Austria

Machine translation (MT) has advanced significantly for high-resource languages, yet specialized domain translation remains a challenge for low-resource languages. This study evaluates the abilityof state-of-the-art multilingual models to translate animal health reports from English to Yorùbá, a crucial task for veterinary communication in underserved regions. We curated a dataset of 1,468 parallel sentences and compared multiple MT models in zero-shot and fine-tuned settings. Our findings indicate substantial limitations in their ability to generalize to domain-specific translation, with common errors arising from vocabulary mismatch, training data scarcity, and morphological complexity. Fine-tuning improves performance, particularly for the NLLB 3.3B model, but challenges remain in preserving technical accuracy. These results underscore the need for more targeted approaches to multilingual and culturally aware LLMs for African languages.

ACL 2025

Beyond Generalization :Evaluating Multilingual LLMs for Yorùbá Animal Health Translation

## Welcome!
"I am excited to welcome you to this year’s edition of the Conference on Empirical Methods in Natural Language Processing! Importantly, it marks the 30th edition of EMNLP. With over 8,000 submissions, more than 3,000 accepted papers, and thousands of attendees, we have come a long way from that first
workshop, which had 14 accepted papers. As the field looks ahead, Suzhou is the fitting location for celebrating this milestone: rooted in a long literary tradition, yet modern and forward-looking, and home to a large share of the NLP community." 

*Message from the General Chair, Dirk Hovy*

EMNLP 2025

Celebrate 30 Years of EMNLP! 
EMNLP 2025 will be held in Suzhou, China from November 5th to November 9th, 2025.

workshop paper

### Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

Message from the General Chair: 
*It is my great pleasure and honor to welcome you to the 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025), held in beautiful Vienna, Austria, from July 27 to August 1, 2025. ACL2025continues our field’s tradition of excellence in scholarship, innovation, and inclusivity, and I am deeply grateful to the many volunteers who have worked tirelessly to bring this event to life.* 
[Read more](https://drive.google.com/file/d/1GI_hvOpjswAuYdUTromfeDiPpCcqidwg/view?usp=sharing)

To access this event page, you need to log in with the **email address you registered with**. Access credentials will be sent to your email from Underline - subject line "Welcome to ACL 2025". Please be sure to check your spam email folder if you do not see an email confirmation right away.

Please log in

To access this event page, you need to log in with the **email address you registered with**. 

Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

The deployment of Large Language Models (LLMs) in real-world applications presents both opportunities and challenges, particularly in multilingual and code-mixed communication settings. This research evaluates the performance of seven leading LLMs in sentiment analysis on a dataset derived from multilingual and code-mixed WhatsApp chats, including Swahili, English and Sheng. Our evaluation includes both quantitative analysis using metrics like F1 score and qualitative assessment of LLMs' explanations for their predictions. We find that, while Mistral-7b and Mixtral-8x7b achieved high F1 scores, they and other LLMs such as GPT-3.5-Turbo, Llama-2-70b, and Gemma-7b struggled with understanding linguistic and contextual nuances, as well as lack of transparency in their decision-making process as observed from their explanations. In contrast, GPT-4 and GPT-4-Turbo excelled in grasping diverse linguistic inputs and managing various contextual information, demonstrating high consistency with human alignment and transparency in their decision-making process. The LLMs however, encountered difficulties in incorporating cultural nuance especially in non-English settings with GPT-4s doing so inconsistently. The findings emphasize the necessity of continuous improvement of LLMs to effectively tackle the challenges of culturally nuanced, low-resource real-world settings and the need for developing evaluation benchmarks for capturing these issues.

Beyond Metrics: Evaluating LLMs’ Effectiveness in Culturally Nuanced, Low-Resource Real-World Scenarios

In this paper, we introduce the first publicly available English-Kpelle dataset for machine translation, comprising over 2,000 sentence pairs drawn from everyday communication, religious texts, and educational materials. By fine-tuning Meta’s No Language Left Behind (NLLB) model on two versions of the dataset, we achieved BLEU scores of up to 30 in the Kpelle-to-English direction, demonstrating the benefits of data augmentation. Our findings align with NLLB-200 benchmarks on other African languages, underscoring Kpelle’s potential for competitive performance despite its low-resource status. Beyond machine translation, this dataset enables broader NLP tasks, including speech recognition and language modeling. We conclude with a roadmap for future dataset expansion, emphasizing orthographic consistency, community-driven validation, and interdisciplinary collaboration to advance inclusive language technology development for Kpelle and other low-resourced Mande languages

Building a Functional Machine Translation Corpus for Kpelle

Low-resource languages face significant challenges in natural language processing due to the scarcity of annotated data, linguistic resources, and the lack of language standardization, which leads to variations in grammar, vocabulary, and writing systems. This issue is particularly observed in many African languages, which significantly reduces their usability. To bridge this barrier, this paper investigates the challenges and limitations of collecting datasets for the Medumba language, a Grassfields Bantu language spoken in Cameroon, in the context of extremely low-resource natural language processing. We mainly focus on the specificity of this language, including its grammatical and lexical structure. Our findings highlight key barriers, including (1) the challenges in typing and encoding Latin scripts, (2) the absence of standardized translations for technical and scientific terms, and (3) the challenge of limited digital resources and financial constraints, highlighting the need to improve data strategies and collaboration to advance computational research on African languages. We hope that our study informs the development of better tools and policies to make knowledge platforms more accessible to extremely low-resource language speakers. We further discuss the representation of the language, data collection, parallel corpus development.

Challenges and Limitations in Gathering Resources for Low-Resource Languages: The Case of Medumba

Despite rapid advancements in multimodal large language models (MLLMs), their ability to process low-resource African languages in document-based visual question answering (VQA) tasks remains limited. This paper evaluates three state-of-the-art MLLMs—GPT-4o, Claude-3.5 Haiku, and Gemini-1.5 Pro—on WAEC/NECO standardized exam questions in Yoruba, Igbo, and Hausa. We curate a dataset of multiple-choice questions from exam images and compare model accuracies across two prompting strategies: (1) using English prompts for African language questions, and (2) using native-language prompts. While GPT-4o achieves over 90% accuracy for English, performance drops below 40% for African languages, highlighting severe data imbalance in model training. Notably, native-language prompting improves accuracy for most models, yet no system approaches human-level performance, which reaches over 50% in Yoruba, Igbo, and Hausa. These findings emphasize the need for diverse training data, fine-tuning, and dedicated benchmarks that address the linguistic intricacies of African languages in multimodal tasks, paving the way for more equitable and effective AI systems in education.

Challenging Multimodal LLM with African Standardized Exams: A Document VQA Evaluations

Building high-quality large language models (LLMs) for enterprise Arabic applications remains challenging due to the limited availability of digitized Arabic data. In this work, we present a data synthesis and refinement strategy to help address this problem, namely, by leveraging synthetic data generation and human-in-the-loop annotation to expand our Arabic training corpus. We further present our iterative post training recipe that is essential to achieving state-of-the-art performance in aligning the model with human preferences, a critical aspect to enterprise use cases. The culmination of this effort is the release of a small, 7B, open-weight model that outperforms similarly sized peers in head-to-head comparisons and on Arabic-focused benchmarks covering cultural knowledge, instruction following, RAG, and contextual faithfulness.

Command R7B Arabic: a small, enterprise-focused, multilingual, and culturally aware Arabic LLM

Pretrained language models (PLMs) for African languages are continually improving, but the reasons behind these advances remain unclear. This paper presents the first systematic investigation into how knowledge about African languages is encoded in PLMs. We train layer-wise probes for six typologically diverse African languages to ana\-lyse how linguistic features are distributed. We also design control tasks, a way to interpret probe performance, for the MasakhaPOS dataset. We find PLMs adapted for African languages to encode more linguistic information about target languages than massively multilingual PLMs. Our results reaffirm previous findings that token-level syntactic information concentrates in middle-to-last layers, while sentence-level semantic information is distributed across all layers. Through control tasks and probing baselines, we confirm that performance reflects the internal know\-ledge of PLMs rather than probe memorisation. Our study applies established interpretability techniques to African-language PLMs. In doing so, we highlight the internal mechanisms underlying the success of strategies like active learning and multilingual adaptation.

Designing and Contextualising Probes for African Languages


The integration of Artificial Intelligence (Al) in agriculture has significantly impacted decision making processes for farmers, particularly in regions such as Kenya, where access to accurate and timely advisory services is crucial. This paper explores the deployment of Retrieval Augmented Generation (RAG) agents powered by fine-tuned quantized language models to enhance Al-driven agricultural advisory services. By optimizing model efficiency through quantization and fine-tuning, our aim is to deliver a specialized language model in agriculture and to ensure real-time, cost-effective and contextually relevant recommendations for smallholder farmers. Our approach takes advantage of localized agricultural datasets and natural language processing techniques to improve the accessibility and accuracy of advisory responses in local Kenyan languages. We show that the proposed model has the potential to improve information delivery and automation of complex and monotonous tasks, making it a viable solution to sustainable agricultural intelligence in Kenya and beyond.

Enhancing AI-Driven Farming Advisory in Kenya with Efficient RAG Agents via Quantized Fine-Tuned Language Models

Generative AI models are primarily accessed through chat interfaces, where user queries often contain typographical errors. While these models perform well in English, their robustness to noisy inputs in low-resource languages like Yorùbá remains underexplored. This work investigates a Yorùbá question-answering (QA) task by introducing synthetic typographical noise into clean inputs. We design a probabilistic noise injection strategy that simulates realistic human typos. In our experiments, each character in a clean sentence is independently altered, with noise levels ranging from 10% to 40%. We evaluate performance across three strong multilingual models using two complementary metrics: (1) a multilingual BERTScore to assess semantic similarity between outputs on clean and noisy inputs, and (2) an LLM-as-judge approach, where the best Yorùbá-capable model rates fluency, comprehension, and accuracy on a 1–5 scale. Results show that while English QA performance degrades gradually, Yorùbá QA suffers a sharper decline. At 40% noise, GPT-4o experiences over a 50% drop in comprehension ability, with similar declines for Gemini 2.0 Flash and Claude 3.7 Sonnet. We conclude with recommendations for noise-aware training and dedicated noisy Yorùbá benchmarks to enhance LLM robustness in low-resource settings.

Evaluating Robustness of LLMs to Typographical Noise in Yorùbá QA

The performance of Automatic Speech Recognition (ASR) depends on the availability of transcribed speech datasets—often scarce or non-existent for many of the world’s languages. This study investigates alternative strategies to bridge the data gap using zero-shot cross-lingual transfer, leveraging transliteration as a method to utilize data from other languages. We experiment with transliteration from various source languages and demonstrate ASR performance in a low-resourced language, Amharic. We find that source data that align with the character distribution of the test data achieves the best performance, regardless of language family. We also experiment with fine-tuning with minimal transcribed data in the target language. Our findings demonstrate that transliteration, particularly when combined with a strategic choice of source languages, is a viable approach for improving ASR in zero-shot and low-resourced settings.

Exploring Transliteration-Based Zero-Shot Transfer for Amharic ASR

Automatic Speech Recognition (ASR) technologies have seen significant advancements, yet many widely spoken languages remain underrepresented. This paper explores the fine-tuning of OpenAI's Whisper Tiny model (39M parameters) for Swahili, a lingua franca for over 100 million people across East Africa. Using a dataset of 5,520 Swahili audio samples, we analyze the model's performance, error patterns, and limitations after fine-tuning. Our results demonstrate the potential of fine-tuning for improving transcription accuracy, while also highlighting persistent challenges such as phonetic misinterpretations, named entity recognition failures, and difficulties with morphologically complex words. We provide recommendations for improving Swahili ASR, including scaling to larger model variants, architectural adaptations for agglutinative languages, and data enhancement strategies. This work contributes to the growing body of research on adapting pre-trained multilingual ASR systems to low-resource languages, emphasizing the need for approaches that account for the unique linguistic features of Bantu languages.

Downloads

Next from ACL 2025

Beyond Metrics: Evaluating LLMs’ Effectiveness in Culturally Nuanced, Low-Resource Real-World Scenarios

Similar lecture

AFRIDOC-MT: Document-level MT Corpus for African Languages

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES