add reference

README.md

@@ -13,12 +13,14 @@ size_categories:
 
 # Introduction
 
-ExioNAICS is the first enterprise-level ML-ready benchmark dataset tailored for GHG emission estimation, bridging sector classification with carbon intensity analysis. In contrast to broad sectoral databases like ExioML, which offer global coverage of 163 sectors across 49 regions, ExioNAICS focuses on enterprise granularity by providing 20, 850 textual descriptions mapped to validated NAICS codes and augmented with 166 sectoral carbon intensity factors. This design enables the automation of Scope 3 emission estimates (e. g., from purchased goods and services) at the firm level, a critical yet often overlooked component of supply chain emissions.
+ExioNAICS is the first enterprise-level ML-ready benchmark dataset tailored for GHG emission estimation, bridging sector classification with carbon intensity analysis. In contrast to broad sectoral databases like ExioML, which offer global coverage of 163 sectors across 49 regions, ExioNAICS focuses on enterprise granularity by providing 20,850 textual descriptions mapped to validated NAICS codes and augmented with 166 sectoral carbon intensity factors. This design enables the automation of Scope 3 emission estimates (e.g., from purchased goods and services) at the firm level, a critical yet often overlooked component of supply chain emissions.
 
 ExioNAICS is derived from the high-quality EE-MRIO dataset, ensuring robust economic and environmental data. By integrating firm-specific text descriptions, NAICS industry labels, and ExioML-based carbon intensity factors, ExioNAICS overcomes key data bottlenecks in enterprise-level GHG accounting. It significantly lowers the entry barrier for smaller firms and researchers by standardizing data formats and linking them to a recognized classification framework.
 
 In demonstrating its usability, we formulate a NAICS classification and subsequent emission estimation pipeline using contrastive learning (Sentence-BERT). Our results showcase near state-of-the-art retrieval accuracy, paving the way for more accessible, cost-effective, and scalable approaches to corporate carbon accounting. ExioNAICS thus facilitates synergy between machine learning and climate research, fostering the **integration** of advanced NLP techniques in eco-economic studies at the enterprise scale.
 
+For further details on the methodology and implementation, please refer to our paper: [Enterprises Level Emission Estimation Dataset with Large Language Models](https://arxiv.org/abs/2502.06874).
+
 # Dataset
 
 ExioNAICS serves as a hybrid textual and numeric dataset, capturing both enterprise descriptions (text modality) and sectoral carbon intensity factors (numeric modality). These data components are linked through NAICS codes, allowing end-to-end modeling of how enterprise descriptions map to sector emission intensities. Key dataset features include:
@@ -26,14 +28,13 @@ ExioNAICS serves as a hybrid textual and numeric dataset, capturing both enterpr
 - Enterprise Description
 - NAICS Description
 - Sectoral Emission Factor
-- Over 20, 000 textual entries
-- Hierarchical coverage: NAICS 2–6 digit codes (20 to 1, 114 categories)
+- Over 20,000 textual entries
+- Hierarchical coverage: NAICS 2–6 digit codes (20 to 1,114 categories)
 
 # NAICS Classification
 
-NAICS Classification is a fundamental component of enterprise-level GHG emission estimation. By assigning each firm to the appropriate sector category, practitioners can reference the corresponding carbon intensity factors, facilitating more accurate reporting. ExioNAICS adopts a natural language processing approach to NAICS classification, treating the task as an information retrieval problem.  
+NAICS Classification is a fundamental component of enterprise-level GHG emission estimation. By assigning each firm to the appropriate sector category, practitioners can reference the corresponding carbon intensity factors, facilitating more accurate reporting. ExioNAICS adopts a natural language processing approach to NAICS classification, treating the task as an information retrieval problem.
 
-Each enterprise description (query) is encoded separately, and matched against NAICS descriptions (corpus) based on the cosine similarity of their embeddings. This methodology leverages a dual-tower architecture, wherein the first tower processes the query (enterprise text) and the second tower processes NAICS descriptions.  
+Each enterprise description (query) is encoded separately, and matched against NAICS descriptions (corpus) based on the cosine similarity of their embeddings. This methodology leverages a dual-tower architecture, wherein the first tower processes the query (enterprise text) and the second tower processes NAICS descriptions.
 
 We apply machine learning to fine-tune a pre-trained Sentence-BERT model. Zero-shot SBERT models may achieve only around 20% Top-1 accuracy on the 1000 classes sector classification task, whereas contrastive fine-tuning raises this to over 75%. Further preprocessing exceeding 77% Top-1 accuracy, such as lowercasing and URL removal, can add incremental gains, leading to state-of-the-art results.
-