Towards More Accurate App Review Response Generation with Fine-Grained Filtering

App reviews serve as one of the essential information sources and decision-making bases for developers in the process of software maintenance. Timely and proactive responses to app reviews can effectively enhance user experience for apps and improve the app’s ratings in the app store. However, existing automatic response generation models ignore the importance of timestamps in historical response records and are unable to capture the historical semantic information in the reviews. To address these issues, we propose FFSGen, a novel framework that automatically generates review responses through a fine-grained filtering and semantic retrieval method. FFSGen leverages semantic similarity to retrieve similar samples, efficiently selects valuable samples, and constructs a multi-step instructional prompt as input to employ an LLM to generate responses corresponding to app reviews. Our experimental results demonstrate that the model outperforms the current state-of-the-art model by 5.66% on BLEU-4. The assessment results on METEOR, ROUGE, and human evaluation further validate that our model can generate more accurate response contents compared to existing best-performing models.