RACCOON: A Retrieval-Augmented Generation Approach for Location Coordinate Capture from News Articles DOI: XXXXXXX.XXXXXXX Conference: Make sure to enter the correct conference title from your rights confirmation emai; June 03–05, 2018; Woodstock, NY ISBN: 978-1-4503-XXXX-X/18/06 CCS: Computing meth

Jonathan Lin, Aditya Joshi, Hye-young Paik, Tri Dung Doung, Deepti Gurdasani · 2025

Read RACCOON: A Retrieval-Augmented Generation Approach for Location Coordinate Capture from News Articles DOI: XXXXXXX.XXXXXXX Conference: Make sure to enter the correct conference title from your rights confirmation emai; June 03–05, 2018; Woodstock, NY ISBN: 978-1-4503-XXXX-X/18/06 CCS: Computing meth by Jonathan Lin, Aditya Joshi, Hye-young Paik, Tri Dung Doung, Deepti Gurdasani (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-11440.txt.

Introduction

Jonathan Lin, Aditya Joshi, Hye-young Paik, Tri Dung Doung, and Deepti Gurdasani built RACCOON in 2025, a tool that finds map coordinates for events reported in news stories. Placing events on a map helps track disease outbreaks and disasters. Doing this by hand is slow, so an automatic method is valuable.

RACCOON looks up candidate places in a location database, then feeds the story and the candidates to a large language model, which picks the coordinates. The team tested it on three data sets, with two models and several rival methods. Removing parts of the system hurt results, which supports the design. They describe it as the first method of its kind for this task.

Catalog record

CREATORJonathan Lin, Aditya Joshi, Hye-young Paik, Tri Dung Doung, Deepti Gurdasani
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE17 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-11440

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

RACCOON: A Retrieval-Augmented Generation Approach for Location Coordinate Capture from News Articles DOI: XXXXXXX.XXXXXXX Conference: Make sure to enter the correct conference title from your rights confirmation emai; June 03–05, 2018; Woodstock, NY ISBN: 978-1-4503-XXXX-X/18/06 CCS: Computing methodologies Natural language processing CCS: Computing methodologies Language resources CCS: Computing methodologies Information extraction CCS: Applied computing Health informatics Jonathan Lin, Aditya Joshi, Hye-young Paik, Tri Dung Doung, Deepti Gurdasani Abstract . Geocoding involves automatic extraction of location coordinates of incidents reported in news articles, and can be used for epidemic intelligence or disaster management. This paper introduces Retrieval-Augmented Coordinate Capture Of Online News articles (RACCOON), an open-source 1 1 1 https://github.com/jonlin223/RACCOON-geocoding geocoding approach that extracts geolocations from news articles. RACCOON uses a retrieval-augmented generation (RAG) approach where candidate locations and associated information are retrieved in the form of context from a location database, and a prompt containing the retrieved context, location mentions and news articles is fed to an LLM to generate the location coordinates. Our evaluation on three datasets, two underlying LLMs, three baselines and several ablation tests based on the components of RACCOON demonstrate the utility of RACCOON. To the best of our knowledge, RACCOON is the first RAG-based approach for geocoding using pre-trained LLMs. 1. Introduction Online content such as news articles and social media posts have been used for applications such as epidemic intelligence ( Joshi et al. 2019 ) . Identifying geographical coordinates corresponding to locations mentioned in the news articles is important for localising incident reports and understanding affected areas for these applications. Geocoding, also known as toponym resolution, is the task of extracting geo-coordinates of location references mentioned in a text 2 2 2 We use the term ‘geocoding’ in the rest of the paper. . Geocoding is challenging due to ambiguity of locations. Location references may share names with non-locations (‘Washington’ can be either a person or place), distinct locations may share names (‘Newcastle’ in the UK and Australia), or locations may have alternate names or abbreviations (‘Royal Melbourne Hospital’ or ‘RMH’ are the same location). Modeled as a typical information extraction task, geocoding has been performed using rule-based, classical machine learning and deep learning-based approaches, as described in the survey by Zhang and Bethard 2024 . However, retrieval-augmented generation (RAG) ( Lewis et al. 2020 ) has been found to improve the accuracy of responses generated by LLMs. Therefore, this paper introduces an approach for geocoding, named as Retrieval-Augmented Coordinate Capture Of Online News articles (RACCOON). The novelty of RACCOON lies in the use of RAG, coupled with techniques such as population re-ranking and candidate retrieval, inspired from past work in geocoding. RACCOON is evaluated on three benchmark datasets, and compared against three baselines (including a rule-based and an LLM prompting-based approach where the latter does not use RAG) via a suite of metrics. Ablation tests on RACCOON include testing alternate LLMs (namely, GPT and Gemini) and the role of other components of RACCOON. Figure 1. Architecture of RACCOON. 2. Related Work Zhang and Bethard 2024 is a survey that decomposes toponym resolution into two categories: ranking and classification approaches. Ranking-based approaches query geospatial databases (gazetteers) and re-rank entries based on specific features, However, these methods often encounter challenges such as limited retriever recall and gazetteer incompleteness ( Grover et al. 2010 ; Wang et al. 2019 ; Halterman 2023 ) . Recent advancements such as GeoNorm ( Zhang and Bethard 2023 ) leverage Transformer-based architectures and employ a two-step retrieval-then-reranking framework. RACCOON further integrates RAG to enhance the performance on geocoding. However, retrieval-based models like GeoNorm face challenges such as retriever recall and gazetteer incompleteness, which RACCOON addresses by incorporating GeoNorm’s two-step method into a Retrieval Augmented Generation (RAG) framework for LLMs. Similarly, GeoNorm involves a train-test setup while RACCOON uses an off-the-shelf pretrained LLM. In contrast, classification approaches make geospatial predictions when the Earth’s surface has been discretised into tiny areas. Geocoding is then models as a classification task with the output being one of these tiles ( DeLozier et al. 2015 ) . While classification approaches do not encounter gazetteer-related issues, they cannot leverage direct geospatial information from gazetteers. By leveraging the pre-trained geospatial knowledge present in LLMs, RACCOON can similarly avoid these gazetteer-related issues. In contrast, ranking-based approaches query a reference geospatial database (gazetteers) and re-rank the entries based on features. Finally, previous studies in geocoding include rule-based ( Grover et al. 2010 ) , machine-learning ( Wang et al. 2019 ) , and deep-learning ( Halterman 2023 ) approaches, with recent models like GeoNorm ( Zhang and Bethard 2023 ) as described above. 3. Methodology RACCOON is implemented using LangChain, and the architecture is shown in Figure 1 . RACCOON uses RAG to connect the GeoNames 3 3 3 https://www.geonames.org/; Accessed on 5th December 2024. gazetteer to an LLM, providing it with specific geospatial knowledge that it might not be able to retrieve when only using an LLM via prompts. RACCOON consists of five key components: (a) country-assisted retrieval, (b) 20 candidate location entries, (c) GeoNames feature types, (d) population heuristic, and (e) state as a context. They are highlighted in Figure 1 , and also indicated in boldface in the subsequent description 4 4 4 Note that, our results show ablation tests on these five components (along with choice of the LLM as a sixth) to validate their value. . Retriever Given a news article as input, a GPT-4o-mini model is used to identify a country for each location mention ( country-assisted retrieval ). The inferred country is passed along with the location mention, and helps to narrow down the candidate locations, as recommended in Zhang and Bethard 2023 . The locations are then searched in a Lucene search engine, based on the BM-25 ranking function. ElasticSearch index forms the background information for the retriever. The index contains location names, coordinates, country and population, English language alternate names, and their abbreviations. RACCOON sorts the retrieved matches using the population heuristic (i.e., based on population), taking 20 top entries as candidate locations. Generation Once the candidate locations have been used, RACCOON uses in-context prompting of the GPT-4o-mini model. The prompt consists of three parts: (a) A sentence containing identifying information about the candidate location such as name, coordinates, along with additional disambiguating features such as GeoNames feature type , country and state ; (B) Location reference; and (C) News article. Specific prompt text can be found in the github repository linked in the abstract. Note that we include the annotated location references from the evaluation datasets in the prompt, assuming that location references are available. This is not unrealistic since NER models can be used to extract location references. GeoVirus GeoWebNews LGL # Articles 229 200 588 # Toponyms 2167 1447 4463 # Unique Considered Locations 685 666 1086 Median Population of Locations 5 5 5 Population figures taken from GeoNames. Locations with no GeoNames annotation excluded. GeoVirus excluded as it is not annotated with GeoNames IDs. N/A 3 522 465 206 922 Table 1. Dataset statistics. 4. Experiment Setup Datasets We evaluated RACCOON on three datasets (statistics in Table 1 ): GeoVirus ( Gritta et al. 2018b ) GeoWebNews ( Gritta et al. 2018a ) and Local-Global

Read the complete text