Prediction of customers' interests with sentiment analysis from e-commerce data in Arabic,Turkish and English

Sentiment analysis has achieved significant advancements in understanding the structure and meaning of language using Transformer-based models in the field of natural language processing, introducing a new dimension to language processing tasks. However, sentiment analysis studies conducted in different languages present challenges stemming from structural features of the language, cultural differences, and linguistic diversity. In structurally complex languages such as English, Turkish, and Arabic, the performance and suitability of models used for sentiment detection vary significantly depending on their ability to adapt to the unique characteristics of these languages. This thesis examines the performance of Machine Learning, Deep Learning, and Transformer-based models used for sentiment analysis on texts written in structurally complex languages such as English, Turkish, and Arabic. The primary aim is to analyze the adaptation processes of these models to different language features, identify their strengths and weaknesses, and evaluate their environmental impacts to propose sustainable artificial intelligence solutions. Multilingual datasets collected from e-commerce platforms underwent preprocessing steps, including text cleaning, labeling, and preparation. These datasets reflect the structural complexity of the languages and enable the evaluation of Transformer-based models' adaptability to these features. The models were comprehensively compared based on accuracy, processing time, and energy consumption criteria. The findings demonstrate that Transformer-based models provide high accuracy rates but require optimization in terms of energy consumption. The BERT-base-multilingual-cased model stood out with an accuracy rate of 95.1% for Arabic, while ElecTRa, ConvBERTurk, and BERTurk models achieved an accuracy rate of 88% for Turkish analyses. This thesis aims to contribute to technical advancements in multilingual natural language processing applications and raise awareness of sustainable artificial intelligence. Comparisons with the existing literature provide a valuable resource to guide future research.