Overview of the SustainEval 2025 Shared Task: Identifying the Topic and Verifiability of Sustainability Report Excerpts

Jakob Prange, Charlott Jakob, Patrick Göttfert, Raphael Huber, Pia Wenzel Neves, Annemarie Friedrich

Abstract: The SustainEval shared task @ GermEval 2025 aims to analyze text from German sustainability reports. The shared task required solving two tasks: classifying a span’s topic into one of 20 reporting criteria and estimating its verifiability on a scale from 0.0 to 1.0. The spans and their corresponding reporting criteria were retrieved from the DNK database. Furthermore, the spans were manually annotated to assess verifiability. This paper details the data collection process and provides an overview of the baselines, participating systems, and results. The submitted systems explore language-specific bidirectional and left-to-right encoders, combined with data augmentation methods. Ensembled BERT models with different sets of hyperparameters work best for content classification, while for verifiability rating, generative pretraining is competitive as well.