DEIR: Dialogue Generation Enhancement via Implicit Intent Detection and Retrieval Models

Addressing the limitations of current dialogue systems in implicit intent understanding and knowledge accuracy, this paper proposes a Dialogue generation Enhancement framework based on Implicit intent detection and Retrieval models (DEIR). This framework departs from the traditional explicit intent classification paradigm by indirectly identifying users' latent needs through an optimized Sentence-BERT model that retrieves "semantically similar questions." It further integrates a lightweight language model with a structured knowledge block backtracking mechanism to enhance the information density and logical coherence of dialogue generation. Our research constructed a knowledge base containing 410,000 knowledge blocks and optimized the retrieval model using 1.33 million high-quality positive and negative samples。Systematic evaluation on the Wizard of Wikipedia dataset demonstrates that DEIR significantly outperforms Standard RAG and advanced baseline models such as HyDE and Rewrite-RAG across six dimensions, including relevance, extensibility, and informativeness, improving comprehensive performance by 7.3%, 3.1%, and 3.2% respectively.Ablation experiments further validate the necessity of the retrieval-enhanced module and the knowledge block backtracking mechanism. The DEIR framework offers an innovative and engineerable solution to the challenges of large language model hallucination and implicit intent understanding.