Towards Controlled AQG: A Hybrid Approach Based on Verbal Valence (LVF) and First-Order Logic

This dataset supports the reproducibility of the experimental results reported in the paper "Towards Controlled Automatic Question Generation: A Hybrid Approach Based on Verbal Valence (LVF) and First-Order Logic" (WANDJI Alain Chamfort, MOUKOUOP NGUENA Ibrahim, University of Yaoundé I, Cameroon). The dataset contains four files: 1. dataset1_questions_500.csv — 495 questions automatically generated in French from 120 LVF/RALI verb entries across 15 semantic classes. Each question is described by its predicate arity (A), Bloom's Taxonomy level (B), word count (W), Complexity Score (SCQ = 0.4×A + 0.6×B), and Estimated Response Time (ERT = W×0.5 + A×B×5.0 seconds). 2. dataset2_response_times.csv — 59,400 response-time records collected from 120 students answering the 495 generated questions. Includes predicted ERT, observed response time (seconds), and correctness of the answer. 3. dataset3_expert_ratings.csv — Difficulty ratings provided by 3 pedagogical experts on a 100-question pilot corpus, used to calibrate the SCQ weighting coefficients (ω₁=0.4, ω₂=0.6). Inter-rater agreement: Cohen's κ = 0.66. 4. dataset4_lvf_verbs.csv — The 120 LVF verb entries used in generation, with their semantic class, predicate arity, and applicable Bloom levels. All data were generated with Python 3.10 (random seed = 42) and are fully reproducible. Key validation metrics: R²(ERT, mean observed time) = 0.999; R²(SCQ, mean expert rating) = 0.984.