We investigate how useful are machine translators based on the fine-tuning of LLMs with very small amounts of training data, typical of extremely low-resource languages such as Indigenous languages. We started by developing translators for the Guarani Mbya and Nheengatu languages by fine-tuning a WMT-19 German-English translator. We then performed a human evaluation of the usefulness of the results of test sets and compared them to their SacreBLUE scores. We had a level of alignment around 60-70\%, although there were about 40\% of very wrong translations.
The results suggest the need of a filter for bad translations as a way to make the translators useful, possibly only in scenarios of human-AI collaboration such as writing-support assistants.

Human Evaluation of the Usefulness of Fine-Tuned English Translators for the Guarani Mbya and Nheengatu Indigenous Languages

Modern natural language processing (NLP) techniques increasingly require substantial amounts of data to train robust algorithms. Building such technologies for low-resource languages requires focusing on data creation efforts and data-efficient algorithms. For a large number of low-resource languages, especially Indigenous languages of the Americas, this data exists in image-based non-machine-readable documents. This includes scanned copies of comprehensive dictionaries, linguistic field notes, children's stories, and other textual material. To digitize these resources, Optical Character Recognition (OCR) has played a major role but it comes with certain challenges in low-resource settings. In this paper, we share the first survey of OCR techniques specific to low-resource data creation settings and outline several open challenges, with a special focus on Indigenous Languages of the Americas. Based on experiences and results from previous research, we conclude with recommendations on utilizing and improving OCR for the benefit of computational researchers, linguists, and language communities.

A Concise Survey of OCR for Low-Resource Languages

Kalaallisut, also known as (West) Greenlandic, poses a number of unique challenges to contemporary natural language processing (NLP). In particular, the language has historically lacked benchmarking datasets and robust evaluation of specific NLP tasks, such as neural machine translation (NMT). 

In this paper, we present a new benchmark dataset for Greenlandic to Danish NMT comprising over 1.2m words of Greenlandic and 2.1m words of parallel Danish translations. We provide initial metrics for models trained on this dataset and conclude by suggesting how these findings can be taken forward to other NLP tasks for the Greenlandic language.

A New Benchmark for Kalaallisut-Danish Neural Machine Translation

The current focus on resource-rich languages poses a challenge to linguistic diversity, affecting minority languages with limited digital presence and relatively old published and unpublished resources. In addressing this issue, this study targets the digitalization of old scanned textbooks written in four Peruvian indigenous languages (Asháninka, Shipibo-Konibo, Yanesha, and Yine) using Optical Character Recognition (OCR) technology. This is complemented with text correction methods to minimize extraction errors. Contributions include the creation of an annotated dataset with 454 scanned page images, for a rigorous evaluation, and the development of a module to correct OCR-generated transcription alignments.

Unlocking Knowledge with OCR-Driven Document Digitization for Peruvian Indigenous Languages

Languages can encode temporal subordination lexically, via subordinating conjunctions, and morphologically, by marking the relation on the predicate. Systematic cross-linguistic variation among the former can be studied using well-established token-based typological approaches to token-aligned parallel corpora. Variation among different morphological means is instead much harder to tackle and therefore more poorly understood, despite being predominant in several language groups. This paper explores variation in the expression of generic temporal subordination ('when'-clauses) among the languages of Latin America and the Caribbean, where morphological marking is particularly common. It presents probabilistic semantic maps computed on the basis of the languages of the region, thus avoiding bias towards the many world's languages that exclusively use lexified connectors, incorporating associations between character <i>n</i>-grams and English <i>when</i>. The approach allows capturing morphological clause-linkage devices in addition to lexified connectors, paving the way for larger-scale, strategy-agnostic analyses of typological variation in temporal subordination.

Mapping 'when'-clauses in Latin American and Caribbean languages: an experiment in subtoken-based typology

This paper outlines the Universal Features tagging of a dependency treebank for Bribri, an Indigenous language of Costa Rica. Universal Features are a morphosyntactic tagging component of Universal Dependencies, which is a framework that aims to provide an annotation system inclusive of all languages and their diverse structures (Nivre et al., 2016; de Marneffe et al., 2021). We used a rule-based system to do a first-pass tagging of a treebank of 1572 words. After manual corrections, the treebank contained 3051 morphological features. We then used this morphologically-tagged treebank to train a UDPipe 2 parsing and tagging model. This model has a UFEATS precision of 80.5 ± 3.6, which is a statistically significant improvement upon the previously available FOMA-based morphological tagger for Bribri. An error analysis suggests that missing TAM and case markers are the most common problem for the model. We hope to use this model to expand upon existing treebanks and facilitate the construction of linguistically-annotated corpora for the language.

Morphological Tagging in Bribri Using Universal Dependency Features

Indigenous languages are historically under-served by Natural Language Processing (NLP) technologies, but this is changing for some languages with the recent scaling of large multilingual models and an increased focus by the NLP community on endangered languages. This position paper explores ethical considerations in building NLP technologies for Indigenous languages, based on the premise that such projects should primarily serve Indigenous communities. We report on interviews with 17 researchers working in or with Aboriginal and/or Torres Strait Islander communities on language technology projects in Australia. Drawing on insights from the interviews, we recommend practices for NLP researchers to increase attention to the process of engagements with Indigenous communities, rather than focusing only on decontextualised artefacts. We hope to encourage dialog about similarities and possible differences in approaches to Indigenous language NLP projects around the world, including in the Americas.

"It's how you do things that matters": Attending to Process to Better Serve Indigenous Communities with Language Technologies

We propose a new paradigm for machine translation that is particularly useful for no-resource languages (those without any publicly available bilingual or monolingual corpora): LLM-RBMT (LLM-Assisted Rule Based Machine Translation). Using the LLM-RBMT paradigm, we design the first language education/revitalization-oriented machine translator for Owens Valley Paiute (OVP), a critically endangered Indigenous American language for which there is virtually no publicly available data. We present a detailed evaluation of the translator's components: a rule-based sentence builder, an OVP to English translator, and an English to OVP translator. We also discuss the potential of the paradigm, its limitations, and the many avenues for future research that it opens up.

LLM-Assisted Rule Based Machine Translation for Low/No-Resource Languages

This article presents an ongoing research on one of the several native languages of the Americas: Amuzgo or jny'on3 nda3 . This language is spoken in Southern Mexico and belongs to the Otomanguean family. Although Amuzgo vitality is stable and there are some available resources, such as grammars, dictionaries, or literature, its digital inclusion is emerging (cf. Eberhard et al. (2024)). In this respect, here is described the creation of a curated dataset in Amuzgo. This resource is intended to contribute the development of tools for scarce resources languages by providing fine-grained linguistic information in different layers: From data collection with native speakers to data annotation. The dataset was built according to the following method: i) data collection in Amuzgo by means of linguistic fieldwork; ii) acoustic data processing; iii) data transcription; iv) glossing and translating data into Spanish; v) semiautomatic alignment of translations; and vi) data systematization. This resource is released as an open access dataset to foster the academic community to explore the richness of this language.

From Field Linguistics to NLP: Creating a curated dataset in Amuzgo language

In this paper we describe our work on Task~2: Creation of Educational Materials. We tried three approaches, but only the third approach yielded improvement over the baseline system. The first system was a fairly generic transformer model. The second system was our own implementation of the edit tree approach from the baseline system. Our final attempt was a version of the baseline system where if no transformation succeeded, we applied transformations from similar morphosyntactic relations. We describe all three here, but, in the end, we only submitted the third system.

The role of morphosyntactic similarity in generating related sentences

We explore threshold vocabulary trimming in Byte-Pair Encoding subword tokenization, a tokenization postprocessing step that replaces rare subwords with their component subwords. The technique is available in popular tokenization libraries but has not been subjected to rigorous scientific scrutiny. While the removal of rare subwords is suggested as best practice in model implementations, both as a means to reduce model size and for improving model performance through robustness, our experiments indicate that, across a large space of hyperparameter settings, vocabulary trimming fails to consistently improve model performance, and is even prone to incurring heavy degradation.

An Analysis of BPE Vocabulary Trimming in Neural Machine Translation

Large Language Models (LLMs) have been reported to outperform existing automatic evaluation metrics in some tasks, such as text summarization and machine translation. However, there has been a lack of research on LLMs as evaluators in grammatical error correction (GEC). In this study, we investigate the performance of LLMs in GEC evaluation by employing prompts designed to incorporate various evaluation criteria inspired by previous research. Our extensive experimental results demonstrate that GPT-4 achieved Kendall's rank correlation of 0.662 with human judgments, surpassing all existing methods. Furthermore, in recent GEC evaluations, we have underscored the significance of the LLMs scale and particularly emphasized the importance of fluency among evaluation criteria.

Underline digital video library

Human Evaluation of the Usefulness of Fine-Tuned English Translators for the Guarani Mbya and Nheengatu Indigenous Languages

A Concise Survey of OCR for Low-Resource Languages

A New Benchmark for Kalaallisut-Danish Neural Machine Translation

Unlocking Knowledge with OCR-Driven Document Digitization for Peruvian Indigenous Languages

Mapping 'when'-clauses in Latin American and Caribbean languages: an experiment in subtoken-based typology

Morphological Tagging in Bribri Using Universal Dependency Features

"It's how you do things that matters": Attending to Process to Better Serve Indigenous Communities with Language Technologies

LLM-Assisted Rule Based Machine Translation for Low/No-Resource Languages

From Field Linguistics to NLP: Creating a curated dataset in Amuzgo language

The role of morphosyntactic similarity in generating related sentences

An Analysis of BPE Vocabulary Trimming in Neural Machine Translation

Large Language Models Are State-of-the-Art Evaluator for Grammatical Error Correction

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES