A Preliminary Evaluation of ChatGPT in Requirements Information Retrieval

Context: Recently, many illustrative examples have shown ChatGPT's impressive ability to perform programming tasks and answer general domain questions. Objective: We empirically evaluate how ChatGPT performs on requirements analysis tasks to derive insights into how generative large language model, represented by ChatGPT, influence the research and practice of natural language processing for requirements engineering. Method: We design an evaluation pipeline including two common requirements information retrieval tasks, four public datasets involving two typical requirements artifacts, querying ChatGPT with fixed task prompts, and quantitative and qualitative results analysis. Results: Quantitative results show that ChatGPT achieves comparable or better $F\beta$ values in all datasets under a zero-shot setting. Qualitative analysis further illustrates ChatGPT's powerful natural language processing ability and limited requirements engineering domain knowledge. Conclusion: The evaluation results demonstrate ChatGPT' impressive ability to retrieve requirements information from different types artifacts involving multiple languages under a zero-shot setting. It is worthy for the research and industry communities to study generative large language model based requirements retrieval models and to develop corresponding tools.