Classifying Sustainability Reports Using Companies Self-Assessments
Charlott Jakob, Vera Schmitt, Salar Mohtaj, Sebastian Möller
Abstract
To achieve the 17 Sustainable Development Goals (SDGs), it is essential to monitor and measure the contributions of various stakeholders, including companies. However, existing methods for automatically analyzing SDG contributions in sustainability reports (SRs) focus on sentence-based classification, discerning SDG context spread over several sentences. This paper proposes an alternative approach that leverages data annotated within SRs. Using SDG icons that companies themselves included in their report pages, we train a multi-label classifier to detect SDG contributions. To assess the approach’s capability to understand SDG contributions across companies, we determined a better generalization using a higher number of different SRs while maintaining the number of page samples. Furthermore, we compared several transformer-based models applicable to long text and achieved the best F0.5 score of 0.65 using Longformer.
