Geoparsing with Large Language Models: Leveraging the linguistic capabilities of generative AI to improve geographic information extraction

<h2>Geoparsing with Large Language Models</h2> <p>The .zip file included in this repository contains all the code and data required to reproduce the results from our paper. Note, however, that in order to run the OpenAI models, users will required an OpenAI API key and sufficient API credits.</p> <div> <h3>Data</h3> <p>The data used for the paper are in the <code>datasetst</code> and <code>results</code> folders.</p> <ul> <li> <p>**Datasets: **This contains the XML files (LGL and Geovirus) and Json files (News2024) used to benchmark the models. It also contains all the data used to fine-tune the gpt-3.5 model, the prompt templates sent to the LLMs, and other data used for mapping and data creation.</p> </li> <li> <p>**Results: **This contains the results for the models on the three datastes. The folder is separated by dataset, with a single <code>.csv</code> file giving the results for each model on each dataset separately. The <code>.csv</code> file is structured so that each row contains either a predicted toponym and an associated true toponym (along with assigned spatial coordinates), if the model correctly identified a toponym; otherwise the true toponym columns are empty for false positives and the predicted columns are empty for false negatives.</p> </li> </ul> <h3>Code</h3> <p>The code is split into two seperate folders <code>gpt_geoparser</code> and <code>notebooks</code>.</p> <ul> <li>**GPT_Geoparser: **this contains the classes and methods used process the XML and JSON articles (<code>data.py</code>), interact with the Nominatim API for geocoding (<code>gazetteer.py</code>), interact with the OpenAI API (<code>gpt_handler.py</code>), process the outputs from the GPT models (<code>geoparser.py</code>) and analyse the results (<code>analysis.py</code>).</li> <li><strong>Notebooks</strong>: This series of notebooks can be used to reproduce the results given in the paper. The file names a reasonably descriptive of what they do within the context of the paper.</li> </ul> <h3>Code/software</h3> <h3>Requirements</h3> <ul> <li>Numpy</li> <li>Pandas</li> <li>Geopy</li> <li>Scitkit-learn</li> <li>lxml</li> <li>openai</li> <li>matplotlib</li> <li>Contextily</li> <li>Shapely</li> <li>Geopandas</li> <li>tqdm</li> <li>huggingface_hub</li> <li>Gnews</li> </ul> <h3>Access information</h3> <p>Other publicly accessible locations of the data:</p> <ul> <li>The LGL and GeoVirus datasets can also be obtained <a href="https://github.com/milangritta/Pragmatic-Guide-to-Geoparsing-Evaluation" target="_blank" rel="noopener">here<span> (opens in new window)</span></a>.</li> </ul> <h3>Abstract</h3> <div> <p>Geoparsing- the process of associating textual data with geographic locations - is a key challenge in natural language processing. The often ambiguous and complex nature of geospatial language make geoparsing a difficult task, requiring sophisticated language modelling techniques. Recent developments in Large Language Models (LLMs) have demonstrated their impressive capability in natural language modelling, suggesting suitability to a wide range of complex linguistic tasks. In this paper, we evaluate the performance of four LLMs - GPT-3.5, GPT-4o, Llama-3.1-8b and Gemma-2-9b - in geographic information extraction by testing them on three geoparsing benchmark datasets: GeoVirus, LGL, and a novel dataset, News2024, composed of geotagged news articles published outside the models' training window. We demonstrate that, through techniques such as fine-tuning and retrieval-augmented generation, LLMs significantly outperform existing geoparsing models. The best performing models achieve a toponym extraction F1 score of 0.985 and toponym resolution accuracy within 161 km of 0.921. Additionally, we show that the spatial information encoded within the embedding space of these models may explain their strong performance in geographic information extraction. Finally, we discuss the spatial biases inherent in the models' predictions and emphasize the need for caution when applying these techniques in certain contexts.</p> </div> <h3>Methods</h3> <div> <p>This contains the data and codes required to reproduce the results from our paper. The LGL and GeoVirus datasets are pre-existing datasets, with references given in the manuscript. The News2024 dataset was constructed specifically for the paper.&nbsp;</p> <p>To construct the News2024 dataset, we first created a list of 50 cities from around the world which have population greater than 1000000. We then used the GNews python package <a href="https://pypi.org/project/gnews/" target="_blank" rel="noopener">https://pypi.org/project/gnews/<span> (opens in new window)</span></a> to find a news article for each location, published between 2024-05-01 and 2024-06-30 (inclusive). Of these articles, 47 were found to contain toponyms, with the three rejected articles referring to businesses which share a name with a city, and which did not otherwise mention any place names.</p> <p>We used a semi autonmous approach to geotagging the articles. The articles were first processed using a Distil-BERT model, fine tuned for named entity recognicion. This provided a first estimate of the toponyms within the text. A human reviewer then read the articles, and accepted or rejected the machine tags, and added any tags missing from the machine tagging process. We then used OpenStreetMap to obtain geographic coordinates for the location, and to identify the toponym type (e.g. city, town, village, river etc). We also flagged if the toponym was acting as a geo-political entity, as these were reomved from the analysis process. In total, 534 toponyms were identified in the 47 news articles.&nbsp;</p> </div> </div>