Can Large Language Models interpret participatory mapping comments at scale? Evidence from Prague’s emotional mapping

Abstract. Citizen contributions from online participatory mapping platforms surface local knowledge and lived experiences that can inform urban decision-making. However, as these initiatives scale, practitioners increasingly struggle to extract meaningful information. While frameworks and methods to analyse the spatial dimension of these contributions exist, there is a lack of approaches leveraging their free-text content. This paper investigates whether small local Large Language Models (LLMs) can help in structuring this data via (i) sentiment polarity analysis and (ii) thematic classification. Our case study is based on an openly available dataset from 2021 for the city of Prague (N≈30000), where residents’ emotional and perceptual relationships to the city were collected. We evaluated the performance of three models across two languages (Czech and translated English text) on 732 randomly sampled and manually annotated citizen comments, using five repeated runs to account for output variability. Our results revealed that OpenEuroLLM-Czech:12b achieved the highest weighted F1 score for sentiment analysis and Gemma3:12b for thematic classification (85.2% and 70.4%, respectively). All LLMs also exceeded a RoBERTa sentiment classification baseline (weighted F1 score 72.1%) on the sentiment task. Overall, this study suggests a potential for using LLMs in participatory mapping, with implications for more engaging and inclusive participation. We also call for the development of ethical guardrails for the responsible use of LLMs in participatory mapping.