E-mail Classification Using Traditional Deep Learning Models and Large Language Models

Nowadays, the reliance on electronic mail (E-mail) as a communication channel has increased significantly, especially in businesses. Companies receive an enormous volume of E-mail daily. Therefore, effective E-mail management has become a critical challenge. This study proposes an E-mail classification approach with the primary aim of automating the process of categorizing E-mails into three relevant departments. The utilized dataset in this study is created, collected, and labeled by the team members. The dataset includes three main categories: customer service, information technology (IT), and human resources (HR). Additionally, the dataset is preprocessed using Natural Language Processing (NLP) techniques. Specifically, the word2vec method is used as the word embedding technique in the data preprocessing phase. Consequently, four models are employed: Gemini and Flan-T5 which are Large Language Models (LLM) along with Bidirectional Gated Recurrent Unit (BiGRU) and Bidirectional Long-Short Term Memory (BiLSTM) which are traditional deep learning models. The results of the study showed that the BiLSTM model outperformed the other models with 85.24% accuracy, followed by the BiGRU model with 83.61%, then the Flan model with 81.67% and lastly the Gemini model with 79.00%. The proposed model will fill the gap and support businesses by effectively automating the process of managing E-mails resulting in improved data-driven decisions and optimized reply process.