A GAN-Enhanced Approach to Bengali Intent Classification with Integrated Entity Recognition

In recent years, there has been growing interest in understanding and semantically analyzing a user’s intent on an automated question-answering system based on their input queries. While numerous studies have explored intent classification in high-resource languages like English and other Western languages, Bengali remains underexplored in this domain. Most existing studies employ the transformer encoder architecture, which processes each token of the input text within the full context of surrounding tokens. However, integrating a Generative Adversarial Network (GAN) with the transformer language model incorporated with Named Entity Recognition (NER) extraction holds the potential to significantly enhance performance, narrowing the gap in effectively interpreting user search intent. In this study, we have proposed BnGANBERT-NER, a novel architecture that combines a pre-trained Bengali BERT model with a GAN for robust intent classification. Moreover, the system incorporates a multilingual NER extractor to identify essential entities such as persons, organizations, and locations, enabling a more comprehensive understanding of user queries. Additionally, an in-depth case study has been conducted on our introduced BnNERIntent25 dataset to validate the model’s effectiveness in semantic analysis, showcasing its adaptability and potential for real-world applications where accurate intent interpretation is crucial. The experimental results reveal that our proposed approach substantially improves both the classification quality and the accuracy of identifying user search intent.