据不足导致翻译生硬怪异，并可能把用户推向主流语言。文章提出AI公司是否应承担支持小语种的责任，即使缺乏商业回报。
论文提出 HexaBee-4B，一个基于 Gemma3 4BIT 通过三阶段训练（维基百科持续预训练、多语言问答与答案翻译联合训练、双向翻译专业化）改编的六语种翻译模型，在阿拉伯语、中文、英语、希伯来语、波斯语和西班牙语全部 30 个翻译方向上相比基线均有 chrF++ 提升。
该论文提出 Language-Aware Skeleton Exploration Framework (LASEF)，系统研究多层数学推理中骨架提示语言的选择问题，发现英语骨架平均略有优势但并非普遍最优，骨架语言是一个依赖上下文的设计变量。
The paper introduces sk-bench, a native-first benchmark for evaluating large language models in Slovak, covering 30 datasets (33 task variants) across ten skill categories to address the lack of native Slovak coverage in existing multilingual benchmarks.
The paper introduces Multilingual GSM-Symbolic, a symbolic-template dataset covering 15 languages and 30k item-matched QA pairs, to quantify what determines cross-lingual capability transfer in LLMs. It finds model size, language resource level, reasoning, and typological distance are key drivers, and shows a framework explaining 92% of between-language variation while predicting unseen-language performance within 6.0pp.
TutlAit v1 presents a crowdsourced Moroccan Tamazight speech dataset of 13,384 audio files (~20.9 hours) paired with Modern Standard Arabic transcriptions and regional accent labels, collected via a purpose-built web application for low-resource speech technology.
A position paper introducing 'Model as a Library' (MaaL), a software architecture that packages small, community-enrolled speech models as versioned on-device dependencies for offline, hallucination-free data collection in low-resource African languages. It proposes keyword spotting to turn closed-vocabulary digital forms into voice-based, on-device data collection for low-literacy populations, and transpiling existing form tools into MaaL schemas.
This paper evaluates how reliably sentiment concept directions extracted from language model representations reproduce across splits, comparing English, Hausa, and Yoruba over four language models. It finds consistent language rank order in direction agreement (English > Hausa > Yoruba) and shows that high probe classification accuracy does not imply directional consistency, though it does not establish tokenizer fertility as the cause.
This paper benchmarks transformer models for Bangla medical Named Entity Recognition, demonstrating that fine-tuned XLM-RoBERTa achieves state-of-the-art performance and outperforms language-specific models, emphasizing the importance of domain diversity in low-resource settings.
EnSiTa is a trilingual multi-domain parallel dataset and benchmark for English, Sinhala, and Tamil, featuring human post-edited training data and extensive experiments on domain-specific machine translation to address low-resource language challenges.
This paper audits a multilingual affective generation benchmark and reveals that system rankings are driven by measurement artifacts rather than genuine performance differences, with annotator variability playing a key role.
The paper introduces BanglaTurn, a benchmark corpus and Whisper-based model for end-of-turn detection in Bangla speech, achieving 84.33% accuracy compared to a 69.28% baseline.
BanglaKontho presents a 20-hour single-speaker Bangla TTS corpus from professional audiobooks, along with an open-source text normalizer and preprocessing pipeline, to address the gap in long-form prosody for low-resource Bangla speech synthesis.
This paper presents COILD, an Indic-centric parallel corpus with over 1.16 million sentence pairs for 20 Indian language pairs and a domain-centric benchmark, demonstrating improvements in machine translation when fine-tuning multilingual models.
This paper introduces VākQA, a Telugu spoken factoid question answering benchmark, and evaluates models using automated and human judges to address low-resource language challenges.
The University of Melbourne submits a domain-balanced machine translation model for low-resource Pacific creoles like Tok Pisin, Bislama, and Solomon Pijin, outperforming open baselines by over 3 chrF++ points in the WMT 2026 CreoleMT shared task.
This paper evaluates the reusability of publicly available speech resources for low-resource languages, focusing on Central Kurdish as a case study.