Intent classification for university administrative services using a bidirectional recurrent neural network modified by a developed Kepler optimization algorithm.

The emergence of conversational agents has transformed university-student interactions, as they offer around-the-clock assistance with immediate responses to questions. Yet building such systems is hard, and the challenges are even greater in multilingual environments, where the complexities inherent to natural language understanding (NLU) are magnified. The conventional methods of NLU use separate models for intent classification (IC) and named entity recognition (NER), resulting in even higher computational costs, more memory usage, and reduced overall performance. To tackle these limitations, we propose a new joint model combining IC and NER with a modified Bidirectional Long Short-Term Memory (BiLSTM) network jointly optimized by a newly developed Kepler optimization (DKO) algorithm. Our model can be implemented in the two languages, Greek and English, to provide assistance to University students. Our model combines deep learning with optimization, which enhances the benefits of NLU accuracy at the most efficient level of use for the two languages. We evaluate our joint model with a standard dataset and show that we outperformed state-of-the-art models in terms of accuracy, precision and recall. All together these initial findings demonstrate the potential for joint NLU models or cognitive assistants in an educational context - which implies significant gains for both student experience and as a cost of support for university and other broad administrative improvements and services. In addition, this contribution makes a contribution to the broader area merging education and AI specifically through including multimodal interaction models and new conversational interface development.