基於堆疊式集成學習之自適應檢索增強生成方法

Abstract

近年來，大型語言模型（LLMs）在自然語言處理任務中取得了顯著的進展，但受限於其固定的參數記憶，難以處理動態且不斷擴展的知識，導致在回答複雜查詢時容易產生事實錯誤或幻覺。檢索增強生成（RAG）通過整合外部知識來源以改善此問題，但傳統 RAG 無差別地對所有查詢進行檢索，造成對簡單查詢的低效處理與計算資源浪費。為此，本研究提出了一個自適應檢索增強生成（ARAG）方法，基於查詢複雜度動態決定是否進行外部知識檢索。該方法從查詢中提取語義特徵、命名實體識別特徵及頁面瀏覽量特徵，並採用堆疊式集成學習（Stacking Ensemble Learning）訓練一個分類器，以預測查詢是否需要進行外部檢索。實驗結果顯示，本方法在 RetrievalQA、TriviaQA 及 NQ-open 資料集上的分類準確率均優於現有基線方法，展示其有效性與穩健性。

In recent years, Large Language Models (LLMs) have achieved remarkable advances in natural language processing tasks, but their parametric knowledge limits their ability to handle dynamic and evolving knowledge, leading to factual errors or hallucinations when answering complex queries. Retrieval-Augmented Generation (RAG) integrates external knowledge sources to address this limitation, but conventional RAG uniformly performs retrieval regardless of query complexity, resulting in inefficiency for simple queries and unnecessary computational overhead. To address this, we propose an Adaptive Retrieval-Augmented Generation (ARAG) method that dynamically determines whether to perform external retrieval based on query complexity. The proposed method extracts semantic features, named entity recognition features, and page view features from the query, and employs a stacking ensemble learning approach to train a classifier that predicts whether retrieval is necessary. Experimental results show that our method achieves higher classification accuracy compared to baseline methods on the RetrievalQA, TriviaQA, and NQ-open datasets, demonstrating its effectiveness and robustness.