Are large-language models enough to generate questionnaires? A use-case with clinical trials

ACL ARR 2026 January Submission309 Authors

Keywords: Question generation, large language models, feature extraction, clinical trials, healthcare

Abstract: We propose DivAC-LLM, a hybrid pipeline that automatically generates physician-facing questionnaires from clinical trial eligibility criteria. The pipeline integrates rule-based systems, shallow machine learning models, fine-tuned biomedical transformers, and prompt-engineered Large Language Models (LLMs). Unlike purely generative approaches, our system explicitly predicts the number of questions derived from each criterion and injects structured cues into LLM prompts to improve control and accuracy. We validate the approach on two datasets: CTQG, a corpus we created from cancer trials, and QG-SQuAD, a modified version of SQuAD for question generation. Experimental results show that combining symbolic and neural components reduces hallucinations and improves semantic alignment. The results highlight that LLMs benefit from lightweight structure and explicit intermediate reasoning when generating interpretable clinical questionnaires.

Paper Type: Long

Research Area: Clinical and Biomedical Applications

Research Area Keywords: healthcare applications,clinical NLP,Question generation,LLM/AI agents,corpus creation

Contribution Types: NLP engineering experiment, Data resources

Languages Studied: English

Submission Number: 309