Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Group Reasoning Emission Estimation Networks

Yanming Guo, Xiao Qian, Kevin Credit, Jin Ma

TL;DR

GREEN addresses the challenge of SME GHG reporting by standardizing enterprise emission estimation through NAICS-aligned carbon intensities. It combines ExioNAICS, SBERT with contrastive fine-tuning, and a Hierarchical Group Reasoning approach to map descriptions to 1,114 NAICS categories via an information retrieval formulation, ultimately computing emissions as $E = R \cdot I$. The framework achieves state-of-the-art NAICS-6 accuracy (83.68% top-1, 91.47% top-10) and a real-world MAPE of $45.88\%$ on a 20-company sample, while providing an open benchmark linking thousands of enterprises to emission factors from ExioML. By delivering an end-to-end, scalable pipeline for enterprise-level carbon accounting, GREEN has the potential to broaden adoption among SMEs and improve transparency in ESG reporting.

Abstract

Accurate greenhouse gas (GHG) emission reporting is critical for governments, businesses, and investors. However, adoption remains limited particularly among small and medium enterprises due to high implementation costs, fragmented emission factor databases, and a lack of robust sector classification methods. To address these challenges, we introduce Group Reasoning Emission Estimation Networks (GREEN), an AI-driven carbon accounting framework that standardizes enterprise-level emission estimation, constructs a large-scale benchmark dataset, and leverages a novel reasoning approach with large language models (LLMs). Specifically, we compile textual descriptions for 20,850 companies with validated North American Industry Classification System (NAICS) labels and align these with an economic model of carbon intensity factors. By reframing sector classification as an information retrieval task, we fine-tune Sentence-BERT models using a contrastive learning loss. To overcome the limitations of single-stage models in handling thousands of hierarchical categories, we propose a Group Reasoning method that ensembles LLM classifiers based on the natural NAICS ontology, decomposing the task into multiple sub-classification steps. We theoretically prove that this approach reduces classification uncertainty and computational complexity. Experiments on 1,114 NAICS categories yield state-of-the-art performance (83.68% Top-1, 91.47% Top-10 accuracy), and case studies on 20 companies report a mean absolute percentage error (MAPE) of 45.88%. The project is available at: https://huggingface.co/datasets/Yvnminc/ExioNAICS.

Group Reasoning Emission Estimation Networks

TL;DR

GREEN addresses the challenge of SME GHG reporting by standardizing enterprise emission estimation through NAICS-aligned carbon intensities. It combines ExioNAICS, SBERT with contrastive fine-tuning, and a Hierarchical Group Reasoning approach to map descriptions to 1,114 NAICS categories via an information retrieval formulation, ultimately computing emissions as . The framework achieves state-of-the-art NAICS-6 accuracy (83.68% top-1, 91.47% top-10) and a real-world MAPE of on a 20-company sample, while providing an open benchmark linking thousands of enterprises to emission factors from ExioML. By delivering an end-to-end, scalable pipeline for enterprise-level carbon accounting, GREEN has the potential to broaden adoption among SMEs and improve transparency in ESG reporting.

Abstract

Accurate greenhouse gas (GHG) emission reporting is critical for governments, businesses, and investors. However, adoption remains limited particularly among small and medium enterprises due to high implementation costs, fragmented emission factor databases, and a lack of robust sector classification methods. To address these challenges, we introduce Group Reasoning Emission Estimation Networks (GREEN), an AI-driven carbon accounting framework that standardizes enterprise-level emission estimation, constructs a large-scale benchmark dataset, and leverages a novel reasoning approach with large language models (LLMs). Specifically, we compile textual descriptions for 20,850 companies with validated North American Industry Classification System (NAICS) labels and align these with an economic model of carbon intensity factors. By reframing sector classification as an information retrieval task, we fine-tune Sentence-BERT models using a contrastive learning loss. To overcome the limitations of single-stage models in handling thousands of hierarchical categories, we propose a Group Reasoning method that ensembles LLM classifiers based on the natural NAICS ontology, decomposing the task into multiple sub-classification steps. We theoretically prove that this approach reduces classification uncertainty and computational complexity. Experiments on 1,114 NAICS categories yield state-of-the-art performance (83.68% Top-1, 91.47% Top-10 accuracy), and case studies on 20 companies report a mean absolute percentage error (MAPE) of 45.88%. The project is available at: https://huggingface.co/datasets/Yvnminc/ExioNAICS.

Paper Structure

This paper contains 19 sections, 2 theorems, 8 equations, 2 figures, 7 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Machine Learning in Sector Classification
  4. Self-Supervised Contrastive Learning Framework
  5. GHG Emission Estimation by Ecological Economic Framework
  6. Method
  7. Open-Source Large-Scale NLP Benchmark Dataset: ExioNAICS
  8. Sentence-BERT with Contrastive Fine-Tuning
  9. Large-Scale Sector Classification via Hierarchical Group Reasoning
  10. Theoretical Performance Analysis
  11. Experiments & Results
  12. Experimental Setup
  13. Comparison of Pre-trained SBERT Backbones
  14. Data Augmentation Analysis
  15. Group Reasoning Hyperparameter $k$
  16. ...and 4 more sections

Key Result

Theorem 1

Let there be a hierarchical classification tree of depth $d$ with uniform branching $b$. Let each level $i$ have accuracy $p_i$. The hierarchical approach has strictly lower Shannon entropy than a single-stage classifier with $b^d$ classes and accuracy $\prod_{i=1}^d p_i$. Formally, $H_{D}(Y) \ge H_

Figures (2)

  • Figure 1: High-level overview of the GREEN framework for enterprise-level emission estimation. (1) We construct ExioNAICS, a large-scale dataset pairing enterprise descriptions with validated NAICS codes and emission-factor data. (2) A contrastive fine-tuning step aligns embeddings of enterprise and NAICS descriptions. (3) A Group Reasoning approach hierarchically classifies each enterprise into a fine-grained NAICS sector. (4) Emission is inferred by multiplying the annual revenue by the sector’s carbon intensity. (5) Validation against official sustainability reports reveals small errors for single-sector firms and larger errors for diversified companies.
  • Figure 2: Architecture of Sentence-BERT (SBERT) for enterprise classification as an IR task. The model encodes both queries (enterprise descriptions) and corpus documents (NAICS definitions) into a shared embedding space, where cosine similarity measures relevance. Fine-tuning uses a contrastive loss separating correct from incorrect matches.

Theorems & Definitions (2)

  • Theorem 1: Hierarchical Classification Entropy
  • Theorem 2: Complexity