Souhail Bakkali

Apprentissage inter-modal et unifié de la vision et du langage pour la compréhension multimodale des documents: Les modèles développés dans cette thèse sont le résultat d'un processus itératif d'analyse et de synthèse entre les théories existantes et nos études réalisées. Plus spécifiquement, nous souhaitons étudier l'apprentissage inter-modal pour la compréhension contextualisée sur les composants des documents à travers le langage et la vision. Cette thèse porte sur l'avancement de la recherche sur l'apprentissage inter-modal et apporte des contributions sur quatre fronts : (i) proposer une approche inter-modale avec des réseaux profonds pour exploiter conjointement les informations visuelles et textuelles dans un espace de représentation sémantique commun afin d'effectuer et de créer automatiquement des prédictions sur les documents multimodaux; (ii) à étudier des stratégies concurrentielles pour s'attaquer aux tâches de classification de documents intermodaux, de récupération basée sur le contenu et de classification few-shot de documents ; (iii) pour résoudre les problèmes liés aux données comme l'apprentissage lorsque les données ne sont pas annotées, en proposant un réseau qui apprend des représentations génériques à partir d'une collection de documents non étiquetés ; et (iv) à exploiter les paramètres d'apprentissage few-shot lorsque les données ne contiennent que peu d’exemples.

The transformer-based deep learning approach represents the current state-of-the-art in machine translation (MT) research. Large-scale pretrained transformer models produce state-of-the-art performance across a wide range of MT tasks for many languages. However, such deep neural network (NN) models are often data-, compute-, space-, power-, and energy-hungry, typically requiring powerful GPUs or large-scale clusters to train and deploy. As a result, they are often regarded as “non-green” and “unsustainable” technologies. Distilling knowledge from large deep NN models (teachers) to smaller NN models (students) is a widely adopted sustainable development approach in MT as well as in broader areas of natural language processing (NLP), including speech, and image processing. However, distilling large pretrained models presents several challenges. First, increased training time and cost that scales with the volume of data used for training a student model. This could pose a challenge for translation service providers (TSPs), as they may have limited budgets for training. Moreover, CO₂ emissions generated during model training are typically proportional to the amount of data used, contributing to environmental harm. Second, when querying teacher models, including encoder–decoder models such as NLLB, the translations they produce for low-resource languages may be noisy or of low quality. This can undermine sequence-level knowledge distillation (SKD), as student models may inherit and reinforce errors from inaccurate labels. In this study, the teacher model’s confidence estimation is employed to filter those instances from the distilled training data for which the teacher exhibits low confidence. We tested our methods on a low-resource Urdu-to-English translation task operating within a constrained training budget in an industrial translation setting. Our findings show that confidence estimation-based filtering can significantly reduce the cost and CO₂ emissions associated with training a student model without drop in translation quality, making it a practical and environmentally sustainable solution for the TSPs.

Text recognition is significantly influenced by font types, especially for complex scripts like Khmer. The variety of Khmer fonts, each with its unique character structure, presents challenges for optical character recognition (OCR) systems. In this study, we evaluate the impact of 19 randomly selected Khmer font types on text recognition accuracy using Pytesseract. The fonts include Angkor, Battambang, Bayon, Bokor, Chenla, Dangrek, Freehand, Kh Kompong Chhnang, Kh SN Kampongsom, Khmer, Khmer CN Stueng Songke, Khmer Savuth Pen, Metal, Moul, Odor MeanChey, Preah Vihear, Siemreap, Sithi Manuss, and iSeth First. Our comparison of OCR performance across these fonts reveals that Khmer, Odor MeanChey, Siemreap, Sithi Manuss, and Battambang achieve high accuracy, while iSeth First, Bayon, and Dangrek perform poorly. This study underscores the critical importance of font selection in optimizing Khmer text recognition and provides valuable insights for developing more robust OCR systems.

Developing effective scene text detection and recognition models hinges on extensive training data, which can be both laborious and costly to obtain, especially for low-resourced languages. Conventional methods tailored for Latin characters often falter with non-Latin scripts due to challenges like character stacking, diacritics, and variable character widths without clear word boundaries. In this paper, we introduce the first Khmer scene-text dataset, featuring 1,544 expert-annotated images, including 997 indoor and 547 outdoor scenes. This diverse dataset includes flat text, raised text, poorly illuminated text, distant and partially obscured text. Annotations provide line-level text and polygonal bounding box coordinates for each scene. The benchmark includes baseline models for scene-text detection and recognition tasks, providing a robust starting point for future research endeavors. The KhmerST dataset is publicly accessible at https://gitlab.com/vannkinhnom123/khmerst.