Despite advancements in Large Language Models, reasoning benchmarks remain centered on high-resource languages, leaving languages like Vietnamese under-evaluated. In this study, we aim to address this gap by evaluating four models: PhoGPT (native), Vistral and VBD-Llama (adapted), and Llama-2 (English-centric), on commonsense reasoning and arithmetic reasoning. As Vietnamese benchmarks for these tasks are lacking, we adapt two analogy datasets from English to Vietnamese and construct two sequence datasets, ensuring a range of structural complexity and difficulty levels. We evaluate diverse prompting strategies, including Chain-of-Thought, role-playing guidance, cross-lingual prompting, and few-shot learning. Our results reveal a baseline proficiency in analogical and arithmetic reasoning among the models, with Vistral and Llama-2 outperforming other models in multiple tasks. The effects of Chain-of-Thought and contextual guidance are limited in Vietnamese, while cross-lingual prompting and few-shot learning show promising performance improvements. The findings underscore the feasibility of adapting benchmarks to less-resourced languages and provide insights into strengths and weaknesses in the performance of Vietnamese LLMs, suggesting directions for model improvements.

Automatic scalar metrics are widely used for machine translation (MT) evaluation, yet their behavior under sociolinguistic variation remains underexplored, particularly in under-resourced and minority-language contexts. We present a small, controlled empirical analysis of reference-based evaluation in Spanish–Basque informal discourse. Register is operationalized as indexical density, capturing dialectal forms, informal lexicon, code-switching and orthographic stylization. Across two MT systems and prompting conditions, sentence-level scores from chrF++, COMET-DA, and XCOMET-XL show a consistent negative association with indexical density under the original informal reference. In a reference-perturbation design that holds MT outputs constant while replacing the informal reference with a standardized Batua version, scores increase systematically, particularly for high-density items, and the density–score association weakens. These results provide controlled evidence that evaluation outcomes in this setting depend in part on reference register configuration. In minority-language and informal domains, reference design choices may influence how translation quality is measured and interpreted.

Manual annotation of pragmastylistic features in sensationalist media is a resource-intensive bottleneck for corpus- based research, particularly for lower-resource languages. This paper evaluates whether Large Language Models (LLMs) can reliably automate this process. We benchmark two proprietary models, OpenAI’s GPT-5 and Google’s Gemini 2.5 Pro, on annotating eight sensationalist linguistic and orthographic features within a corpus of 508 Serbian celebrity magazine headlines. Our methodology involves a systematic comparison of five prompting strategies: zero-shot, few-shot (1, 3, and 5 examples), and chain-of-thought. Results demonstrate that LLMs can achieve high alignment with a manually curated gold standard, reaching a peak macro-F1 score of 98.76%. Notably, the most effective and cost-efficient configuration was GPT-5 using a simple zero-shot prompt. Qualitative error analysis reveals that remaining inaccuracies are systematic, primarily involving pragmatic conventions, discourse scope, and quoted speech. We conclude that LLMs are viable for first-pass annotation of well-defined features in Serbian, though implicit and genre-dependent cues require further study. To support reproducibility and future research on underrepresented languages, we provide our full prompting setup, evaluation procedures, and a detailed cost comparison.

We explore the use of large language models (LLMs) for morphological disambiguation in Belarusian, a low-resource language. The pipeline has two stages: a rule-based analyzer generates candidate lemmas and grammatical tags, which an LLM then disambiguates in context. Initial evaluation of ChatGPT, Claude, and Gemini on a gold-standard sample shows high accuracy. We scale this approach to a 375K-word corpus using Gemini and compare the results against a neural baseline (Stanza). Manual review of discrepancies suggests that the LLM-based approach outperforms the baseline, offering a solution for corpus annotation in Belarusian.

We present mobile and desktop keyboards for Idu Mishmi, an endangered Trans-Himalayan language spoken by approximately 11,000 people in Arunachal Pradesh, India. A Latin-based orthography, the Idu Azobra, was developed in 2018, but no digital input tools existed to use it. The orthography requires characters absent from standard keyboards, including schwa (ə), retracted vowels (ə̱, o̱, u̱), nasalized vowels, and accented forms, several of which involve multi-codepoint Unicode sequences that default keyboards do not support. Developed with the Idu Mishmi community, the keyboards comprise: (1) an Android mobile keyboard, published on the Google Play Store, and (2) a Windows desktop keyboard distributed as a single portable executable. Both tools support the complete character inventory, and operate fully offline with zero network permissions. The Android keyboard has been adopted by community leaders and teachers who currently know and actively use the Idu Azobra orthography. The Windows keyboard is currently undergoing testing with community leaders. We describe the design, implementation, and deployment as a replicable model for other endangered language communities.

We investigate multilingual sentiment analysis and interpretability across high- and low-resource languages, focusing on Amharic, English, German, and Hausa. Our study evaluates encoder-only transformer models for both sequence-level sentiment classification and token-level attribution using Captum. Additionally, we assess zero- and few-shot decoder-only models for sequence-level sentiment prediction. Our results show that few-shot decoder-only models outperform encoder-only models on token-level sentiment classification in most languages, with the exception of Hausa, where a multilingual encoder-based model leads. For sequence-level sentiment classification, encoder-only models generally achieve strong performance across most languages, but decoder-only models are highly competitive, and may even surpass encoders, in the high-resource settings (German, English) and low-resource scenarios, depending on the prompting strategy. These findings highlight the utility of combining fine-tuned transformer models with prompt- based large language models to build interpretable sentiment analysis systems across both low- and high-resource languages. The SAINT dataset, annotation guideline, and evaluation scripts can be found at https://github.com/uhh-hcds/SAINT.

Safety evaluation of Large Language Models (LLMs) has largely focused on high-resource languages, leaving low-resource languages critically underserved. We present AlbanianLLMSafety, the first publicly available safety evaluation dataset for LLMs in Albanian, a linguistically distinct low-resource language with approximately 7.5 million speakers across Albania, Kosovo, North Macedonia, and the diaspora. The dataset contains 2,951 prompts spanning 11 safety categories, including self-harm, violence, racist content, child exploitation, and radicalization, with an average of 268 prompts per category. Each prompt is provided in Albanian with an English reference translation and a detailed category label. This resource addresses a significant gap in safety evaluation infrastructure for low-resource languages and provides an essential benchmark for developing safer, more inclusive LLMs. The dataset will be provided upon request to support safety evaluation, fine-tuning, red-teaming, and guardrail development for Albanian-speaking communities.

Visual Question Answering (VQA) bridges the gap between computer vision and natural language processing, yet progress remains largely confined to high-resource languages. For low-resource languages like Urdu, research is severely hindered by the total absence of large-scale reasoning-based datasets. To address this critical gap, we introduce the first synthetic Urdu VQA dataset modeled after the CLEVR framework, specifically designed to evaluate complex, multi-step visual reasoning. We conduct a rigorous comparative analysis using both transformer-based architectures (VisualBERT, LXMERT, ViLT) and neuro-symbolic models. Our results demonstrate that the neuro-symbolic approach achieves a superior accuracy of 85.3%, outperforming the strongest transformer baseline by 7.1% while maintaining competitive processing efficiency. This work establishes a primary benchmark for Urdu VQA, demonstrating that hybrid reasoning architectures provide a robust and scalable solution for advancing multimodal AI in under-resourced linguistic contexts.

We present a new database of Romance clitics across nine varieties. The database includes speech, transcriptions, and linguistic annotations. The database concentrates on clitics, and includes varieties of Romance with stressed clitics. Specifically, the database includes data from the regions of Corsica, Pyrénées-Atlantiques, Sardinia, Liguria, Basilicata, Campania, Mallorca, Menorca, and Formentera. A publicly accessible interface allows easy searching. The database and, separately, the interface code will be made publicly available.

This paper introduces GreekCommonGen, the first benchmark designed for generative commonsense reasoning in Greek. The dataset is created by automatically translating the original English CommonGen corpus and subsequently refining the outputs through manual post-editing to ensure linguistic and semantic quality. We conduct a comprehensive evaluation of a range of approaches/models on this benchmark, exploring the impact of different prompting strategies, decoding methods, and model sizes/architectures. Our findings provide valuable insights into the challenges of commonsense generation in Greek, paving the way for future research in the field.

In this early-stage study, we investigate whether transfer learning from lexifier or substrate languages can improve text-to-speech (TTS) performance for low-resource creoles. We conducted a controlled experiment using two creoles of distinct lexical origins: Nigerian Pidgin (English-based) and Guadeloupean Creole (French-based). Single-speaker TTS datasets of approximately 30 minutes each were recorded and used to fine-tune pretrained models for English, French, and Yoruba. Objective metrics and informal subjective evaluations were employed to assess synthesis quality. Though partially inconclusive, our results suggest that the French-based models outperform others for both creoles, while Yoruba-based models yield weaker performance. These findings may suggest that lexical similarity or historic influences alone do not fully predict transfer learning effectiveness, and that phonotactic compatibility and orthographic depth may also be relevant factors. Our work provides insight into TTS model development for creoles and other low-resource languages, and highlights avenues for further research on leveraging relevant linguistic and orthographic features for model development.

Kazakh is underrepresented in resources for evaluating the safety behavior of large language models. We present KZ-SafetyPrompts, a Kazakh prompt dataset for safety evaluation across eleven categories covering common risk areas such as self-harm, violence, child exploitation, sexual content, racist content, radicalization, and regulated goods or illegal activities. The dataset contains 5,717 prompts written natively in Kazakh (Cyrillic), organized by category, with English translations for cross-lingual analysis. Prompts resemble realistic user queries, often in a teen or child style, and are phrased as intent prompts without procedural instructions. We document the writing protocol, labeling procedures (including borderline-case decision rules), and quality-control steps (schema standardization, completeness checks, and deduplication). We also align the categories with widely used safety taxonomies to support integration with existing evaluation pipelines. Baseline results with GPT-4o show an overall refusal rate of 28.2%, varying from 5.5% to 53.8% across categories, indicating that Kazakh prompts expose category-specific safety gaps not captured by English-only evaluation.

Human judgements of word similarity have been a core benchmark for the intrinsic evaluation of word embedding models, and continue to be used for assessing the capabilities of large language models. While word similarity benchmarks have been collected for a range of languages, none existed for Greek. We develop a Modern Greek variant of the SimLex-999 word similarity dataset by gathering similarity judgements from 90 native speakers of Greek. We then use this as a benchmark for intrinsically evaluating several Greek language models.

The value of high-quality datasets for training essential language tools has long been recognised for NLP research. Despite the importance of such datasets, most language data available for training consists of large, automatically curated corpora, often scraped from web content. The quality of such datasets is often an unknown factor. This presents a problem for already low-resourced languages (such as Irish), as existing datasets may not provide adequate, representative language data for training effective models. This paper examines existing monolingual and parallel Irish text corpora to evaluate the quality of the language data, through manual review, automatic metrics, and LLMs as judges.

High-quality bilingual resources remain a critical bottleneck for advancing multilingual NLP in low-resource settings, particularly for Bangla. To mitigate this gap, we introduce BiST, a rigorously curated Bangla–English corpus for sentence-level grammatical classification, annotated across two fundamental dimensions: syntactic structure (Simple, Complex, Compound, Complex-Compound) and tense (Present, Past, Future). The corpus is compiled from open-licensed encyclopedic sources and naturally composed conversational text, followed by systematic preprocessing and automated language identification, resulting in 30,534 sentences, including 17,465 English and 13,069 Bangla instances. Annotation quality is ensured through a multi-stage framework with three independent annotators and dimension-wise Fleiss’ Kappa (κ) agreement, yielding reliable and reproducible labels with κ values of 0.82 and 0.88 for structural and temporal annotation, respectively. Statistical analyses demonstrate realistic structural and temporal distributions, while baseline evaluations show that dual-encoder architectures leveraging complementary language-specific representations consistently outperform strong multilingual encoders. Beyond benchmarking, BiST provides explicit linguistic supervision that supports grammatical modeling tasks, including controlled text generation, automated feedback generation, and cross-lingual representation learning. The corpus establishes a unified resource for bilingual grammatical modeling and facilitates linguistically grounded multilingual research.

This study presents a multi-dialect, pragmatically annotated Spanish corpus designed to address persistent gaps in the representation of regional varieties and communicative functions in existing linguistic and NLP resources. The corpus focuses exclusively on Spanish dialects spoken in the Americas, selecting one representative dialect per country and incorporating a single neutral Castilian variety for comparative purposes. Dialects are organized into five regional groups: Mexican, Central American, Caribbean, South American, and Rioplatense Spanish. Corpus development follows a multi-stage workflow in which a seed lexicon composed of openly licensed material from sources such as Wikipedia, Project Gutenberg, and curated random and synthetic data is used to initiate the LLM-based text generation. Each base sentence is expanded into dialect-specific variants and annotated with pragmatic and domain labels, producing a fully parallel dataset that supports cross dialect comparison. A multi-stage correction pipeline combining automated scripts, controlled LLM-based editing, and manual review ensures syntactic well-formedness and dialectal authenticity while eliminating language-switching and hallucination errors. The final version of the corpus covers 20 dialects and contains, 40,000 annotated sentences, released in both JSON and plain-text formats for use in a wide range of NLP tasks.

Hawaiian (ʻŌlelo Hawaiʻi) is an endangered Polynesian language whose broadcast archives represent a critical yet underutilized resource for language documentation. We present the first evaluation of vision-language models (VLMs) for structured entity extraction from television chyrons, investigating the performance gap between Hawaiian-language content and mainland U.S. comparisons. Using our new HiChy dataset of 3,925 manually annotated images, we demonstrate that Hawaiian content remains significantly more challenging for current VLMs: for the best-performing model (Qwen2.5-VL-7B), character error rates roughly double from 0.064 on mainland data to 0.130 on Hawaiian content. We extend the task to key information extraction (KIE), finding that while models can perform structured parsing, they struggle specifically with names of Hawaiian linguistic origin, a difficulty that persists even when controlling for geographic source. Across five evaluated models spanning local quantized inference and commercial APIs, we find that OCR accuracy and structured extraction capability do not necessarily correlate: the best OCR model (Gemini 3 Flash) underperforms locally-deployed alternatives on KIE, while even a 2.2B-parameter model (SmolVLM2) achieves functional extraction. Our results provide a baseline for AI-assisted archival processing of underrepresented language media and highlight the need for models that better account for the orthographic and cultural specificities of Hawaiian.

The world’s languages are commonly categorised in terms of available language technologies such as speech recognition and machine translation. On this view, “under-resourced languages” suffer from language barriers which cut people off from markets, healthcare, human rights, and AI. The “solution” is more funding for language technologies, opening the way to a utopia of digital language equality and AI-enabled mobility. Yet the world’s linguistic diversity is not a set of language objects to be pushed up a cline from emerging to thriving. It consists of polyglossic communities who have long used vernaculars for local functions and dominant languages for external functions. I present a new theory of linguistic diversity which places the world of vernaculars alongside the world of institutional languages, and articulates diverse language technology agendas that lie within and between these worlds.

Translating very low-resource languages is a challenge that has been approached using available linguistic cues. Among them, interlinear glosses are linguistic annotations that can essentially bridge the gap between two languages thanks to both grammatical and lexical information. We perform a case study on a simulated low-resource condition for Turkish, a morphologically rich language, with a pipeline approach, following (Zhou et al., 2020). A source sentence is passed through a morphological analyzer and a bilingual dictionary to obtain a gloss-like representation. We then evaluate the current capacity of Neural Machine Translation systems and Large Language Models in performing the translation task from interlinear glosses into fluent English translations. We notably evaluate how performance scales with multilingual glossed data and how translation is affected by pseudo-glosses. Pivoting with glosses remains a better approach than a direct translation for languages with limited parallel data for training. Although glosses remain helpful resources, translations are sensitive to their quality, especially for lexical information.

Multilingual large language models are known to perform very well on high-resource languages, while their ability to process severely under-resourced languages remains underexplored. We investigate multilingual LLM translation performance on Fuzhounese, an under-resourced Sinitic language without a standardized orthography and almost no digital presence. Having adopted some methodological insights from the HKCanto-Eval benchmark, this paper presents a bidirectional translation framework based on a dataset of 305 sentences (300 constructed English sentences and 5 additional reference translations), that assesses the comprehension and generation of Fuzhounese, evaluated using automatic metrics and human Likert-scale judgments. The results reveal poor performance on Fuzhounese in both translation directions: BERTScore and chrF++ values consistently stay low when models are faced with comprehension tasks, while for generation tasks, scores are generally more than twofold lower than those for Mandarin or Cantonese. These findings highlight structural biases in multilingual LLMs toward high-resource languages and stress the need for resource-aware modeling and evaluation approaches in multilingual NLP systems.

Nepali Natural Language Processing (NLP) models are typically trained and evaluated on Standard Nepali, which can introduce bias against regional dialects. This study investigates the performance of Named Entity Recognition (NER) and Part-of-Speech (POS) tagging models on Achhami, a Far-Western dialect of Nepal. A parallel corpus of 300 sentence pairs was created, covering news, cultural topics, and everyday conversations. Achhami translations were produced by native speakers to preserve linguistic authenticity. The evaluation compared fine-tuned Transformer models with large language models using zero-shot prompting. Across both tasks, all models showed consistent performance degradation on the Achhami dialect. For NER, F1 scores decreased by 2.12 to 3.97 percent. Claude 3.5 Haiku achieved the best NER performance, while the monolingual NepBertA model unexpectedly outperformed multilingual alternatives, challenging assumptions about multilingual advantages. POS tagging results showed a similar pattern, with accuracy dropping notably on Achhami data. Large language models also showed comparable weaknesses, with accuracy reductions ranging from 2.9 to 7.0 percent. These findings quantify Kathmandu-centered bias in Nepali NLP and highlight the importance of dialectally diverse training data for building more inclusive and equitable language technologies.

Spontaneous speech in multilingual communities often involves rapid code-mixing (CM) and natural disfluencies, yet such patterns are rarely reflected in available training data for under-resourced languages. This gap limits the development of robust automatic speech recognition (ASR) systems. To address this, we introduce BEHE-CMDisfl, a fully synthetic Bengali–English and Hindi–English disfluent code-mixed speech corpus generated through a controlled Large Language Model (LLM) and Text-to-Speech (TTS) pipeline. The dataset explicitly incorporates conversational phenomena such as filled pauses, repetitions, and restarts. We evaluate its usefulness under two ASR settings. In a micro-resource scenario (∼1.3 hours), a GMM-HMM Kaldi baseline achieved a 37.74% Word Error Rate (WER) after phonetic normalization to reduce transliteration inconsistencies, and successfully retained disfluency markers in decoding. We also examined adaptation of a modern foundation model. In zero-shot testing, openai/whisper-small failed on the code-mixed speech due to severe hallucinations and looping behavior. After applying parameter-efficient fine-tuning (LoRA) for 1,000 steps, the model stabilized, reduced insertion errors, captured rapid language switching more reliably, and achieved a WER of 21.37%. These findings show that synthetic data combined with efficient fine-tuning offers a practical path for ASR development in complex low-resource disfluent CM settings.

Multilingual sentence-embedding models are widely used for cross-lingual retrieval; however, their performance drops significantly in low-resource languages. The Urdu language, which is considered a low-resource language by the NL community, poses this challenge, despite being spoken by over 246 million people worldwide. Its distribution in training corpora results in poor alignment with English within shared embedding spaces. To resolve this misalignment without model fine-tuning, we apply Procrustes transformation, which is an orthogonal post-hoc alignment method with a closed-form solution. We utilize SQuAD and UQA datasets to learn a rotation matrix from a small set of sentence pairs and evaluate its effect across five multilingual embedding models (MiniLM, DistilUSE, E5-Base, LaBSE, and E5-Large) and perform geometric alignment, cross-lingual retrieval, and question-answering tasks on these models. We find that cosine distances between parallel pairs decrease by up to 38.67%, and retrieval accuracy improves by 12.49% points in Recall@1. We also analyze that models with better pre-trained cross-lingual representations exhibit a saturation effect, showing minimal retrieval change even as geometric tightening increases. Our error analysis reveals that morphologically complex queries and colloquial expressions remain challenging, indicating representational limitations beyond the scope of a linear transformation. These findings demonstrate that a computationally inexpensive alignment step can meaningfully improve cross-lingual retrieval for low-resource languages, with implications for retrieval-augmented generation (RAG) in resource-constrained settings.

Formosan languages are a critically endangered group of Austronesian languages spoken in Taiwan, with severely limited representation in natural language processing (NLP) research and no support in existing language identification (LID) tools. We present the first systematic evaluation of machine learning models for the language identification of Kavalan, a Formosan language with fewer than 300 known speakers. We construct two benchmarks: a deployment-oriented benchmark with languages commonly confused with Kavalan by existing tools, and a linguistically motivated benchmark of typologically related Formosan languages. We evaluate random forest, support vector machine (SVM), and three pre-trained multilingual models using repeated stratified cross-validation. SVM models with character n-gram features achieve the strongest performance on both benchmarks, with a macro F1 of 0.993 on the deployment benchmark and a macro F1 of 0.906 on the Formosan benchmark, while remaining computationally inexpensive and effective with a low amount of data. Pre-trained multilingual models degrade significantly on the Formosan benchmark, with XLM-RoBERTa falling to a macro F1 of 0.505. These results demonstrate that traditional n-gram-based approaches are effective with low-resource Formosan LID and establish a foundation for downstream NLP tasks supporting the documentation and revitalization of low-resource Formosan languages.

Lombard is an endangered and under-resourced Gallo-Italic language variety that exists with Standard Italian. As with other language varieties of Italy, code-switching and code-mixing is common between Lombard and Italian in everyday conversation and with English, online. This linguistic complexity, and the lack of a unified written standard, poses challenges for Natural Language Processing tools. We introduce Rebelòt, a novel multi-domain, token-level annotated dataset for Lombard-Italian-English code-mixing. Furthermore, we develop and evaluate three variants of a token-level Language Identification (LID) tool based on a pre-trained encoder architecture, fine-tuned using both authentic data from our corpus and synthetically generated code-mixed text. Our evaluation demonstrates that the optimal model variant achieves an accuracy of over 0.99 on token-level prediction, and substantially outperforms widely used off-the-shelf LID baselines at sentence-level.

This paper introduces the Spontaneous Persian Speech (SPS) dataset designed for automatic speech recognition (ASR) tasks and a methodology laying the groundwork for addressing the shortage of spontaneous speech data. The corpus aims to support research on natural and conversational Persian, which remains under-represented in current ASR resources. The dataset consists of 694 minutes of audio from a total of 65 speakers, including 34 male and 31 female speakers. It contains 526,585 tokens. The audio segmentation step produces intervals of 1.24 to 3.25 seconds, each containing 3 to 9 words. The recordings cover a variety of environments, from inside cars to homes and shopping areas, including both busy and quiet settings. We use the SPS dataset to fine-tune Whisper and the performance increases significantly for both the small and medium models based on Word Error Rate (WER). This could be an initiative toward building domain-oriented datasets for specific ASR tasks.

Small Language Models (SLMs), typically ranging from a few million to 10–15 billion parameters, offer a promising solution towards constraints imposed by platform size–particularly mobile and IoT devices–and by the requirements of many organizations such as SMEs, which need solutions that ensure data privacy while remaining cost-effective. Their compactness and efficiency provide digital sovereignty and flexibility, though with more limited general-purpose capabilities. This makes them especially sensitive when working with underrepresented languages, such as Catalan, due to interference from majority languages that can increase bias risk. This paper evaluates state-of-the-art SLMs in a real-world Catalan use case: an AI assistant for older adults, assessing both user interactions and structured function call generation. Our work, which contributes to the Anonymized-Project initiative for deploying a connected SLM-based infrastructure under the Model Context Protocol, demonstrates that some SLMs are able to deliver high-quality performance even in resource-constrained, linguistically minority environments.

The Amazon basin along the Peru–Brazil border hosts extraordinary linguistic diversity, including many Indigenous languages whose speaker communities span national frontiers. Despite sustained documentation work, most remain extremely low-resource languages (ELRLs) for Natural Language Processing (NLP): reusable corpora are scarce, orthographies vary across countries and institutions, and basic tools such as tokenizers, taggers, and morphological analyzers are largely unavailable. We present a resource-oriented survey of five Indigenous languages of the Western Amazon—Matsés, Amahuaca, Kashinawa, Ticuna, and Kukama-Kukamiria—aimed at supporting more realistic NLP and speech work in extreme low-resource settings. Using a systematic search across academic venues, language archives, and public code/model repositories, we identify and cross-check available materials spanning lexical resources, text corpora, linguistic annotation, and speech collections. For each item we record practical reuse information, including the relevant task or modality, source location, and any stated access, licensing, or usage conditions. Our findings show strong cross-language asymmetries and fragmentation: most materials concentrate in documentation artifacts and lexicons, while standardized datasets with clear access and reuse conditions suitable for training and evaluation remain rare. We conclude with concrete recommendations to improve discoverability, normalize orthographic variation, and prioritize resource creation that maximizes interoperability across tools and benchmarks.

AfriVoices-KE is a large-scale multilingual speech dataset comprising approximately 3,000 hours of audio across five Kenyan languages: Dholuo, Kikuyu, Kalenjin, Maasai, and Somali. The dataset includes 750 hours of scripted speech and 2,250 hours of spontaneous speech, collected from 4,777 native speakers across diverse regions and demographics. This work addresses the critical underrepresentation of African languages in speech technology by providing a high-quality, linguistically diverse resource. Data collection followed a dual methodology: scripted recordings drew from compiled text corpora, translations, and domain-specific generated sentences spanning eleven domains relevant to the Kenyan context, while unscripted speech was elicited through textual and image prompts to capture natural linguistic variation and dialectal nuances. A customized mobile application enabled contributors to record using smartphones. Quality assurance operated at multiple layers, encompassing automated signal-to-noise ratio validation prior to recording and human review for content accuracy. Though the project encountered challenges common to low-resource settings, including unreliable infrastructure, device compatibility issues, and community trust barriers, these were mitigated through local mobilizers, stakeholder partnerships, and adaptive training protocols. AfriVoices-KE provides a foundational resource for developing inclusive automatic speech recognition and text-to-speech systems, while advancing the digital preservation of Kenya’s linguistic heritage.

Humor recognition is a major challenge in Natural Language Processing (NLP) due to its subtle and context-dependent nature. Despite advances, Large Language Models (LLMs) still struggle with this task, especially in Brazilian Portuguese, where no dedicated benchmarks exist. This paper presents HuNeBR, a new benchmark of 475 annotated humorous texts from Northeastern Brazilian comedians. The benchmark evaluates LLMs on three tasks: identifying punchlines, classifying texts into eight comic styles, and explaining humor. This is the first benchmark to evaluate LLMs on the in-depth interpretation of humorous texts in Brazilian Portuguese, going beyond the binary tasks of traditional humor benchmarks. Both general-purpose and Portuguese-specialized LLMs were evaluated under zero-shot and few-shot settings. The findings indicate that LLMs perform very well at identifying punchlines, show inconsistent results in classifying comic styles, and produce humor interpretations that mostly align with human judgments. Among the models assessed, general-purpose multilingual systems like GPT-4 and Gemini 2.5 Flash achieved the top overall performance, whereas Sabiá 3.1, a model specialized in Brazilian Portuguese, demonstrated competitive results across all three tasks, highlighting the value of locally trained models in capturing linguistic and cultural subtleties.

Esperanto is a widespread constructed language, known for its regular grammar and productive word formation. Besides having substantial resources available thanks to its online community, it remains relatively underexplored in the context of modern machine translation (MT) approaches. In this work, we present the first comprehensive evaluation of open-source MT systems for Esperanto, comparing rule-based systems, encoder–decoder models, and LLMs across model sizes. We evaluate translation quality across six language directions involving English, Spanish, Catalan, and Esperanto using multiple automatic metrics as well as human evaluation. Our results show that the NLLB family achieves the best performance in all language pairs, followed closely by our trained compact models and a fine-tuned general-purpose LLM. Human evaluation confirms this trend, with NLLB translations preferred in approximately half of the comparisons, although noticeable errors remain. In line with Esperanto’s tradition of openness and international collaboration, we release our code and best-performing models publicly.