Votre recherche

Latent diffusion approaches to sign language production (SLP) rely on an initial stage that learns an encoding of sign pose sequences, enabling generative modeling in the resulting latent space. The autoencoder used in this stage is typically evaluated in terms of reconstruction quality using geometric metrics common in SLP. While informative, these metrics do not fully capture latent space properties that may influence the training and performance of the downstream generative model. In this work, we investigate how architectural and training objective design choices in a variational autoencoder (VAE) for sign pose encoding affect latent space structure, and how these differences translate into the performance of a latent diffusion model for text-to-sign generation. Our experiments on Phoenix14T dataset show that variations in generative performance, measured through back-translation BLEU scores, can sometimes be better explained by differences in latent space properties than by VAE reconstruction accuracy alone.

This paper presents a fine-tuning approach for automatic speech recognition (ASR) that exploits linguistic similarities between dialects to address data scarcity and imbalance in dialectally rich, low-resource languages. The approach is evaluated on Occitan, a low-resource language composed of six main dialects with unevenly distributed resources. For the better-resourced dialects, Languedocien and Gascon, a two-step fine-tuning is applied to improve recognition performance by exploiting inter-dialectal information. For the less-represented dialects, we assess the effectiveness of knowledge transfer from the better-resourced variants, analyzing the model’s ability to generalize to sparsely represented dialects. The results show that leveraging linguistic proximity between dialects is an effective strategy to enhance ASR performance in low-resource settings with high dialectal variability.

Alsatian is a regional low-resource language spoken in a majority-language context. In order to create a voice dataset suited for training automatic speech recognition and speech-to-text models, we launched a crowdsourcing campaign on the platform Mozilla Common Voice. We describe sociolinguistic issues we ran into, such as participants’ perception of their own language and its role in the AI landscape, which are vital to address to raise the participation in the crowdsourcing effort. We found that the participants are often confused about NLP and AI tools, and have a strong interested in preserving their language.

This paper presents a historical parallel corpus of languages spoken in mainland France. It consists of a collection of versions of the Parable of the Prodigal Son, collected during the 19th century. The paper aims to present the interest of such a corpus, its constitution-through XML/TEI encoding, semi-automatic alignment and projection on linguistic maps-and its potential uses for the study of these low-resource languages.

French minority and regional languages remain under-resourced in digital formats, limiting cultural preservation and computational studies. The COLaF project develops corpora and NLP tools for French and languages spoken in France in synchrony and diachrony. This paper presents its TEI schema, which allows for the representation of linguistic variation across temporal, geographic and sociolinguistic dimensions. Using a combination of <language>, <person> and <place> elements, the schema links text segments to metadata describing speakers, locations and varieties. Glottolog identifiers are used for non-standard languages. This paper advocates for the creation of a non-BCP 47-language attribute in TEI to enable the interoperability and TEI-compliance of corpora working on non-standard languages. The schema supports diverse document types from 17th-century monographs to web forums and has been applied to several corpora.

Automatic Text Recognition (ATR) has become a key component of digital editorial workflows, enabling the conversion of scanned documents into TEI. Beyond text recognition, ATR pipelines rely on Document Layout Analysis (DLA) to segment pages into labeled zones that support document reconstruction. Controlled vocabularies such as SegmOnto have already been widely used for this purpose, notably in projects like Gallicorpora and SETA. This paper presents LADaS (Layout Analysis Dataset with SegmOnto), an extension of SegmOnto designed for deeper alignment with the TEI Guidelines. LADaS introduces a two-level annotation system combining SegmOnto broad layout zones with visually identifiable subzones mapped on the TEI. The paper describes the process of building this controlled vocabulary and its documentation, as well as the associated annotated dataset and a proposed pipeline to convert historical scanned documents into deeper encoded TEI files.

We introduce an open-access and openly-licensed diachronic dataset for Document Layout Analysis (DLA) targeting heterogeneous historical and contemporary materials 4 . The dataset comprises 8,287 manually annotated pages spanning 1600-2024, covering both digitised and born-digital documents across diverse genres including magazines, academic papers, monographs, plays, and administrative records. Annotations follow a two-level vocabulary derived from SegmOnto, extended to paragraph-level semantic regions consistent with the Text Encoding Initiative (TEI), and provide 45 fine-grained classes over 115,007 instances. The data is organised in modular subsets, enabling domain-specific configurations. We establish baselines using two families of object detectors (YOLOv11 and D-FINE) and study three regimes: in-domain training, label-set simplification, and large-scale pretraining (Objects365). We additionally conduct a cross-dataset transfer study against DocLayNet and M 6 Doc using a unified label mapping and a per-decade temporal evaluation. Results show that models trained on existing modern-centric DLA datasets degrade sharply on pre-20th century material (mean F1 dropping by up to 0.25), while in-domain training on our dataset preserves performance across centuries. This highlights the need for diachronic benchmarks and demonstrates the value of combining CV-based detection with DH-oriented annotation standards.

We introduce ForumOccitania, a new Occitan corpus of posts from an online forum, covering a range of topics and dialects. While some existing datasets for this low-resource language include labels of varieties within the dialect continuum, we go one step further by providing metadata pertaining to sociolinguistic factors of language variation (dialect, geographical location, age, proficiency), extracted from self-declared user profiles. We carry out statistical and qualitative analyses, as well as preliminary experiments on unsupervised dialect identification. Our results show that (i) most of the contents is written in Occitan, with the classical spelling conventions, and by young new speakers, (ii) posts display a strong presence of dialectal features from four major Occitan varieties (Lemosin, Lengadocian, Gascon, Provençau), and (iii) a simple topic modelling approach introduced by Kuparinen and Scherrer (2024) effectively detects salient features of these four varieties, but also reveals finer-grained diatopical variation tendencies.

Occitan is a Romance language spoken mostly in the South of France and characterised by rich dialectal variation, which can pose problems for certain NLP tools. This shortfall is largely attributable to the scarcity of dialect-annotated corpora, in a context where linguistic classification within the Occitan dialect continuum is still debated and major nomenclatures, such as ISO 639, fail to provide granular codes for varieties below the generic “Occitan” label. In this paper, we introduce OcWikiDialects, a new dataset comprising articles from the Occitan Wikipedia. The corpus features rich metadata, including dialect labels, and is segmented at both paragraph and sentence levels. Combined with previously released datasets, we explore approaches for Occitan dialect identification by training three types of model on up to 8 labels: linear SVM classifiers based on word and character n-grams, FastText classifiers based on pretrained vectors, and BERT-based neural classifiers adapted through fine-tuning. Evaluations across in- and out-of-domain test sets demonstrate the substantial impact of our new dataset for the task. However, a peak macro-averaged F1 score of 58.15 underscores persistent challenges for underrepresented Occitan varieties, supported by our per-dialect analysis. Code, dataset and models are available: https://github.com/DEFI-COLaF/OcWikiDialects.

Of the over 7,000 languages spoken in the world, commercial language identification (LID) systems only reliably identify a few hundred in written form. Research-grade systems extend this coverage under certain circumstances, but for most languages coverage remains patchy or nonexistent. This position paper argues that this situation is largely self-imposed. In particular, it arises from a persistent framing of LID as decontextualized text classification, which obscures the central role of prior probability estimation and is reinforced by institutional incentives that favor global, fixed-prior models. We argue that improving coverage for tail languages requires rethinking LID as a routing problem and developing principled ways to incorporate environmental cues that make languages locally plausible.

Language identification (LID) is a fundamental step in curating multilingual corpora. However, LID models still perform poorly for many languages, especially on the noisy and heterogeneous web data often used to train multilingual language models. In this paper, we introduce CommonLID, a community-driven, human-annotated LID benchmark for the web domain, covering 109 languages. Many of the included languages have been previously under-served, making CommonLID a key resource for developing more representative high-quality text corpora. We show CommonLID's value by using it, alongside five other common evaluation sets, to test eight popular LID models. We analyse our results to situate our contribution and to provide an overview of the state of the art. In particular, we highlight that existing evaluations overestimate LID accuracy for many languages in the web domain. We make CommonLID and the code used to create it available under an open, permissive license.

For decades, the pidgin-creole life cycle hypothesis was considered the standard account for the development of both pidgins and creoles. Key aspects of this account include the belief that pidgins emerge from communicative necessity in multilingual environments, while creoles are the outcomes of the stabilization of pidgins, especially via nativization. This account has been critiqued from various directions, with French-based Creoles often serving as the crux of the debate. For these languages, a major point of contention has been the similarity between the languages of the Americas and the Indian Ocean despite the apparent absence of a stable precursor pidgin. For example, McWhorter (2000) defends the life cycle model by hypothesizing that French- based Creoles ultimately descend from an unattested 17th century slave castle pidgin. In contrast, Mufwene (2020) maintains that pidgins mostly formed after the development of creoles during the during a colonial transitional period in Africa, Asia, and the Pacific. Despite these canonical debates regarding French-based Creoles, French-based pidgins remain underexplored. While Skirgård (2013) has compiled and analyzed a corpus of Français Tirailleur, a variety closely associated with West African soldiers, there is still debate about whether it represents a “real” contact variety from the late 19th and early 20th centuries Parkvall (2018) or a stereotypical projection of otherness (Vigouroux (2017). In contrast, there is more consensus that Tây Bồi, or Vietnamese Pidgin French, developed in French Indochina during roughly the same time period and fell out of use when the French military presence ended. Although Français Tirailleur and Tây Bồi share many features, they are often assumed to be mostly independent developments, and to our knowledge, no extensive comparison has been conducted. In this work, we first review descriptions of Tây Bồi and Français Tirailleur to highlight their close structural affinity, Then, we introduce textual data showing that they were largely consolidated by 1890. Finally, we revisit some contemporary accounts of language education in French colonies. Taken together, these documents suggest that, far from being spontaneous solutions to communicative barriers, Français Tirailleur and Tây Bồi were heavily shaped by top- down instruction of ostensibly simplified French. Rather than disproving the existence of pidginized varieties, the codevelopment of regionalized Frenches in West Africa and Southeast Asia underscores the contradictions and instability of thecolonial-era two-track approach to language education.

Description du processus de développement et d’anonymisation d’un corpus en occitan issu d’un forum en ligne et riche en métadonnées liées à la variation dialectale.

Automatic Speech Recognition (ASR) has advanced considerably in recent years, largely fueled by breakthroughs in self-supervised learning and the increasing availability of large-scale multilingual speech corpora. Models such as Whisper have demonstrated remarkable performance across numerous languages. However, despite its strong multilingual coverage, performance remains highly variable across languages. In this study, we fine-tune Whisper in four regional languages spoken in France, each selected to represent distinct linguistic and resource conditions: Basque, Occitan, Alsatian, and Shimaore. Performance is increased by fine-tuning for each language, but to significantly varying degrees. This is due to a combination of varying data quantity and quality, language similarity to an already well-covered language, and transcription standardisation: all things that low-resource languages may lack.

Automatic language identification is frequently framed as a multi-class classification problem. However, when creating digital corpora for less commonly written languages, it may be more appropriate to consider it a data mining problem. For these varieties, one knows ahead of time that the vast majority of documents are of little interest. By minimizing resources spent on classifying such documents, we can create corpora covering previously overlooked languages faster than existing pipelines. To demonstrate the effectiveness of the targeted mining perspective, we introduce a new pipeline that can filter a single snapshot in two hours. We also provide web corpora for several French-based Creoles.

Efforts towards better machine translation (MT) for Creole languages have historically been isolated, due to Creole languages' geographic and linguistic diversity. However, most speakers of Creole languages stand to benefit from improved MT for low-resource languages. To galvanize collaboration for Creole MT across the NLP community, we introduce the First Shared Task for Creole Language Machine Translation at WMT25. This Shared Task consists of two systems tracks and one data track, for which we received submissions from five participating teams. Participants experimented with a wide variety of systems and development techniques. Our evaluation campaign gave rise to improvements in MT performance in several languages, and particularly large improvements in new testing genres, though some participants found that reusing subsets of pretraining data for specialized post-training did not yield significant improvements. Our campaign also yielded new test sets for Mauritian Creole and a vast expansion of public training data for two Creole languages of Latin America.

One of the main challenges in neural sign language production (SLP) lies in the high intra-class variability of signs, arising from signer morphology and stylistic variety in the training data. To improve robustness to such variations, we propose two enhancements to the standard Progressive Transformers (PT) architecture (Saunders et al., 2020). First, we encode poses using bone rotations in quaternion space and train with a geodesic loss to improve the accuracy and clarity of angular joint movements. Second, we introduce a contrastive loss to structure decoder embeddings by semantic similarity, using either gloss overlap or SBERT-based sentence similarity, aiming to filter out anatomical and stylistic features that do not convey relevant semantic information. On the Phoenix14T dataset, the contrastive loss alone yields a 16% improvement in Probability of Correct Keypoint over the PT baseline. When combined with quaternion-based pose encoding, the model achieves a 6% reduction in Mean Bone Angle Error. These results point to the benefit of incorporating skeletal structure modeling and semantically guided contrastive objectives on sign pose representations into the training of Transformer-based SLP models.

This paper introduces a preprocessing pipeline for keypoints extracted using MediaPipe, aiming to improve pose annotation consistency in sign language datasets. We evaluate its effectiveness using a sign similarity task based on phonological features, without relying on gloss annotations. Similarity is measured using Dynamic Time Warping (DTW) across videos from sign language dictionaries. Although such similarity analyses can support various sign language processing applications -such as lexical search, clustering, and data enrichment -the main contribution of this work is to standardise pose features across heterogeneous sources, including different signers and backgrounds. Experiments on two dictionary datasets show that our pipeline significantly improves similarity measurements, with promising benefits for other sign language processing tasks.

Cet article dresse un état de l'art de la traduction automatique et de son évaluation pour les langues à variation dialectale, et en particulier pour les continuums dialectaux. Pour illustrer cet état de l'art, nous proposons une série d'expériences préliminaires sur le continuum occitan, afin de dresser un état des performances des systèmes existants pour la traduction depuis et vers plusieurs variétés d'occitan. Nos résultats indiquent d'une part des performances globalement satisfaisantes pour la traduction vers le français et l'anglais. D'autre part, des analyses mélangées à des outils d'identification de langues sur les prédictions vers l'occitan mettent en lumière la capacité de la plupart des systèmes évalués à générer des textes dans cette langue (y compris en zero-shot ), mais révèlent aussi des limitations en termes d'évaluation de la diversité dialectale dans les traductions proposées.

Automatic language identification is frequently framed as a multi-class classification problem. However, when creating digital corpora for less commonly written languages, it may be more appropriate to consider it a data mining problem. For these varieties, one knows ahead of time that the vast majority of documents are of little interest. By minimizing resources spent on classifying such documents, we can create corpora much faster and with better coverage than using established pipelines. To demonstrate the effectiveness of the language mining perspective, we introduce a new pipeline and corpora for several French-based Creoles.