An intelligent transformer based framework for bilingual financial complaint classification

The manual processing of large volumes of bilingual consumer complaint data poses significant operational challenges, leading to delayed resolution times and inefficient resource utilization. This study presents the development and evaluation of an intelligent automated system for classifying bilingual financial customer complaints. A comparative analysis was conducted using a balanced synthetic bilingual dataset of 25,000 records derived from the Consumer Financial Protection Bureau database. Traditional machine learning approaches, including Logistic Regression and LightGBM, were implemented using TF-IDF feature representations. In addition, a pre trained multilingual transformer model XLM-RoBERTa was fine-tuned for the classification task. Experimental results demonstrate that the fine-tuned XLM-RoBERTa model outperformed conventional models, achieving an overall accuracy of 88.38%. These findings highlight the effectiveness of contextualized multilingual transformer models in addressing complex bilingual natural language processing tasks, particularly in the domain of financial complaint classification.