Enhancing Topic Modeling Through Adaptive Similarity Thresholding: The Case of Large-Scale Citizen Contributions

Abstract

In this paper, we propose a topic modeling method aimed at identifying the most salient themes in a large-scale collection of citizen contributions. Our method explicitly distinguishes between different proposals expressed within the same theme by assigning them to separate clusters, rather than grouping all occurrences into a single cluster. We leverage multiple language models to compute sentence embeddings which are organized into similarity graphs using model-specific similarity thresholds. Community detection algorithms are then applied to these graphs to identify semantically coherent clusters. We conduct experiments to determine the optimal threshold for each model and to evaluate different combinations of embedding models and community detection algorithms in order to identify the most effective configuration. The proposed configurations are evaluated on two large-scale corpora of citizen contributions and compared with state-of-the-art baselines using a comprehensive set of evaluation metrics, including topic modeling measures (topic diversity, NPMI) and clustering validity indices (Davies-Bouldin, intra- and inter-cluster distances). The results show that the selected configuration improves cluster quality, outperforms state-of-the-art baselines, and more accurately captures the diversity of proposals within each theme, providing an effective framework for large-scale thematic analysis.