A Coarse-to-Fine Model for Geolocating Chinese Addresses

Abstract

Address geolocation aims to associate address texts to the geographic locations. In China, due to the increasing demand for LBS applications such as take-out services and express delivery, automatically geolocating the unstructured address information is the key issue that needs to be solved first. Recently, a few approaches have been proposed to automate the address geolocation by directly predicting geographic coordinates. However, such point-based methods ignore the hierarchy information in addresses which may cause poor geolocation performance. In this paper, we propose a hierarchical region-based approach for geolocating Chinese addresses. We model the address geolocation as a Sequence-to-Sequence (Seq2Seq) learning task, that is, the input sequence is a textual address, and the output sequence is a GeoSOT grid code which exactly represents multi-level regions covered by the address. A novel coarse-to-fine model, which combines BERT and LSTM, is designed to learn the task. The experimental results demonstrate that our model correctly understands the Chinese addresses and achieves the highest geolocation accuracy among all the baselines.

Knowledge Graph Representation of Multi-Source Urban Storm Surge Hazard Information Based on Spatio-Temporal Coding and the Hazard Events Ontology Model

“…In the future, we need to add more microblog text data to extract more disaster entities, location entities, and time entities from them. We will also develop a deep learning-based Seq2Seq model to achieve the conversion of location entities into GeoSOT spatial coding, based on [38]. We aim to explore the integration of remote sensing datasets with our existing static spatial data, aiming to enhance the spatial precision and achieve a finer scale in hazard modeling.…”

A Hierarchy-Aware Geocoding Model Based on Cross-Attention within the Seq2Seq Framework

“…To explore the effectiveness of the model and the contribution of each component, we conducted ablation experiments on key modules, including the cross-attention mechanism and layer normalization. Furthermore, to enable meaningful comparisons with the method in [48], we focused on core module differences, given our adoption of more advanced backbone and grid partitioning techniques, along with the integration of residual connections and layer normalization modules. Specifically, we replaced the HGAM's cross-attention module with the self-attention module used in [40] to achieve a fair comparison.…”

“…Given the limited character categories used in the grid label sequence, we posit that the character-by-character prediction of the Seq2Seq model effectively avoids the problem of dimensionality explosion, thereby enhancing geocoding performance. Qian et al [48] combined a GeoSOT grid-division system with a sequence-tosequence framework to design a coarse-to-fine model to solve text geolocation problems. However, the Z-order curve used by GeoSOT [49] suffers from a local order mutation at its zigzag corners.…”

We Know You Are Living in Bali: Location Prediction of Twitter Users Using BERT Language Model

“…In contrast to these previous studies, in our work, we performed location prediction using tweets written by users in Indonesia; therefore, our study is generally specific to Indonesian-language tweets. We then implemented a pretrained bidirectional encoder representations from transformers (BERT) language model to predict geolocation, adopting an approach used by Qian et al [35] for Chinese tweets and Scherrer et al [36] for Helsinki-Ljubljana tweets.…”