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.

We present ongoing research on automatic post-processing approaches to enhance the readability of noisy speech transcripts in low-resource languages, with a focus on conversational speech in Latvian. We compare transformer-based sequence-labeling models and large language models (LLMs) for the standard punctuation and capitalization restoration task, while also considering automatic correction of mispronounced words and disfluency, and partial inverse text normalization. Our results show that very small LLMs (approx. 2B parameters), fine-tuned on a modest text corpus, can achieve near state-of-the-art performance, rivaling orders of magnitude larger LLMs. Additionally, we demonstrate that a fine-tuned Whisper model, leveraging acoustic cues, outperforms text-only systems on challenging conversational data, even for a low-resource language. Error analysis reveals recurring pitfalls in sentence boundary determination and disfluency handling, emphasizing the importance of consistent annotation and domain adaptation for robust post-processing. Our findings highlight the feasibility of developing efficient post-processing solutions that significantly refine ASR output in low-resource settings, while opening new possibilities for editing and formatting speech transcripts beyond mere restoration of punctuation and capitalization.

Open speech corpora of substantial size are seldom available for less-spoken languages, and this was recently the case also for Latvian with its 1.5M native speakers. While there exist several closed Latvian speech corpora of 100+ hours, used to train competitive models for automatic speech recognition (ASR), there were only a few tiny open datasets available at the beginning of 2023, the 18-hour Latvian Common Voice 13.0 dataset being the largest one. In the result of a successful national crowdsourcing initiative, organised jointly by several institutions, the size and speaker diversity of the Latvian Common Voice 17.0 release have increased more than tenfold in less than a year. A successful follow-up initiative was also launched for Latgalian, which has been recognized as an endangered historic variant of Latvian with 150k speakers. The goal of these initiatives is not only to enlarge the datasets but also to make them more diverse in terms of speakers and accents, text genres and styles, intonations, grammar and lexicon. They have already become considerable language resources for both improving ASR and conducting linguistic research. Since we use the Mozilla Common Voice platform to record and validate speech samples, this paper focuses on (i) the selection of text snippets to enrich the language data and to stimulate various intonations, (ii) an indicative evaluation of the acquired corpus and the first ASR models fine-tuned on this data, (iii) our social campaigns to boost and maintain this initiative.

The latest large language models (LLM) have significantly advanced natural language processing (NLP) capabilities across various tasks. However, their performance in low-resource languages, such as Latvian with 1.5 million native speakers, remains substantially underexplored due to both limited training data and the absence of comprehensive evaluation benchmarks. This study addresses this gap by conducting a systematic assessment of prominent open-source LLMs on natural language understanding (NLU) and natural language generation (NLG) tasks in Latvian. We utilize standardized high school centralized graduation exams as a benchmark dataset, offering relatable and diverse evaluation scenarios that encompass multiple-choice questions and complex text analysis tasks. Our experimental setup involves testing models from the leading LLM families, including Llama, Qwen, Gemma, and Mistral, with OpenAI’s GPT-4 serving as a performance reference. The results reveal that certain open-source models demonstrate competitive performance in NLU tasks, narrowing the gap with GPT-4. However, all models exhibit notable deficiencies in NLG tasks, specifically in generating coherent and contextually appropriate text analyses, highlighting persistent challenges in NLG for low-resource languages. These findings contribute to efforts to develop robust multilingual benchmarks and improve LLM performance in diverse linguistic contexts.

This paper presents the Latvian Language Learner Corpus (LaVA) developed at the Institute of Mathematics and Computer Science, University of Latvia. LaVA corpus contains 1015 essays (190k tokens and 790k characters excluding whitespaces) from foreigners studying at Latvian higher education institutions and who are learning Latvian as a foreign language in the first or second semester, reaching the A1 (possibly A2) Latvian language proficiency level. The corpus has morphological and error annotations. Error analysis and the statistics of the LaVA corpus are also provided in the paper. The corpus is publicly available at: http://www.korpuss.lv/id/LaVA.

LNCC is a diverse collection of Latvian language corpora representing both written and spoken language and is useful for both linguistic research and language modelling. The collection is intended to cover diverse Latvian language use cases and all the important text types and genres (e.g. news, social media, blogs, books, scientific texts, debates, essays, etc.), taking into account both quality and size aspects. To reach this objective, LNCC is a continuous multi-institutional and multi-project effort, supported by the Digital Humanities and Language Technology communities in Latvia. LNCC includes a broad range of Latvian texts from the Latvian National Library, Culture Information Systems Centre, Latvian National News Agency, Latvian Parliament, Latvian web crawl, various Latvian publishers, and from the Latvian language corpora created by Institute of Mathematics and Computer Science and its partners, including spoken language corpora. All corpora of LNCC are re-annotated with a uniform morpho-syntactic annotation scheme which enables federated search and consistent linguistics analysis in all the LNCC corpora, as well as facilitates to select and mix various corpora for pre-training large Latvian language models like BERT and GPT.

We propose an approach for generating an accurate and consistent PropBank-annotated corpus, given a FrameNet-annotated corpus which has an underlying dependency annotation layer, namely, a parallel Universal Dependencies (UD) treebank. The PropBank annotation layer of such a multi-layer corpus can be semi-automatically derived from the existing FrameNet and UD annotation layers, by providing a mapping configuration from lexical units in [a non-English language] FrameNet to [English language] PropBank predicates, and a mapping configuration from FrameNet frame elements to PropBank semantic arguments for the given pair of a FrameNet frame and a PropBank predicate. The latter mapping generally depends on the underlying UD syntactic relations. To demonstrate our approach, we use Latvian FrameNet, annotated on top of Latvian UD Treebank, for generating Latvian PropBank in compliance with the Universal Propositions approach.

The paper presents quality focused approach to a learner corpus development. The methodology was developed with multiple design considerations put in place to make the annotation process easier and at the same time reduce the amount of mistakes that could be introduced due to inconsistent text correction or carelessness. The approach suggested in this paper consists of multiple parts: comparison of digitized texts by several annotators, text correction, automated morphological analysis, and manual review of annotations. The described approach is used to create Latvian Language Learner corpus (LaVA) which is part of a currently ongoing project Development of Learner corpus of Latvian: methods, tools and applications.

Text to speech (TTS) systems are necessary for all languages to ensure accessibility and availability of digital language services. Recent advances in neural speech synthesis have eText to speech (TTS) systems are necessary for any language to ensure accessibility and availability of digital language services. Recent advances in neural speech synthesis have enabled the development of such systems with a data-driven approach that does not require significant development of language-specific tools. However, smaller languages often lack speech corpora that would be sufficient for training current neural TTS models, which require at least 30 hours of good quality audio recordings from a single speaker in a noiseless environment with matching transcriptions. Making such a corpus manually can be cost prohibitive. This paper presents an unsupervised approach to obtain a suitable corpus from unannotated recordings using automated speech recognition for transcription, as well as automated speaker segmentation and identification. The proposed method and software tools are applied and evaluated on a case study for developing a corpus suitable for Latvian speech synthesis based on Latvian public radio archive data.nabled the development of such systems with a data-driven approach that does not require much language-specific tool development. However, smaller languages often lack speech corpora that would be sufficient for training current neural TTS models, which require approximately 30 hours of good quality audio recordings from a single speaker in a noiseless environment with matching transcriptions. Making such a corpus manually can be cost prohibitive. This paper presents an unsupervised approach to obtain a suitable corpus from unannotated recordings using automated speech recognition for transcription, as well as automated speaker segmentation and identification. The proposed methods and software tools are applied and evaluated on a case study for developing a corpus suitable for Latvian speech synthesis based on Latvian public radio archive data.

We describe an extensive and versatile lexical resource for Latvian, an under-resourced Indo-European language, which we call Tezaurs (Latvian for ‘thesaurus’). It comprises a large explanatory dictionary of more than 250,000 entries that are derived from more than 280 external sources. The dictionary is enriched with phonetic, morphological, semantic and other annotations, as well as augmented by various language processing tools allowing for the generation of inflectional forms and pronunciation, for on-the-fly selection of corpus examples, for suggesting synonyms, etc. Tezaurs is available as a public and widely used web application for end-users, as an open data set for the use in language technology (LT), and as an API ― a set of web services for the integration into third-party applications. The ultimate goal of Tezaurs is to be the central computational lexicon for Latvian, bringing together all Latvian words and frequently used multi-word units and allowing for the integration of other LT resources and tools.