امين فتحي

The use of AI-based models is widely preva‎lent across various domains. However, constructing a high-quality AI model tailored for specific tasks heavily relies on human expertise, posing a barrier to its widespread application. This issue has underscored the importance of developing approaches and methodologies aimed at automating the search for neural networks with minimal human intervention. Initial efforts in this direction, relying on evolutionary algorithms and reinforcement learning, have been computationally intensive and time-consuming. Innovations and research in automated search primarily focus on three main phases: defining the search space, determining search algorithms, and establishing eva‎luation methods. Moreover, automated search methods for neural networks have predominantly been applied in image processing tasks, while their application in natural language processing tasks such as machine translation, sentiment analysis, and text summarization remains underexplored. This research aims to propose a method for automating architecture search in machine translation tasks, applicable to other natural language processing tasks as well. To achieve this, we require an appropriate search space to ultimately present models with efficient performance and scores suitable for machine translation (specifically from colloquial Persian to English) using a metaheuristic search algorithm. Such a model will effectively compete with large language models facing significant challenges in smaller, commercially viable tasks. The presented method in this thesis demonstrates potential for use in other language processing tasks and is highlighted as an innovation in this field. The search space is designed using MT5 model architectures, encompassing 512 different configurations for models. Additionally, due to hardware limitations, the Bee Colony Optimization algorithm has been employed as the search algorithm. Results on the TEP dataset indicate improved machine translation performance, eva‎luated using the SacreBLEU metric.

جست‌وجوي شبكه‌هاي عصبي , پردازش زبان‌هاي طبيعي , ترجمه ماشيني , يادگيري عميق

Neural architecture search , Natural language processing , Machine translation , Deep learning

amin fathi

dr. nasser mozayani