A Multi-Strategy QA Method Combining with Embedded User Feedback

Question answering (QA) is an important interaction to achieve professional and effective results from the knowledge sources. Most existing work focuses on training the QA model once before release using a single method, but lay less emphasis on combination of multiple method with the interactive feedback data. In this paper, a multi-strategy QA method combining with embedded user feedback is proposed, for which the main goal is to augment the comprehension of various query. By leveraging historical feedback question answering data to fine-tune the embedding model, the question answering system enhances its understanding of user needs and preferences, which contributes to improve retrieval or generation results. Considering the complexity and diversity of the query content, the responses can be generated by utilizing large language model (LLM) combined with specific domain knowledge documents. Experiments on the export tax rebate datasets show that the multi-strategy QA method provides more accurate and specific results than other open-source pre-trained large models, which illustrates that the proposed method is effective and practical for specific domain question answering.