Skip to content
Technical University of Munich

AI4BuildingESG

Breaking into the black box of ESG in the building sector: a machine learning approach

Principal Investigators 

Prof. Dr. Bing Zhu
Professorship for Real Estate Development, Department of Civil, Geo and Environmental Engineering, TUM School of Engineering and Design

Prof. Dr. Debarghya Ghoshdastidar
Professorship for Theoretical Foundations of Artificial Intelligence, Department of Informatics, TUM School of Computation, Information and Technology

Motivation and Goals  

Environment, Social, and Governance (ESG) assessment in the building sector remains a black box hampering sustainable development. This project provides AI/ML solutions to (i) solve the lack and missing of ESG data; (ii) perform sentiment analysis of ESG text data; (iii) predict ESG scores for firms and optimize its weighting scheme to resolve the arbitrariness and opaqueness problems; (iv) strategize financial investment by evaluating the performance of firms based on the ESG scores.

Recent Results 

ESG Data Challenge

  • Built ESG-PhraseBank, an ESG benchmark dataset for sentiment analysis
  • Processed corporate data from GRESB on asset and firm levels for method evaluation

ML Solutions

  • Developed ESGBERT, a state-of-the-art sentiment-scoring language model based on Transformers (the model behind ChatGPT) for ESG data using ESG-PhraseBank.
    Result: Outperforms comparable models like FinBERT by >17%.
  • Optimized for a reduced number of parameters in ESGBERT.
  • Built a tool to obtain high-quality data (to be open-sourced)

Work in Progress

  • Financial performance analysis using ESGBERT
  • Build an ML model to predict missing values in the ESG data using the GRESB processed data

Selected documentation

Working Papers

Eisel, Niklas et al. „Towards Better ESG Sentiment Analysis: A Phrase-Structure Model for Detecting Contextual Sentiment”

Sabanayagam, Mahalakshmi et al. „ESGBERT: A Pretrained Language Model for ESG Sentiment Analysis”

To top