Rechercher

This paper is a position paper concerning corpus-building strategies in minoritized languages in the Global North. It draws attention to the structure of the non-technical community of speakers, and concretely addresses how their needs can inform the design of technical solutions. Celtic Breton is taken as a case study for its relatively small speaker community, which is rather well-connected to modern technical infrastructures, and is bilingual with a non-English language (French). I report on three different community internal initiatives that have the potential to facilitate the growth of NLP-ready corpora in FAIR practices (Findability, Accessibility, Interoperability, Reusability). These initiatives follow a careful analysis of the Breton NLP situation both inside and outside of academia, and take advantage of preexisting dynamics. They are integrated to the speaking community, both on small and larger scales. They have in common the goal of creating an environment that fosters virtuous circles, in which various actors help each other. It is the interactions between these actors that create qualityenriched corpora usable for NLP, once some low-cost technical solutions are provided. This work aims at providing an estimate of the community’s internal potential to grow its own pool of resources, provided the right NLP resource gathering tools and ecosystem design. Some projects reported here are in the early stages of conception, while others build on decade-long society/research interfaces for the building of resources. All call for feedback from both NLP researchers and the speaking communities, contributing to building bridges and fruitful collaborations between these two groups.

Apertium translation pair for Occitan and French

L'outil VAGO est un système expert de détection du vague lexical qui mesure aussi le degré de subjectivité du discours, ainsi que son niveau de détail. Dans cet article, nous construisons un clone neuronal de VAGO, fondé sur une architecture de type BERT, entraîné à partir des scores du VAGO symbolique sur un corpus de presse française (FreSaDa). L'analyse qualitative et quantitative montre la fidélité de la version neuronale. En exploitant des outils d'explicabilité (LIME), nous montrons ensuite l'intérêt de cette version neuronale d'une part pour l'enrichissement des lexiques de la version symbolique, et d'autre part pour la production de versions dans d'autres langues.

Ce travail présente des contributions récentes à l'effort de doter l'occitan de ressources et outils pour le TAL. Plusieurs ressources existantes ont été modifiées ou adaptées, notamment un tokéniseur à base de règles, un lexique morphosyntaxique et un corpus arboré. Ces ressources ont été utilisées pour entraîner et évaluer des modèles neuronaux pour la lemmatisation. Dans le cadre de ces expériences, un nouveau corpus plus large (2 millions de tokens) provenant du Wikipédia a été annoté en parties du discours, lemmatisé et diffusé.

Apertium translation pair for Breton and French

One of the challenges with finetuning pretrained language models (PLMs) is that their tokenizer is optimized for the language(s) it was pretrained on, but brittle when it comes to previously unseen variations in the data. This can for instance be observed when finetuning PLMs on one language and evaluating them on data in a closely related language variety with no standardized orthography. Despite the high linguistic similarity, tokenization no longer corresponds to meaningful representations of the target data, leading to low performance in, e.g., part-of-speech tagging. In this work, we finetune PLMs on seven languages from three different families and analyze their zero-shot performance on closely related, non-standardized varieties. We consider different measures for the divergence in the tokenization of the source and target data, and the way they can be adjusted by manipulating the tokenization during the finetuning step. Overall, we find that the similarity between the percentage of words that get split into subwords in the source and target data (the split word ratio difference) is the strongest predictor for model performance on target data.

This chapter presents a survey of the current state of technologies for the automatic processing of the French language. It is based on a thorough analysis of existing tools and resources for French, and also provides an accurate presentation of the domain and its main stakeholders (Adda et al. 2022). The chapter documents the presence of French on the internet and describes in broad terms the existing technologies for the French language. It also spells out general conclusions and formulates recommendations for progress towards deep language understanding for French.

We present lemmatization experiments on the unstandardized low-resourced languages Low Saxon and Occitan using two machine-learningbased approaches represented by MaChAmp and Stanza. We show different ways to increase training data by leveraging historical corpora, small amounts of gold data and dictionary information, and discuss the usefulness of this additional data. In the results, we find some differences in the performance of the models depending on the language. This variation is likely to be partly due to differences in the corpora we used, such as the amount of internal variation. However, we also observe common tendencies, for instance that sequential models trained only on gold-annotated data often yield the best overall performance and generalize better to unknown tokens.

We study the capabilities of speech processing systems trained simply to predict large amounts of transcripts of audio on the internet. When scaled to 680,000 hours of multilingual and multitask supervision, the resulting models generalize well to standard benchmarks and are often competitive with prior fully supervised results without the need for any dataset specific fine-tuning. When compared to humans, the models approach their accuracy and robustness. We are releasing models and inference code to serve as a foundation for further work on robust speech processing.

Dans l’esprit d’une majorité de Français, les langues dites régionales ne seraient que des « patois », de vulgaires déformations du français, de vagues idiomes tout juste bons à décrire des banalités. Pourquoi devraient-ils s’émouvoir de leur effacement ? Or, tous les linguistes le savent : le basque, le breton, l’alsacien, le corse, le picard et les autres, n’ont rien à envier au français, à l’anglais, à l’arabe ou au mandarin. La seule différence entre les « petites langues » et les autres, c’est que les premières n’ont pas eu la chance de devenir des langues officielles d’un État. Cet ouvrage affiche une ambition assumée : réconcilier la France avec sa diversité. Pour que le français reste notre langue commune, sans devenir notre langue unique.

This paper presents OcWikiDisc, a new freely available corpus in Occitan, as well as language identification experiments on Occitan done as part of the corpus building process. Occitan is a regional language spoken mainly in the south of France and in parts of Spain and Italy. It exhibits rich diatopic variation, it is not standardized, and it is still low-resourced, especially when it comes to large downloadable corpora. We introduce OcWikiDisc, a corpus extracted from the talk pages associated with the Occitan Wikipedia. The version of the corpus with the most restrictive language filtering contains 8K user messages for a total of 618K tokens. The language filtering is performed based on language identification experiments with five off-the-shelf tools, including the new fasttext's language identification model from Meta AI's No Language Left Behind initiative, released in July 2022.

We investigate methods to develop a parser for Martinican Creole, a highly under-resourced language, using a French treebank. We compare transfer learning and multi-task learning models and examine different input features and strategies to handle the massive size imbalance between the treebanks. Surprisingly, we find that a simple concatenated (French + Martinican Creole) baseline yields optimal results even though it has access to only 80 Martinican Creole sentences. POS embeddings work better than lexical ones, but they suffer from negative transfer.

Driven by the goal of eradicating language barriers on a global scale, machine translation has solidified itself as a key focus of artificial intelligence research today. However, such efforts have coalesced around a small subset of languages, leaving behind the vast majority of mostly low-resource languages. What does it take to break the 200 language barrier while ensuring safe, high quality results, all while keeping ethical considerations in mind? In No Language Left Behind, we took on this challenge by first contextualizing the need for low-resource language translation support through exploratory interviews with native speakers. Then, we created datasets and models aimed at narrowing the performance gap between low and high-resource languages. More specifically, we developed a conditional compute model based on Sparsely Gated Mixture of Experts that is trained on data obtained with novel and effective data mining techniques tailored for low-resource languages. We propose multiple architectural and training improvements to counteract overfitting while training on thousands of tasks. Critically, we evaluated the performance of over 40,000 different translation directions using a human-translated benchmark, Flores-200, and combined human evaluation with a novel toxicity benchmark covering all languages in Flores-200 to assess translation safety. Our model achieves an improvement of 44% BLEU relative to the previous state-of-the-art, laying important groundwork towards realizing a universal translation system. Finally, we open source all contributions described in this work, accessible at https://github.com/facebookresearch/fairseq/tree/nllb.