ESGSenticNet: A Neurosymbolic Knowledge Base for Corporate Sustainability Analysis

Keane Ong, Rui Mao, Frank Xing, Ranjan Satapathy, Johan Sulaeman, Erik Cambria, Gianmarco Mengaldo

Abstract

Evaluating corporate sustainability performance is essential to drive sustainable business practices, amid the need for a more sustainable economy. However, this is hindered by the complexity and volume of corporate sustainability data, not least by the effectiveness of the NLP tools used to analyse them. We identify three primary challenges—immateriality, complexity, and subjectivity—that exacerbate the difficulty of extracting insights from sustainability disclosures. To address these issues, we introduce ESGSenticNet, a publicly available knowledge base for sustainability analysis. Constructed from a neurosymbolic framework integrating specialised concept parsing, GPT-4o inference, and semi-supervised label propagation, ESGSenticNet comprises 44,000 knowledge triplets. Experiments indicate that ESGSenticNet more effectively captures relevant and actionable sustainability information compared to state-of-the-art baselines, outperforming them on ESG relatedness and action orientation by 26% and 31% respectively, while requiring no training or tuning for deployment.

Introduction: The Challenges of ESG Analysis

As companies face mounting pressure to disclose sustainability efforts, the volume of Environmental, Social, and Governance (ESG) reports has exploded. However, automated analysis remains difficult due to three primary "sins" of current NLP approaches:

  1. Irrelevance: General topic models often pick up noise (e.g., "financial year") that lacks sustainability context.
  2. Immateriality: Stakeholders prioritize actions (e.g., "reducing emissions"), but most tools only flag keywords (e.g., "carbon"), missing the context of corporate behavior.
  3. Limited Organization: Extracted topics often fail to align with established global frameworks like GRI or SASB, making data difficult for regulators to utilize.

Comparative Analysis of ESG Analysis Methods

To address these gaps, ESGSenticNet was designed to overcome the limitations of traditional topic models and existing lexicons. The following table highlights how ESGSenticNet addresses the core issues of irrelevance, immateriality, and complexity.

Comparison of ESGSenticNet against traditional topic models and existing lexicons across key performance dimensions.

As shown in the comparison above, while topic models suffer from high complexity and irrelevance, and existing lexicons lack action-oriented depth, ESGSenticNet provides a low-complexity, highly relevant, and structured alternative aligned with real-world reports.

Methodology: A Neurosymbolic Approach

ESGSenticNet employs a neurosymbolic framework, blending the reliability of symbolic AI (rules) with the flexibility of neural AI (LLMs). The construction process involves:

  • ESG Concept Parser (Symbolic): A custom parser uses dependency parsing and part-of-speech tagging to identify "concepts"—specifically verb-noun-noun or adjective-noun combinations (e.g., "minimize paper waste").
  • Quality Control & Seed Annotation (Neural): GPT-4o filters parsed phrases for coherence and labels a subset according to a hierarchical ESG taxonomy, acting as a "sustainability expert."
  • Label Propagation (Graph-based): To scale the knowledge base, labels from "seed" concepts are propagated across a semantic graph to similar unlabelled concepts using a graph-based algorithm.

The resulting knowledge base consists of triplets in the format: (concept, relation, ESG topic), such as ("halve carbon emission", supports, "emissions control").

Experimental Evaluation

The study utilized 1,998 sustainability reports from 589 Singapore Exchange (SGX)-based companies (2015–2023). ESGSenticNet was benchmarked against classic models (NMF, LDA), neural topic models (BERTopic, ECRTM), and existing ESG dictionaries (Baier, Naiara).

Key Results

ESGSenticNet significantly outperformed all baselines across specialized metrics:

  • ESG Relatedness: Achieved 76% (a 26% improvement), ensuring identified terms are truly ESG-relevant.
  • ESG Action Orientation: Reached 84% (a 34% improvement), proving it captures meaningful corporate efforts rather than just buzzwords.
  • Diversity: Identified 2,555 unique ESG terms using "Flexible" matching, which matches verbs and noun phrases separately, compared to only 212 in the Baier dictionary.
  • Human Validation: Experts agreed with the system's categorizations with over 86% accuracy.

Conclusion and Future Work

ESGSenticNet offers a scalable, low-complexity solution for non-technical stakeholders to analyze sustainability disclosures. Future work will focus on expanding the knowledge base to include multilingual support, sector-specific materiality assessments, and dynamic updates to reflect evolving global sustainability frameworks and emerging topics like biodiversity.

Summary updated

Abstract

Evaluating corporate sustainability performance is essential to drive sustainable business practices, amid the need for a more sustainable economy. However, this is hindered by the complexity and volume of corporate sustainability data (i.e. sustainability disclosures), not least by the effectiveness of the NLP tools used to analyse them. To this end, we identify three primary challenges - immateriality, complexity, and subjectivity, that exacerbate the difficulty of extracting insights from sustainability disclosures. To address these issues, we introduce ESGSenticNet, a publicly available knowledge base for sustainability analysis. ESGSenticNet is constructed from a neurosymbolic framework that integrates specialised concept parsing, GPT-4o inference, and semi-supervised label propagation, together with a hierarchical taxonomy. This approach culminates in a structured knowledge base of 44k knowledge triplets - ('halve carbon emission', supports, 'emissions control'), for effective sustainability analysis. Experiments indicate that ESGSenticNet, when deployed as a lexical method, more effectively captures relevant and actionable sustainability information from sustainability disclosures compared to state of the art baselines. Besides capturing a high number of unique ESG topic terms, ESGSenticNet outperforms baselines on the ESG relatedness and ESG action orientation of these terms by 26% and 31% respectively. These metrics describe the extent to which topic terms are related to ESG, and depict an action toward ESG. Moreover, when deployed as a lexical method, ESGSenticNet does not require any training, possessing a key advantage in its simplicity for non-technical stakeholders.
BibTeX
@misc{ong2025esgsenticnet,
  author = {Keane Ong and Rui Mao and Frank Xing and Ranjan Satapathy and Johan Sulaeman and Erik Cambria and Gianmarco Mengaldo},
  title = {{ESGSenticNet: A Neurosymbolic Knowledge Base for Corporate Sustainability Analysis}},
  year = {2025},
  eprint = {2501.15720},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2501.15720},
}
Figure