Enhancing Query Understanding in Educational Question Answering Systems Through Neural Models

Accurate query type prediction is a critical component in knowledge graph-based Question Answering (QA) systems, enabling efficient query processing and facilitating the generation of query-specific responses. This study focuses on developing a query classification framework tailored for educational QA systems, categorizing queries into Boolean, Factoid, List, and Multi-line types. A curated dataset of labelled queries is constructed and utilized to evaluate the performance of various deep learning architectures, including LSTM, Bi-LSTM, GRU, and Bi-GRU. Early approaches to query-type classification relied primarily on rule-based heuristics, while traditional machine learning classifiers depended on features such as n-grams and syntactic patterns, and more recent neural approaches typically focused on restricted query categories. In contrast, our work introduces a domain-specific, fine-grained query type prediction framework using deep learning approaches. Among the evaluated models, Bi-GRU demonstrated superior generalization and accuracy, achieving a test accuracy of 72%. Comparative analysis with existing works highlights the effectiveness of the proposed approach in capturing diverse query intents within an academic context. Despite its effectiveness, challenges remain in handling ambiguous or incomplete queries, which are addressed through recommendations for multi-label classification and active learning strategies as future enhancements. This research contributes to the development of more robust and adaptable QA systems, particularly in the domain of education.