---
title: ESG Ontology-Based Sentiment Mapping
emoji: 🌱
colorFrom: green
colorTo: blue
sdk: gradio
app_file: app.py
pinned: false
---

# Advanced ESG Analysis & Interpretability Dashboard

This project provides a powerful, interactive dashboard for in-depth analysis of Environmental, Social, and Governance (ESG) reports. It goes beyond simple keyword matching by leveraging a custom ontology, advanced NLP models, and a suite of analytical tools to uncover nuanced insights, detect sentiment drift, and identify potential greenwashing.

The application is built with Gradio and integrates several advanced features, from section-aware context modeling to knowledge graph generation and bias analysis.

## Key Features

The dashboard integrates a multi-stage pipeline to deliver deep, explainable insights:

*   **Ontology-Powered Aspect Mapping**: Maps text from ESG reports to a custom, hierarchical ESG ontology (`esg_ontology.owl`), ensuring that analysis is grounded in a structured, domain-specific framework.
*   **Section-Aware Context Modeling**: Intelligently distinguishes between different sections of a report (e.g., "Commitments" vs. "Results") and applies contextual weights to mitigate optimism bias in forward-looking statements.
*   **Cross-Document Consistency Analysis**: Compares two documents (e.g., reports from consecutive years) to automatically detect *sentiment drift* for specific ESG aspects, flagging potential inconsistencies or greenwashing.
*   **Weakly Supervised Aspect Expansion**: Uses `KeyBERT` to discover emerging ESG topics and keywords from the input text that are not yet in the ontology, suggesting them for future inclusion.
*   **Knowledge Graph Generation**: Exports the analysis results into a machine-readable RDF Turtle file (`esg_knowledge_graph.ttl`), representing documents, aspects, and sentiments as a semantic graph.
*   **Fine-Tuning Ready**: The architecture is designed to seamlessly integrate a fine-tuned sentence transformer model for improved domain adaptation and mapping accuracy. A training script template (`train_finetune.py`) is provided.
*   **Interpretability Dashboard**:
    *   **Sentiment Distribution**: Visualizes the breakdown of positive, negative, and neutral sentiments across documents.
    *   **Bias & Confidence Analysis**: A scatter plot that correlates "Optimism Bias" with "Mapping Confidence" to help users critically evaluate the sentiment associated with each aspect.
    *   **Ontology Viewer**: Displays the structure of the ESG ontology directly in the dashboard.

Hybrid stage (7 + 8 + 10)
*   The ontology backbone (Stage 7) structures the ESG knowledge.
*   The weakly supervised extraction (Stage 8) expands it dynamically.
*   The dashboard layer (Stage 10) communicates interpretability metrics.
*   Once you fine-tune embeddings on ESG corpora (Stage 9), your system will reach full maturity — improving confidence and semantic clustering precision.