Rule-constrained NER & GPT: automated geolocation mapping from contributions on digital participation platforms

Digital participation platforms enhance the efficiency of handling public concerns by aggregating citizens' contributions along with corresponding location tags. However, two key challenges persist: (1) frequent discrepancies between the locations as described in the contribution text and the locations marked on the map and (2) the limited capability of the Named Entity Recognition (NER) process in mapping geolocations. This article proposes a novel framework that takes citizens' contributions, including associated annotated geometries, as input, and outputs the identified geolocations and geometry mapping results from either rule-constrained NER models or Generative Pretrained Transformer (GPT). This enables a comparison between the original geometries as indicated by citizens and the results generated by our framework based on the contribution text. We selected contributions from the Digital Participation System (DIPAS) in Hamburg, Germany, for experimental analysis. The experimental results show that, out of 100 randomly selected contributions, the mapping accuracy of the original geometries is 60%, that of the confirmed rule-constrained NER model is 77%, and that of GPT is 50%. The rule-based NER model performs the best in our case. These findings suggest that our approach could serve as a geolocation mapper for digital participation platforms, such as DIPAS, potentially alleviating the need to calibrate contributions manually.