Generative Interpretation: Toward Human-Like Evaluation for Educational Question-Answer Pair Generation

Through experimental analysis, it is revealed that GI outperforms existing evaluation methods in terms of human alignment, and even shows comparable performance with GPT3.5, only with BART-large.