TESTING NETWORK CLUSTERING ALGORITHMS WITH NATURAL LANGUAGE PROCESSING

Ixandra Achitouv^*David ChavalariasBruno Gaume

• Centre National de la Recherche Scientifique (CNRS), Paris, France

We propose a hybrid methodology to evaluate the alignment between structural communities inferred from interaction networks and the linguistic coherence of users' textual production in online social networks. Using Twitter data on climate change discussions, we compare different Community Detection Algorithms (CDAs) by training Natural Language Processing Classification Algorithms (NLPCA), such as BERTweet-based models, on the communities they generate. The classification accuracy serves as a proxy for how semantically coherent the CDA-induced groups are. Rather than assuming CDA outputs as ground truth, our approach uses NLPCA accuracy as a relative scoring function to rank CDAs based on their alignment with linguistic identity. This comparative framework provides a self-consistent evaluation of community coherence without requiring manually annotated labels. Our key results show that the best CDA/NLPCA pairs can predict a user's community with over 85% accuracy using only three succinct sentences. We also introduce a coverage-precision trade-off metric to assess community-level performance. Limitations include potential noise in CDA-generated labels and the need for deeper robustness checks.