Πολυγλωσσικοί ταξινομητές μέσω νευρωνικής μηχανικής μετάφρασης με έμφαση σε μεθόδους εκμάθησης μέσω μεταφοράς και ομοσπονδιακής μάθησης

Μανιάς, Γεώργιος

Multilingual classifiers based on neural machine translation with an emphasis on transfer and federated learning methods

Doctoral Thesis

Author

Μανιάς, Γεώργιος

Date

2025

Abstract

In recent years, there has been a remarkable increase in the interest of the scientific and business communities in the field of Artificial Intelligence – AI. The ability of modern devices to implement highly sophisticated AI techniques has increased the importance of using and developing applications that will use AI techniques and algorithms to process and analyze modern data. Undoubtedly, progress in AI makes computer systems more human-friendly, which requires the inclusion of natural language as an integral means of broader communication and interface between humans and computers. Today, the primary goal of research in the field of Natural Language Processing – NLP is to enable and facilitate communication between humans and computers without resorting to complex procedures. In particular, the introduction into our daily lives of a large number of applications that fall within the field of Generative Artificial Intelligence – GAI, such as ChatGPT by OpenAI, Copilot by Microsoft, and Google Bard, have introduced revolutionary methods in the fields of NLP, Natural Language Understanding – NLU, and Natural Language Generation – NLG of more human-friendly linguistic content. The latter has also brought significant improvements in human-machine interaction, allowing for faster production of text and other audiovisual material. Modern chatbots, automatic translation machines, and search engines are just some of these applications. These applications apply algorithms and techniques that fall within the field of AI, which is a subfield of AI and is one of the most advanced and growing research and business areas of the last five years. Although AI is not a new field of research, recent technological developments in the context of computer science, the increasing need for meaningful human-computer communication, the availability of big data, high-power computing resources and modern algorithms have triggered an exponential increase in interest in the development of new applications and systems. However, one of the biggest challenges in AI is the inability to generalize systems so that they can be effectively applied to multiple and heterogeneous areas of interest. Emerging Machine Learning – ML techniques and approaches, such as Transfer Learning, aim to address these limitations. However, the ever-increasing production and diversity of data create additional challenges in interoperability and their use in different sectors. Moreover, one of the major problems faced by modern applications in the field of NLP has to do with their inability to generalize and apply to different areas of interest. The modern production of huge volumes of data and the diversity that distinguishes them places certain limitations on their application and interoperability between different sectors, such as health, public administration, industry, etc. Sentiment analysis or entity recognition applications that have been created for a specific sector, e.g. in the healthcare sector, appear to have lower performance if used in a different sector, e.g. sentiment analysis of social media users regarding an electoral process or a specific product. This demonstrates the need to develop techniques that are not limited by the content and domain in which they are developed, but are based on the concepts of interoperability and data federation. Of particular importance is the need to develop systems that support multilingual implementations. Language differences constitute a significant obstacle to the horizontal application of NLP algorithms, as most implementations focus on specific languages, resulting in limited application and performance in different linguistic or cultural contexts. The implementation of multilingual systems is a critical requirement for the creation of interoperable and federated approaches that do not depend on language, thus ensuring their wider use and integration. The need for multilingual implementations is not limited only to addressing language barriers but also contributes to expanding the usefulness of systems at a global level, enhancing accessibility, inclusiveness and equitable development. Systems with multilingual support enhance the ability to manage data from distributed and heterogeneous sources, unifying knowledge and blurring language boundaries, thus offering more accuracy and flexibility in different application areas. ΠΕΡΙΛΗΨΗ ΜΑΝΙΑΣ ΓΕΩΡΓΙΟΣ – ΔΙΔΑΚΤΟΡΙΚΗ ΔΙΑΤΡΙΒΗ 18 Furthermore, the production and management of big data from more and more distributed sources and devices creates the need to develop federated and interoperable approaches that are not influenced and limited by a single language. In recent years, the use of Federated Learning, where devices are used as computing nodes that perform calculations on their local data in order to update a global model, aims to solve the issue of data federation, while simultaneously enhancing the security of user data. Finally, the needs and trends of modern intercultural and multilingual societies increasingly demonstrate the need to create multilingual solutions and applications that will have the ability to operate in a broader federal context and will make full and effective use of Neural Machine Translation – NMT techniques. So far, the research and business community have focused on the application of tools and techniques based on the English language. Few studies and tools target less important languages, such as Greek. For example, despite the continuous advances in the development of NMT algorithms and techniques for sentiment analysis in high-frequency and high-use languages, such as English, few applications and studies have been carried out in less used and more difficult languages, such as Greek. Based on the above challenges, this doctoral thesis focuses on the development of scientific approaches that will promote multilingual interoperability and generalization of AI applications. The goal is to create innovative systems that can manage data from heterogeneous and distributed sources, ensuring effective interaction between humans and machines regardless of language or cultural context, and enhance the decision-making process by providing reliable results and additional knowledge to end users. Through the integration of techniques such as transfer learning, pre-trained language models, federated learning and NMT, the thesis aspires to contribute to addressing contemporary research and technological limitations, laying the foundation for solutions that will enhance fair and inclusive development in the field of AI.

Department

Σχολή Τεχνολογιών Πληροφορικής και Επικοινωνιών. Τμήμα Ψηφιακών Συστημάτων

Number of pages

259

Language

Greek

URI

https://dione.lib.unipi.gr/xmlui/handle/unipi/17474

Collections

Τμήμα Ψηφιακών Συστημάτων

Show full item record

Except where otherwise noted, this item's license is described as
Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα