Scalable Ontology-Driven Retrieval with Latency-Aware Gated Fusion for Real-Time Question Answering

Abstract

There is a basic trade-off between knowledge graph-based question answering systems, which are guaranteed to be logically correct under the conditions of accurate parsing and entity linking, and computationally intractable; and neural retrieval, which is computationally fast, but subject to semantic drift and hallucinations. In this paper, the current tension is considered through the introduction of Latency-Aware Gated Fusion (LAGF), a principled approach to dynamically establishing a balance between neural and symbolic inference paths based on estimated query complexity. LAGF predicts the expected latency of symbolic reasoning via a lightweight multi-layer perceptron trained on query-derived features and uses this prediction to adaptively weight contributions from neural and symbolic paths through a learned sigmoid gate. The method operates via four coordinated phases: (1) query-complexity feature extraction, (2) latency prediction via MLP regression (MAE 28ms, R² 0.81), (3) adaptive gating via learned sigmoid, and (4) candidate-set fusion with calibrated scoring. We evaluate on LC-QuAD 2.0 (N=4,441) and QALD-9 (N=150, directional validation). Results show that LAGF achieves 86.8% F1 on LC-QuAD 2.0 with P99 latency of 410ms and 9.4 QPS throughput, statistically significantly outperforming symbolic-only (68.4% F1, p<0.001), neural-only (81.2% F1, p<0.001), and static fusion (84.5% F1, p=0.008) baselines. The error rate of 1.8% Logical Correctness shows great ontology alignment. According to ablation studies, the latency prediction (+2.3 F1), GNN-based soft reasoning (+7.5 F1), and adaptive gating over fixed routing play a critical role. The complete definition of constraint and full reproducibility specifications are presented in anonymous supplementary material.