Improved Framework Model to Train and Evaluate Difficulty of Interview Question Using Generative AI

Research on question difficulty estimation is in demand for accurate assessments tools and other models for different purposes. Recent years include many implementations on this topic based on BERT[Bidirectional Encoder Representations], RNN[recurrent neural networks] and other classification models. Traditional methods for question difficulty estimation have primarily focused on linguistic and structural analysis and trained on large pre-labeled datasets of questions and their difficulty level. This paper presents an approach, which combines these conventional techniques with Generative AI, for more accurate question difficulty estimation. The principle behind our method is that as the AI system dives deeper into documentation to formulate questions, the questions generated are likely to be more complex or rare and, therefore, deemed more difficult. By utilizing this multi-dimensional approach, our research aims to refine the question difficulty estimator to better reflect the true complexity of technical questions. Finally, this model is anticipated to contribute to more tailored and better evaluations for CSE domain interviews.