Flexible Directional Relations via Multi‐Label Classification

ABSTRACT The semantics of direction inherently carry ambiguity, and individuals often exhibit significant differences in their understanding and expression of direction, for example, different people may use different directional relations like West or Northwest to relate two locations. Modeling these personalized preferences can not only improve the accuracy of directional semantics understanding but also enhance the human–computer interaction experience. Existing models rely on fixed and elaborate rules and strategies to calculate directional relations from geometric attributes of geographical entities, making them hard to adapt to different users or scenarios. To resolve this issue and derive directional relations in a more flexible way, we propose employing feature engineering and statistical learning methods to uncover implicit directional preferences in data as a multi‐label classifier. Experimental results demonstrate that this classifier significantly outperforms traditional rule‐based methods in terms of adaptability and capturing individual preferences.