Social networking service (SNS) posts increasingly influence urban travel choices, but they are challenging to use directly for multi stop itinerary planning because the information is informal, unstructured, and often distributed across mixed language captions. In train-oriented cities, users may find several places in a single post yet still face the burden of identifying locations, structuring them as geospatial data, and deciding an efficient visiting order.

This talk presents an open-general workflow that addresses both problems. First, we use and adapt LLM based extraction to convert unstructured SNS posts (in this talk we focus on extraction of café spots) into structured geospatial point of interest data. The workflow is designed for Japanese and English Instagram style captions. We introduce Segmentation Aware Geospatial Extraction (SAGE), a segmentation based extraction prompting technique, to identify multiple cafe locations from noisy SNS text. This method addresses structural failure modes that prompting alone cannot resolve, especially the segmentation of long, mixed language captions containing multiple points of interest. Second, we use dynamic programming to compute the best visiting order over the extracted locations in a timetable aware rail network. The routing stage is formulated as a traveling salesperson type problem, which allows us to guarantee the optimal route for the extracted set under exact solving conditions. In particular, the routing stage maps extracted points of interest to nearby train stations and builds a generalized travel time matrix that combines walking access with timetable-based train travel. Because train services follow scheduled departures and arrivals, the underlying network is treated as a time dependent graph in which edge availability and travel cost vary with time. The itinerary is then solved as an asymmetric traveling salesperson type problem, where non symmetric costs arise from transfer penalties, service frequencies, and direction dependent travel times. Using an exact optimization method such as dynamic programming or mixed integer programming, the framework guarantees the optimal route for the extracted set under the defined travel cost assumptions.

The talk emphasizes how open source LLMs can be used effectively not only for text understanding but also for practical geospatial data structuring, and how optimization provides a principled foundation for reliable itinerary generation. The overall contribution is a reproducible open geospatial workflow connecting SNS text, structured point of interest extraction, station mapping, and optimal rail itinerary planning.