HIYO-encoder: a dual-retrieval model based on expanded question generation for addressing hallucinations in question answering

Large language model (LLM) performance has shown promise using retrieval-augmented generation (RAG) techniques. But conventional RAG methods can rely mostly on direct user searches, so neglecting complex semantic links between the query and retrieval text, so reducing retrieval accuracy and raising hallucination risk. To tackle this, we propose HIYOEncoder, a plug-and-play adaptive encoding method leveraging higher query generating to improve retrieval accuracy. By means of large language models, HIYO-Encoder generates expanded searches and develops semantic centers for both user queries and retrieval databases, hence enhancing the alignment between the retrieved content and the original query. This twin-search approach lowers hallucinations and maximizes retrieval performance. We evaluate HIYO-Encoder on question-answering benchmarks, demonstrating its effectiveness in reducing hallucination rates and improving retrieval quality. Additionally, we conduct ablation studies to validate the contributions of individual components.