Austria

This research article examines the effectiveness of various pretraining strategies for developing machine translation models tailored to low-resource languages. Although this work considers several low-resource languages, including Afrikaans, Swahili, and Zulu, the translation model is specifically developed for Lingala, an under-resourced African language, building upon the pretraining approach introduced by Reid and Artetxe (2021), originally designed for high-resource languages. Through a series of comprehensive experiments, we explore different pretraining methodologies, including the integration of multiple languages and the use of both monolingual and parallel data during the pretraining phase. Our findings indicate that pretraining on multiple languages and leveraging both monolingual and parallel data significantly enhance translation quality. This study offers valuable insights into effective pretraining strategies for low-resource machine translation, helping to bridge the performance gap between high-resource and low-resource languages. The results contribute to the broader goal of developing more inclusive and accurate NLP models for marginalized communities and underrepresented populations. The code and datasets used in this study are publicly available to facilitate further research and ensure reproducibility, with the exception of certain data that may no longer be accessible due to changes in public availability.

ACL 2025

Pretraining Strategies using Monolingual and Parallel Data for Low-Resource Machine Translation

**Welcome to ACL-IJCNLP 2021!**

The great event is jointly organized by the Association for Computational Linguistics (ACL) and Asian Federation of Natural Language Processing (AFNLP). 

As in previous years, the program of the conference includes a poster session, tutorials, workshops and demonstrations in addition to the main conference.


We were able to keep the registration fees similar to those charged for the virtual ACL 2020. The one fee allows attendance at the main conference and any/all tutorials and workshops. These fees would be $125 Regular Early and $175 Regular Late; $50 Student Early and $75 Student Late. Early registration closes at midnight July 11, 2021 (Eastern Daylight Time).

**Reminder:** It is ACL’s policy that at least one author of each accepted paper (including ACL Finding papers) must register for the conference.

**Reminder2:** Underline site will open closer to the event. If you already registered you will receive access detail

Registration is now open

ACL-IJCNLP 2021

The great event is jointly organized by the Association for Computational Linguistics (ACL) and Asian Federation of Natural Language Processing (AFNLP).

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)

Medication errors are among the leading causes of avoidable harm in healthcare systems across the world. A large portion of these errors stem from inefficient information retrieval processes and lack of comprehension of drug information. In low-resource settings, these issues are exacerbated by limited access to updated and reliable sources, technological constraints, and linguistic barriers. Innovations to improve the retrieval and comprehension of drug-related information are therefore poised to reduce medication errors and improve patient outcomes. This research employed open-source Retrieval-Augmented Generation (RAG) integrated with multilingual translation and Text-to-Speech (TTS) systems. Using open-source tools, a corpus was created from prominent sources of medical information in Nigeria and stored as high-level text embeddings in a Chroma database. Upon user query, relevant drug information is retrieved and synthesized using a large language model. This can be translated into Yoruba, Igbo, and Hausa languages, and converted into speech through the TTS system, addressing the linguistic accessibility gap. Evaluation of the system by domain experts indicated impressive overall performance in translation, achieving an average accuracy of 73%, and the best performance observed in Hausa and Yoruba. TTS results were moderately effective (mean = 57%), with Igbo scoring highest in speech clarity (68%). However, tonal complexity, especially in Yoruba, posed challenges for accurate pronunciation, highlighting the need for language-specific model fine-tuning. Addressing these linguistic nuances is essential to optimize comprehension and practical utility in diverse healthcare settings. The results demonstrates system’s the potential to improve access to drug information, enhance comprehension, and reduce linguistic barriers. These technologies could substantially mitigate medication errors and improve patient safety. This study offers valuable insights and practical guidelines for future implementations aimed at strengthening global medication safety practices.

Retrieval-Augmented Generation Meets Local Languages for Improved Drug Information Access and Comprehension.

The rapid advancement of large language models (LLMs) has revolutionized natural language processing, yet a significant challenge persists: the under representation of low-resource languages. This paper introduces \textsc{SabiYarn}, a novel 125M parameter decoder-only language model specifically designed to address this gap for Nigerian languages. Our research demonstrates that a relatively small language model can achieve remarkable performance across multiple languages even in a low-resource setting when trained on carefully curated task-specific datasets. We introduce a multitask learning framework designed for computational efficiency, leveraging techniques such as sequence packing to maximize token throughput per batch. This allows \textsc{SabiYarn} to make the most of a limited compute budget while achieving strong performance across multiple NLP tasks. This paper not only highlights the effectiveness of our approach but also challenges the notion that only massive models can achieve high performance in diverse linguistic contexts, outperforming models over 100 times its parameter size on specific tasks such as translation (in both directions), Named Entity Recognition, Text Diacritization, and Sentiment Analysis in the low-resource languages it was trained on. SabiYarn-125M represents a significant step towards democratizing NLP technologies for low-resource languages, offering a blueprint for developing efficient, high-performing models tailored to specific linguistic regions. Our work paves the way for more inclusive and culturally sensitive AI systems, potentially transforming how language technologies are developed and deployed in linguistically diverse areas like Nigeria and beyond.

SabiYarn: Advancing Low Resource Languages with Multitask NLP Pretraining

The development of Large Language Models (LLMs) for African languages has been hindered by the lack of large-scale textual data. Previous research has shown that relatively small language models, when trained on synthetic data generated by larger models, can produce fluent, short English stories, providing a data-efficient alternative to large-scale pretraining. In this paper, we apply a similar approach to develop and evaluate small language models for generating children’s stories in isiZulu and Yoruba, using synthetic datasets created through translation and multilingual prompting. We train six language-specific models varying in dataset size and source, and based on the GPT-2 architecture. Our results show that models trained on synthetic low-resource data are capable of producing coherent and fluent short stories in isiZulu and Yoruba. Models trained on larger synthetic datasets generally perform better in terms of coherence and grammar, and also tend to generalize better, as seen by their lower evaluation perplexities. Models trained on datasets generated through prompting instead of translation generate similar or more coherent stories and display more creativity, but perform worse in terms of generalization to unseen data. In addition to the potential educational applications of the automated story generation, our approach has the potential to be used as the foundation for more data-efficient low-resource language models.

Story Generation with Large Language Models for African Languages

Understanding written content can vary significantly based on the linguistic complexity of the text. In the context of Amharic, a morphologically rich and low-resource language, the use of complex vocabulary and less frequent expressions often hinders understanding, particularly among readers with limited literacy skills. Such complexity poses challenges for both human comprehension and NLP applications. Addressing this complexity in Amharic is therefore important for text readability and accessibility. In this study, we developed a text complexity annotation tool using curated list of 1,113 complex Amharic terms. Utilizing this tool, we collected and annotated a dataset comprising 20,000 sentences. Based on the annotated corpus, we developed a text complexity classification model using both traditional and deep learning approaches. For traditional machine learning models, the dataset was vectorized using the Bag-of-Words representation. For deep learning and pre-trained models, we implemented embedding layers based on Word2Vec and BERT, trained on a vocabulary consisting of 24,148 tokens. The experiment is conducted using Support Vector Machine and Random Forest for classical machine learning, and Long Short-Term Memory, Bidirectional LSTM, and BERT for deep learning and pre-trained models. The classification accuracies achieved were 83.5% for SVM, 80.3% for RF, 84.1% for LSTM, 85.0% for BiLSTM, and 89.4% for the BERT-based model. Among these, the BERT-based approaches shows optimal performance for text complexity classifications which have ability to capture long-range dependencies and contextual relationships within the text.

Supervised Machine Learning based Amharic Text Complexity Classification Using Automatic Annotator Tool

In this paper, we propose a comprehensive framework for the classification of Swahili news articles using a combination of classical machine learning techniques, deep neural networks, and transformer-based models. By balancing two diverse datasets sourced from Harvard Dataverse and Kaggle, our approach addresses the inherent challenges of imbalanced data in low-resource languages. Our experiments demonstrate the effectiveness of the proposed methodology and set the stage for further advances in Swahili natural language processing.

Swahili News Classification: Performance, Challenges, and Explainability Across ML, DL, and Transformers

The purpose of this work is to share an English-Yorùbá evaluation dataset for openbook reading comprehension with open-ended questions to assess the performance of models both in a high- and a low-resource language. The dataset contains 358 questions and answers on 338 English documents and 208 Yorùbá documents. Experiments show a consistent disparity in performance between the two languages, with Yorùbá falling behind English for automatic metrics even if documents are much shorter for this language. For a small set of documents with comparable length, performance of Yorùbá drops by 2.5 times and this comparison is validated with human evaluation. When analyzing performance by length, we observe that Yorùbá decreases performance dramatically for documents that reach 1500 words while English performance is barely affected at that length. Our dataset opens the door to showcasing if English LLM reading comprehension capabilities extend to Yorùbá, which for the evaluated LLMs is not the case.

Y-NQ: English-Yorùbá Evaluation dataset for Open-Book Reading Comprehension with Open-Ended Questions

This paper presents Yankari, a large-scale monolingual dataset for the Yoruba language, aimed at addressing the critical gap in Natural Language Processing (NLP) resources for this important West African language. Despite being spoken by over 30 million people, Yoruba has been severely underrepresented in NLP research and applications. We detail our methodology for creating this dataset, which includes careful source selection, automated quality control, and rigorous data cleaning processes. The Yankari dataset comprises 51,407 documents from 13 diverse sources, totaling over 30 million tokens. Our approach focuses on ethical data collection practices, avoiding problematic sources and addressing issues prevalent in existing datasets. We provide thorough automated evaluations of the dataset, demonstrating its quality compared to existing resources. The Yankari dataset represents a significant advancement in Yoruba language resources, providing a foundation for developing more accurate NLP models, supporting comparative linguistic studies, and contributing to the digital accessibility of the Yoruba language.

Yankari: Monolingual Yoruba Dataset

Most of the 40+ languages spoken in Togo are severely under-represented in Natural Language Processing (NLP) resources. We present YodiV3, a comprehensive approach to developing NLP for ten Togolese languages (plus two major lingua francas) covering machine translation, speech recognition, text-to-speech, and language identification. We introduce Eyaa-Tom, a new multi-domain parallel corpus (religious, healthcare, financial, etc.) for these languages. We also propose the Lom metric, a scoring framework to quantify the AI-readiness of each language in terms of available resources. Our experiments demonstrate that leveraging large pretrained models (e.g.NLLB for translation, MMS for speech) with YodiV3 leads to significant improvements in low-resource translation and speech tasks. This work highlights the impact of integrating diverse data sources and pretrained models to bootstrap NLP for under-served languages, and outlines future steps for expanding coverage and capability.

YodiV3: NLP for Togolese Languages with Eyaa-Tom Dataset and the Lom Metric

Large-scale multilingual evaluations, such as MEGA, often include only a handful of African languages due to the scarcity of high-quality evaluation data and the limited discoverability of existing African datasets. This lack of representation hinders comprehensive LLM evaluation across a diverse range of languages and tasks. To address these challenges, we introduce AFROBENCH—a multi-task benchmark for evaluating the performance of LLMs across 64 African languages, 15 tasks and 22 datasets. AFROBENCH consists of nine natural language understanding datasets, six text generation datasets, six knowledge and question answering tasks, and one mathematical reasoning task. We present results comparing the performance of prompting LLMs to fine-tuned baselines based on BERT and T5-style models. Our results suggest large gaps in performance between high-resource languages, such as English, and African languages across most tasks; but performance also varies based on the availability of monolingual data resources. Our findings confirm that performance on African languages continues to remain a hurdle for current LLMs, underscoring the need for additional efforts to close this gap.

AfroBench: How Good are Large Language Models on African Languages?

The fact that everyone with a social media account can create and share content, and the increasing public reliance on social media platforms as a news and information source bring about significant challenges such as misinformation, fake news, harmful content, etc. Although human content moderation may be useful to an extent and used by these platforms to flag posted materials, the use of AI models provides a more sustainable, scalable, and effective way to mitigate these harmful contents. However, low-resourced languages such as the Somali language face limitations in AI automation, including scarce annotated training datasets and lack of language models tailored to their unique linguistic characteristics. This paper presents part of our ongoing research work to bridge some of these gaps for the Somali language. In particular, we created two human-annotated social-media-sourced Somali datasets for two downstream applications, fake news & toxicity classification, and developed a transformer-based monolingual Somali language model (named SomBERTa) -- the first of its kind to the best of our knowledge. SomBERTa is then fine-tuned and evaluated on toxic content, fake news and news topic classification datasets. Comparative evaluation analysis of the proposed model against related multilingual models (e.g., AfriBERTa, AfroXLMR, etc) demonstrated that SomBERTa consistently outperformed these comparators in both fake news and toxic content classification tasks while achieving the best average accuracy (87.99%) across all tasks. This research contributes to Somali NLP by offering a foundational language model and a replicable framework for other low-resource languages, promoting digital and AI inclusivity and linguistic diversity.

Downloads

Next from ACL 2025

Retrieval-Augmented Generation Meets Local Languages for Improved Drug Information Access and Comprehension.

Similar lecture

Adapting High-resource NMT Models to Translate Low-resource Related Languages without Parallel Data