LenseESG — Explainable Multimodal ESG Risk Intelligence

ESG ratings can differ significantly between providers because different organisations use different sources, definitions, weightings and methodologies. A company may therefore receive a risk score without decision-makers being able to see clearly which evidence drove it.

This project explored whether structured environmental, social, governance and controversy indicators could be combined with company disclosures and news-derived text to produce a more informative and explainable risk-assessment workflow.

Financial, sustainability and risk teams need more than a prediction. They need to understand the evidence, assumptions and limitations behind an assessment before using it in a decision.

The project used a final modelling dataset covering 542 companies. The dataset is not representative of every global company; it was assembled for research and benchmarking purposes.

• Environmental risk indicators
• Social risk indicators
• Governance risk indicators
• Controversy information
• Corporate or SEC-style textual content
• NewsAPI-derived ESG signals
• Additional company-level information where available

• Missing-data handling
• Feature preparation
• Model benchmarking
• Class-imbalance handling
• Leakage checking
• Train/test evaluation
• Structured versus text-only versus fusion comparison

The multimodal classification experiment achieved approximately 0.71 weighted F1 in the project evaluation and performed better than the weaker text-only approach.

Structured models also provided strong regression baselines, while SHAP and transformer-attention analysis helped identify influential risk variables and relevant textual signals.

• SHAP feature attribution
• Feature-level risk drivers
• Attention heat-map analysis
• Translating model behaviour for non-technical stakeholders
• Human oversight

• During experimentation, leakage-prone configurations involving total ESG risk as an input were identified and corrected. This reinforced the importance of distinguishing genuinely predictive evidence from variables that reveal or closely reproduce the target.

• ESG labels are provider-dependent.
• Company disclosures may contain selective or promotional language.
• News coverage is uneven.
• A 542-company dataset limits generalisation.
• Attention is not a complete explanation of model reasoning.
• Provider disagreement cannot be solved by one model.
• Further temporal and external validation would be required before practical use.

• Data definitions can matter as much as model choice.
• A higher metric does not automatically produce a trustworthy system.
• Multimodal systems require careful validation.
• Leakage checks are essential.
• Explainability must be connected to user decisions.
• Human review remains necessary in risk-sensitive work.

• Expand the company and time coverage.
• Improve text quality and document segmentation.
• Add evidence retrieval.
• Compare additional domain-specific language models.
• Evaluate calibration and uncertainty.
• Conduct user testing with ESG or risk professionals.
• Build a transparent analyst-facing dashboard.