Extracting and recognizing texts from historical maps presents significant challenges due to complex layouts, varied typographic conventions, and the entanglement of multiple sequences. In this paper, we present a modular neural framework for linking and ordering text segments together. This task goes beyond simple word recognition; it enables to recover the complete text sequences. Our solution, based on an Attentive Pointer, successfully manages the presence of distractor words. It leverages both positional and Bézier directional features. We demonstrate the effectiveness of our framework with two practical applications. First, we prove its scalability by applying it to the 1890s Ordnance Survey of London, retrieving 285,846 text sequences. Second, we validate the practical effectiveness of the sequenced placenames by geocoding them and showcasing their capability to automate city maps realignment. Our approach is scalable, trainable, and generic. It supports hierarchical integration and multimodal feature fusion by design, making it an extensible and modular framework for further advancements.