Reinforcement learning-guided archival question answering with adaptive curriculum learning

Abstract The digitization of archival resources has created unprecedented demands for intelligent retrieval systems capable of handling complex queries across heterogeneous documentary collections. This paper proposes a reinforcement learning-guided question answering model specifically designed for archival question answering scenarios, incorporating a hierarchical reward mechanism that evaluates answer quality across multiple dimensions including factual accuracy, semantic relevance, and provenance clarity. Additionally, an adaptive curriculum learning training strategy is developed to enhance learning efficiency by dynamically adjusting sample difficulty based on model competency progression. The curriculum framework introduces a multi-dimensional query difficulty assessment system encompassing reasoning depth, entity count, coverage scope, and semantic ambiguity. On a Chinese municipal-and-provincial archival corpus of 28,546 query-answer pairs, the proposed approach reaches 78.9% accuracy and an F1 score of 0.824, a margin of 5.7 percentage points over the strongest re-implemented baseline. Adaptive curriculum learning does not shorten per-epoch computation; rather, it accelerates convergence, so that the wall-clock time needed to reach the same validation plateau falls by roughly 28% relative to the ablated variant trained without a curriculum. The gap widens sharply on multi-hop reasoning queries, where the relative gain over the same re-implemented baseline exceeds 17% on this corpus. Because these results rest on a single monolingual archival corpus and on baselines re-implemented under a shared protocol, the evidence should be read as supportive of the proposed design in this setting rather than as an unqualified demonstration of broad applicability.