Abstract: Currently, retrieval-based answer recommendation dominates in professional Q&A communities with extensive corpora. To deal with poor questions suffering from "answer starvation", this study introduces a method using the memory generation adversarial network. It samples similar past questions to enhance the representation of lower-quality queries. An end-to-end memory network refines the candidate answer features. The adversarial network, guided by supervised learning, narrows the gap between these features and the ideal answer traits. Experiments reveal that the proposed model outperforms the baseline, enhancing the NDCG@5 metric by 2.1%, 2.8%, and 3.1% on datasets of Super User, Stack Overflow and Ask Ubuntu, respectively.