Description:The C3QG system comprises a pipeline as follows:
Input Layer: Accepts raw text paragraphs, tokenizes them, and processes them into model-compatible formats with instruction prefixes (e.g., "generate a medium difficulty question").
Transformer Model Execution: Uses FLAN-T5 pretrained and fine-tuned on domain-specific QG datasets to generate multiple candidate questions through beam search or top-p sampling.
Post-Processing Module: Filters duplicate, ungrammatical, or redundant outputs using confidence thresholds and paraphrasing detection.
Attribution Engine: Applies SHAP to the encoder-decoder attention weights to compute feature importance scores for each token.
Output Formatting: Returns a ranked list of generated questions with SHAP attribution maps and semantic metrics.
Technical Advantages:
• Transparent and controllable question generation
• Compatible with a wide range of educational texts
• Enables human-in-the-loop validation through token-level attribution
• Adaptive difficulty tuning for different learner levels
• Quantitative evaluation via semantic-aware metrics

Example Use-Case:
Consider a high-school learning platform integrated with C3QG. A teacher uploads a batch of 9th-grade biology textbook paragraphs. The system generates multiple-choice and short-answer questions in real time, each tailored to Bloom’s taxonomy levels (e.g., "Understand", "Apply"). Students can select difficulty modes, and for each question, a heatmap reveals the most influential words in the passage, enhancing transparency. Teachers receive semantic metrics (e.g., question complexity, alignment score) and can refine question batches for homework or assessment. In a pilot study, C3QG reduced manual workload by 60% while maintaining educational quality.
, C , Claims:We claim:
1. A computer-implemented system for generating contextually controlled and explainable natural language questions comprising:
o a data curation module configured to tokenize, clean, and augment input datasets;
o a transformer-based question generation engine based on fine-tuned FLAN-T5 architecture;
o a multi-metric evaluation module using BERTScore, METEOR, and SHAP values for assessing semantic accuracy, fluency, and interpretability;
o an adaptive difficulty control unit for modifying output question complexity based on learner profiles.
2. The system of claim 1, wherein SHAP-based interpretability highlights token-wise contribution in the generation process.
3. The system of claim 1, wherein the training is optimized using GPU-accelerated hardware including NVIDIA T4 Tensor Cores.
4. A method for generating explainable and efficient questions comprising:
• preprocessing input context through tokenization and noise reduction;
• fine-tuning transformer-based models for question synthesis;
• evaluating generated outputs with semantic similarity and interpretability metrics;
• adjusting output question difficulty dynamically based on predefined learner categories.
5. The method of claim 4, wherein the model architecture integrates hyperparameter optimization routines to ensure computational efficiency.
6. A non-transitory computer-readable storage medium storing instructions which, when executed by one or more processors, cause the system to perform the steps of the method as claimed in any of claims 4 or 5.