Proceedings of the Third Workshop on Resources and Representations for Under-Resourced Languages and Domains (RESOURCEFUL-2025)

Selle kollektsiooni püsiv URIhttps://hdl.handle.net/10062/107107

Sirvi

Nüüd näidatakse 1 - 20 25

The Third Workshop on Resources and Representations for Under-Resourced Languages and Domains (RESOURCEFUL-2025)
(University of Tartu Library, 2025-03) Tudor, Crina Madalina; Debess, Iben Nyholm; Bruton, Micaella; Scalvini, Barbara; Ilinykh, Nikolai; Holdt, Špela Arhar
Federated Meta-Learning for Low-Resource Translation of Kirundi
(University of Tartu Library, 2025-03) Sang, Kyle Rui; Rabbani, Tahseen; Zhou, Tianyi; Tudor, Crina Madalina; Debess, Iben Nyholm; Bruton, Micaella; Scalvini, Barbara; Ilinykh, Nikolai; Holdt, Špela Arhar
In this work, we reframe multilingual neural machine translation (NMT) as a federated meta-learning problem and introduce a translation dataset for the low-resource Kirundi language. We aggregate machine translation models () locally trained on varying (but related) source languages to produce a global meta-model that encodes abstract representations of key semantic structures relevant to the parent languages. We then use the Reptile algorithm and Optuna fine-tuning to fit the global model onto a target language. The target language may live outside the subset of parent languages (such as closely-related dialects or sibling languages), which is particularly useful for languages with limitedly available sentence pairs. We first develop a novel dataset of Kirundi-English sentence pairs curated from Biblical translation. We then demonstrate that a federated learning approach can produce a tiny 4.8M Kirundi translation model and a stronger NLLB-600M model which performs well on both our Biblical corpus and the FLORES-200 Kirundi corpus.
Multi-label Scandinavian Language Identification (SLIDE)
(University of Tartu Library, 2025-03) Fedorova, Mariia; Frydenberg, Jonas Sebulon; Handford, Victoria; Langø, Victoria Ovedie Chruickshank; Willoch, Solveig Helene; Midtgaard, Marthe Løken; Scherrer, Yves; Mæhlum, Petter; Samuel, David; Tudor, Crina Madalina; Debess, Iben Nyholm; Bruton, Micaella; Scalvini, Barbara; Ilinykh, Nikolai; Holdt, Špela Arhar
Identifying closely related languages at sentence level is difficult, in particular because it is often impossible to assign a sentence to a single language. In this paper, we focus on multi-label sentence-level Scandinavian language identification (LID) for Danish, Norwegian Bokmål, Norwegian Nynorsk, and Swedish. We present the Scandinavian Language Identification and Evaluation, SLIDE, a manually curated multi-label evaluation dataset and a suite of LID models with varying speed–accuracy tradeoffs. We demonstrate that the ability to identify multiple languages simultaneously is necessary for any accurate LID method, and present a novel approach to training such multi-label LID models.
"I Need More Context and an English Translation": Analysing How LLMs Identify Personal Information in Komi, Polish, and English
(University of Tartu Library, 2025-03) Ilinykh, Nikolai; Szawerna, Maria Irena; Tudor, Crina Madalina; Debess, Iben Nyholm; Bruton, Micaella; Scalvini, Barbara; Ilinykh, Nikolai; Holdt, Špela Arhar
Automatic identification of personal information (PI) is particularly difficult for languages with limited linguistic resources. Recently, large language models (LLMs) have been applied to various tasks involving low-resourced languages, but their capability to process PI in such contexts remains under-explored. In this paper we provide a qualitative analysis of the outputs from three LLMs prompted to identify PI in texts written in Komi (Permyak and Zyrian), Polish, and English. Our analysis highlights challenges in using pre-trained LLMs for PI identification in both low- and medium-resourced languages. It also motivates the need to develop LLMs that understand the differences in how PI is expressed across languages with varying levels of availability of linguistic resources.
The Application of Corpus-Based Language Distance Measurement to the Diatopic Variation Study (on the Material of the Old Novgorodian Birchbark Letters)
(University of Tartu Library, 2025-03) Afanasev, Ilia; Lyashevskaya, Olga; Tudor, Crina Madalina; Debess, Iben Nyholm; Bruton, Micaella; Scalvini, Barbara; Ilinykh, Nikolai; Holdt, Špela Arhar
The paper presents a computer-assisted exploration of a set of texts, where qualitative analysis complements the linguistically-aware vector-based language distance measurements, interpreting them through close reading and thus proving or disproving their conclusions. It proposes using a method designed for small raw corpora to explore the individual, chronological, and gender-based differences within an extinct single territorial lect, known only by a scarce collection of documents. The material under consideration is the Novgorodian birchbark letters, a set of rather small manuscripts (not a single one is more than 1000 tokens) that are witnesses of the Old Novgorodian lect, spoken on the territories of modern Novgorod and Staraya Russa at the first half of the first millennium CE. The study shows the existence of chronological variation, a mild degree of individual variation, and almost absent gender-based differences. Possible prospects of the study include its application to the newly discovered birchbark letters and using an outgroup for more precise measurements.
Voices of Luxembourg: Tackling Dialect Diversity in a Low-Resource Setting
(University of Tartu Library, 2025-03) Hosseini-Kivanani, Nina; Schommer, Christoph; Gilles, Peter; Tudor, Crina Madalina; Debess, Iben Nyholm; Bruton, Micaella; Scalvini, Barbara; Ilinykh, Nikolai; Holdt, Špela Arhar
Dialect classification is essential for preserving linguistic diversity, particularly in low-resource languages such as Luxembourgish. This study introduces one of the first systematic approaches to classifying Luxembourgish dialects, addressing phonetic, prosodic, and lexical variations across four major regions. We benchmarked multiple models, including state-of-the-art pre-trained speech models like Wav2Vec2, XLSR-Wav2Vec2, and Whisper, alongside traditional approaches such as Random Forest and CNN-LSTM. To overcome data limitations, we applied targeted data augmentation strategies and analyzed their impact on model performance. Our findings highlight the superior performance of CNN-Spectrogram and CNN-LSTM models while identifying the strengths and limitations of data augmentation. This work establishes foundational benchmarks and provides actionable insights for advancing dialectal NLP in Luxembourgish and other low-resource languages.
Assessing the Similarity of Cross-Lingual Seq2Seq Sentence Embeddings Using Low-Resource Spectral Clustering
(University of Tartu Library, 2025-03) Moll, Nelson; Rabbani, Tahseen; Tudor, Crina Madalina; Debess, Iben Nyholm; Bruton, Micaella; Scalvini, Barbara; Ilinykh, Nikolai; Holdt, Špela Arhar
In this work, we study the cross-lingual distance of machine translations through alignment of seq2seq representations over small corpora. First, we use the M2M100 model to collect sentence-level representations of The Book of Revelation in several languages. We then perform unsupervised manifold alignment (spectral clustering) between these collections of embeddings. As verses between translations are not necessarily aligned, our procedure falls under the challenging, but more realistic non-correspondence regime. The cost function associated with each alignment is used to rank the relative (machine) similarity of one language to another. We then perform correspondent alignment over another cluster of languages, this time using FLORES+ parallel NLLB model embeddings. Our experiments demonstrate that the representations of closely-related languages group closely, and are cheap to align (requiring $<$1000 sentences) via our strategy.
From Words to Action: A National Initiative to Overcome Data Scarcity for the Slovene LLM
(University of Tartu Library, 2025-03) Holdt, Špela Arhar; Antloga, Špela; Munda, Tina; Pori, Eva; Krek, Simon; Tudor, Crina Madalina; Debess, Iben Nyholm; Bruton, Micaella; Scalvini, Barbara; Ilinykh, Nikolai; Holdt, Špela Arhar
Large Language Models (LLMs) have demonstrated significant potential in natural language processing, but they depend on vast, diverse datasets, creating challenges for languages with limited resources. The paper presents a national initiative that addresses these challenges for Slovene. We outline strategies for large-scale text collection, including the creation of an online platform to engage the broader public in contributing texts and a communication campaign promoting openly accessible and transparently developed LLMs.
Post-OCR Correction of Historical German Periodicals using LLMs
(University of Tartu Library, 2025-03) Danilova, Vera; Aangenendt, Gijs; Tudor, Crina Madalina; Debess, Iben Nyholm; Bruton, Micaella; Scalvini, Barbara; Ilinykh, Nikolai; Holdt, Špela Arhar
Optical Character Recognition (OCR) is critical for accurate access to historical corpora, providing a foundation for processing pipelines and the reliable interpretation of historical texts. Despite advances, the quality of OCR in historical documents remains limited, often requiring post-OCR correction to address residual errors. Building on recent progress with instruction-tuned Llama 2 models applied to English historical newspapers, we examine the potential of German Llama 2 and Mistral models for post-OCR correction of German medical historical periodicals. We perform instruction tuning using two configurations of training data, augmenting our small annotated dataset with two German datasets from the same time period. The results demonstrate that German Mistral enhances the raw OCR output, achieving a lower average word error rate (WER). However, the average character error rate (CER) either decreases or remains unchanged across all models considered. We perform an analysis of performance within the error groups and provide an interpretation of the results.
VerbCraft: Morphologically-Aware Armenian Text Generation Using LLMs in Low-Resource Settings
(University of Tartu Library, 2025-03) Avetisyan, Hayastan; Broneske, David; Tudor, Crina Madalina; Debess, Iben Nyholm; Bruton, Micaella; Scalvini, Barbara; Ilinykh, Nikolai; Holdt, Špela Arhar
Understanding and generating morphologically complex verb forms is a critical challenge in Natural Language Processing (NLP), particularly for low-resource languages like Armenian. Armenian's verb morphology encodes multiple layers of grammatical information, such as tense, aspect, mood, voice, person, and number, requiring nuanced computational modeling. We introduce VerbCraft, a novel neural model that integrates explicit morphological classifiers into the mBART-50 architecture. VerbCraft achieves a BLEU score of 0.4899 on test data, compared to the baseline's 0.9975, reflecting its focus on prioritizing morphological precision over fluency. With over 99\% accuracy in aspect and voice predictions and robust performance on rare and irregular verb forms, VerbCraft addresses data scarcity through synthetic data generation with human-in-the-loop validation. Beyond Armenian, it offers a scalable framework for morphologically rich, low-resource languages, paving the way for linguistically informed NLP systems and advancing language preservation efforts.
Annotating Attitude in Swedish Political Tweets
(University of Tartu Library, 2025-03) Lindahl, Anna; Tudor, Crina Madalina; Debess, Iben Nyholm; Bruton, Micaella; Scalvini, Barbara; Ilinykh, Nikolai; Holdt, Špela Arhar
There is a lack of Swedish datasets annotated for emotional and argumentative language. This work therefore presents an annotation procedure and a dataset of Swedish political tweets. The tweets are annotated for positive and negative attitude. Challenges with this type of annotation is identified and described. The evaluation shows that the annotators do not agree on where to annotate spans, but that they agree on labels. This is demonstrated with a new implementation of the agreement coefficient Krippendorff's unitized alpha.
On the Usage of Semantics, Syntax, and Morphology for Noun Classification in IsiZulu
(University of Tartu Library, 2025-03) Sayed, Imaan; Mahlaza, Zola; van der Leek, Alexander; Mopp, Jonathan; Keet, C. Maria; Tudor, Crina Madalina; Debess, Iben Nyholm; Bruton, Micaella; Scalvini, Barbara; Ilinykh, Nikolai; Holdt, Špela Arhar
There is limited work aimed at solving the core task of noun classification for Nguni languages. The task focuses on identifying the semantic categorisation of each noun and plays a crucial role in the ability to form semantically and morphologically valid sentences. The work by Byamugisha (2022) was the first to tackle the problem for a related, but non-Nguni, language. While there have been efforts to replicate it for a Nguni language, there has been no effort focused on comparing the technique used in the original work vs. contemporary neural methods or a number of traditional machine learning classification techniques that do not rely on human-guided knowledge to the same extent. We reproduce Byamugisha (2022)’s work with different configurations to account for differences in access to datasets and resources, compare the approach with a pre-trained transformer-based model, and traditional machine learning models that relyon less human-guided knowledge. The newly created data-driven models outperform the knowledge-infused models, with the best performing models achieving an F1 score of 0.97.
First Steps in Benchmarking Latvian in Large Language Models
(University of Tartu Library, 2025-03) Skadina, Inguna; Bakanovs, Bruno; Darģis, Roberts; Tudor, Crina Madalina; Debess, Iben Nyholm; Bruton, Micaella; Scalvini, Barbara; Ilinykh, Nikolai; Holdt, Špela Arhar
The performance of multilingual large language models (LLMs) in low-resource languages, such as Latvian, has been under-explored. In this paper, we investigate the capabilities of several open and commercial LLMs in the Latvian language understanding tasks. We evaluate these models across several well-known benchmarks, such as the Choice of Plausible Alternatives (COPA) and Measuring Massive Multitask Language Understanding (MMLU), which were adapted into Latvian using machine translation. Our results highlight significant variability in model performance, emphasizing the challenges of extending LLMs to low-resource languages. We also analyze the effect of post-editing on machine-translated datasets, observing notable improvements in model accuracy, particularly with BERT-based architectures. We also assess open-source LLMs using the Belebele dataset, showcasing competitive performance from open-weight models when compared to proprietary systems. This study reveals key insights into the limitations of current LLMs in low-resource settings and provides datasets for future benchmarking efforts.
A Simple Audio and Text Collection-Annotation Tool Targeted to Brazilian Indigenous Language Native Speakers
(University of Tartu Library, 2025-03) Polleti, Gustavo Padilha; Cozman, Fabio; Gerardi, Fabricio; Tudor, Crina Madalina; Debess, Iben Nyholm; Bruton, Micaella; Scalvini, Barbara; Ilinykh, Nikolai; Holdt, Špela Arhar
In this paper we present an audio and text annotation tool for native speakers, with a particular focus on Brazilian indigenous languages. Our tool simplifies the process of language resource annotation and employs gamefication techniques typically found in language learning games. Then we describe the annotation tool and present preliminary results for the Bororo language. We discuss the limitations of our tool, highlighting ethical and practical implementation concerns.
DUDU: A Treebank for Ottoman Turkish in UD Style
(University of Tartu Library, 2025-03) Yılandiloğlu, Enes; Siewert, Janine; Tudor, Crina Madalina; Debess, Iben Nyholm; Bruton, Micaella; Scalvini, Barbara; Ilinykh, Nikolai; Holdt, Špela Arhar
This paper introduces a recently released Ottoman Turkish (ota) treebank in Universal Dependencies (UD) style, DUDU. The DUDU Treebank consists of 1,064 automatically annotated and manually corrected sentences. The texts were manually collected from various academic or literary sources available on the Internet. Following preprocessing, the sentences were annotated using a MaCHAMP-based neural network model utilizing the large language model (LLM) architecture and manually corrected. The treebank became publicly available with the 2.14 release, and future steps involve expanding the treebank with more data and refining the annotation scheme. The treebank is the first and only treebank that utilizes the IJMES transliteration alphabet. The treebank not only gives insight on Ottoman Turkish lexically, morphologically, and syntactically, but also provides a small but robust test set for future computational models for Ottoman Turkish.
WikiQA-IS: Assisted Benchmark Generation and Automated Evaluation of Icelandic Cultural Knowledge in LLMs
(University of Tartu Library, 2025-03) Arnardóttir, Þórunn; Einarsson, Elías Bjartur; Ingvarsson Juto, Garðar; Helgason, Þorvaldur Páll; Einarsson, Hafsteinn; Tudor, Crina Madalina; Debess, Iben Nyholm; Bruton, Micaella; Scalvini, Barbara; Ilinykh, Nikolai; Holdt, Špela Arhar
This paper presents WikiQA-IS, a novel question-answering dataset focusing on Icelandic culture and history, along with an automated pipeline for dataset generation and evaluation. Leveraging GPT-4 to create questions and answers based on Icelandic Wikipedia articles and news sources, we produced a high-quality corpus of 2,000 question-answer pairs. We introduce an automatic evaluation method using GPT-4o as a judge, which shows strong agreement with human evaluations. Our benchmark reveals varying performances across different language models, with closed-source models generally outperforming open-weights alternatives. This work contributes a resource for evaluating language models' knowledge of Icelandic culture and offers a replicable framework for creating similar datasets in other cultural contexts.
Automatic Validation of the Non-Validated Spanish Speech Data of Common Voice 17.0
(University of Tartu Library, 2025-03) Hernández Mena, Carlos Daniel; Scalvini, Barbara; Lág, Dávid í; Tudor, Crina Madalina; Debess, Iben Nyholm; Bruton, Micaella; Scalvini, Barbara; Ilinykh, Nikolai; Holdt, Špela Arhar
Mozilla Common Voice is a crowdsourced project that aims to create a public, multilingual dataset of voice recordings for training speech recognition models. In Common Voice, anyone can contribute by donating or validating recordings in various languages. However, despite the availability of many recordings in certain languages, a significant percentage remains unvalidated by users. This is the case for Spanish, where in version 17.0 of Common Voice, 75\% of the 2,220 hours of recordings are unvalidated. In this work, we used the Whisper recognizer to automatically validate approximately 784 hours of recordings which are more than the 562 hours validated by users. To verify the accuracy of the validation, we developed a speech recognition model based on a version of NVIDIA-NeMo’s Parakeet, which does not have an official Spanish version. Our final model achieved a WER of less than 4\% on the test and validation splits of Common Voice 17.0. Both the model and the speech corpus are publicly available on Hugging Face.
FoQA: A Faroese Question-Answering Dataset
(University of Tartu Library, 2025-03) Simonsen, Annika; Nielsen, Dan Saattrup; Einarsson, Hafsteinn; Tudor, Crina Madalina; Debess, Iben Nyholm; Bruton, Micaella; Scalvini, Barbara; Ilinykh, Nikolai; Holdt, Špela Arhar
We present FoQA, a Faroese extractive question-answering (QA) dataset with 2,000 samples, created using a semi-automated approach combining Large Language Models (LLMs) and human validation. The dataset was generated from Faroese Wikipedia articles using GPT-4-turbo for initial QA generation, followed by question rephrasing to increase complexity and native speaker validation to ensure quality. We provide baseline performance metrics for FoQA across multiple models, including LLMs and BERT, demonstrating its effectiveness in evaluating Faroese QA performance. The dataset is released in three versions: a validated set of 2,000 samples, a complete set of all 10,001 generated samples, and a set of 2,395 rejected samples for error analysis.
OCR Error Post-Correction with LLMs in Historical Documents: No Free Lunches
(University of Tartu Library, 2025-03) Kanerva, Jenna; Ledins, Cassandra; Käpyaho, Siiri; Ginter, Filip; Tudor, Crina Madalina; Debess, Iben Nyholm; Bruton, Micaella; Scalvini, Barbara; Ilinykh, Nikolai; Holdt, Špela Arhar
Optical Character Recognition (OCR) systems often introduce errors when transcribing historical documents, leaving room for post-correction to improve text quality. This study evaluates the use of open-weight LLMs for OCR error correction in historical English and Finnish datasets. We explore various strategies, including parameter optimization, quantization, segment length effects, and text continuation methods. Our results demonstrate that while modern LLMs show promise in reducing character error rates (CER) in English, a practically useful performance for Finnish was not reached. Our findings highlight the potential and limitations of LLMs in scaling OCR post-correction for large historical corpora.
Beyond a Means to an End: A Case Study in Building Phonotactic Corpora for Central Australian Languages
(University of Tartu Library, 2025-03) Muradoglu, Saliha; Gray, James; Simpson, Jane Helen; Proctor, Michael; Harvey, Mark; Tudor, Crina Madalina; Debess, Iben Nyholm; Bruton, Micaella; Scalvini, Barbara; Ilinykh, Nikolai; Holdt, Špela Arhar
Linguistic datasets are essential across fields: computational linguists use them for NLP development, theoretical linguists for statistical arguments supporting hypotheses about language, and documentary linguists for preserving examples and aiding grammatical descriptions. Transforming raw data (e.g., recordings or dictionaries) into structured forms (e.g., tables) requires non-trivial decisions within processing pipelines. This paper highlights the importance of these processes in understanding linguistic systems. Our contributions include: (1) an interactive dashboard for four central Australian languages with custom filters, and (2) demonstrating how data processing decisions influence measured outcomes.

Sirvi

Viimati lisatud