Recently, a few studies have been conducted to construct data-driven job dispatching methods for hybrid flow shop such as semiconductor and display manufacturing systems. The data-driven job dispatching models can be used to simulate the production scheduling or dispatching without knowing the original dispatching rules. To learn historical job dispatching data, they adopt machine learning and deep learning methods for classification, which specifically aim at choosing the dispatched job among candidate jobs. However, such classification-based job dispatching engines often take a long inference time despite high accuracy because they require a lot of comparison between among jobs. This paper proposes a three-stage modelling approach that filters prioritized jobs using a learning-to-rank technique before performing a pairwise comparison. The method was evaluated using two major semiconductor process datasets obtained from a commercial simulation-based scheduling engine. The experimental results demonstrate that the proposed method outperforms the traditional pairwise dispatching model in terms of speed while maintaining high dispatching accuracy.