💬 NLP

Scope3Trace: Evidence-Based Identification and Extraction of Scope 3 GHG Emissions from Sustainability Reports

The paper introduces Scope3Trace, an evidence-grounded framework that combines document processing, LLM-assisted localization, and hybrid rule-LLM extraction to reliably and transparently identify and extract Scope 3 GHG emissions from heterogeneous sustainability reports, accompanied by a new dual-level multimodal dataset.

Original authors: Siyuan Zheng, Yifan Duan, Chao Xue, Flora D. Salim

Published 2026-07-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Siyuan Zheng, Yifan Duan, Chao Xue, Flora D. Salim

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to solve a giant, global puzzle: figuring out exactly how much carbon pollution every single company on Earth is responsible for. This isn't just about the smoke coming out of their own chimneys; it's about the invisible cloud of pollution created by everything they buy, sell, and ship. In the world of climate science, this massive cloud is called "Scope 3" emissions. Think of Scope 1 and 2 as the smoke you can see right in front of you—like a factory's own smokestack or the electricity bill for its lights. But Scope 3 is the "ghost in the machine." It's the pollution from the steel used to build the factory, the trucks delivering the products, and even the waste customers throw away.

The problem is that while companies are getting better at reporting their own smoke, they are terrible at reporting this ghostly Scope 3 cloud. Their reports are messy, scattered across thousands of different PDF documents, written in confusing language, and often hidden inside complex tables. It's like trying to find a specific needle in a haystack made of other haystacks, where the needles are made of invisible ink. Scientists and regulators care deeply about this because you can't fix a problem if you can't measure it accurately. If we can't track these hidden emissions, we can't know if companies are actually helping the planet or just pretending to.

Enter Scope3Trace, a new digital detective team designed to hunt down these hidden numbers. The researchers built a smart system that acts like a super-powered librarian combined with a forensic accountant. Instead of hiring a team of humans to read millions of pages (which would take forever and cost a fortune), they created an automated pipeline that uses Artificial Intelligence (AI) to do the heavy lifting. But here is the clever part: they didn't just let the AI guess. They built a "rule-based" safety net. The AI reads the document, finds the numbers, and then a set of strict rules checks the math and the context to make sure the AI didn't get confused.

The system works in three main steps, like a well-oiled machine. First, it grabs the PDF reports and turns them into readable text, even if they are scanned images. Second, it uses a smart AI to find the exact pages and tables where the emissions are mentioned, ignoring the boring fluff. Third, it reconstructs the messy tables and pulls out the specific numbers for Scope 1, 2, and 3. Crucially, for every number it extracts, the system grabs a "receipt"—a tiny snippet of the original text proving where that number came from. This means you can always trace a number back to its source, just like checking a receipt at a store.

The results are impressive. When the researchers tested their system against a "gold standard" set of data that humans had carefully checked, Scope3Trace got it right almost every time, scoring a near-perfect 0.99 out of 1.0 on accuracy. This is significantly better than using just a raw AI model or simple rule-based tools alone. The team also used this system to build a massive new dataset called Scope3Trace, which includes over 54,000 records of building-level data and organization-level reports from countries like Australia and across Europe. They found that while companies report their total emissions, they rarely break it down by specific building. In fact, only about 3% of the building-level data in their new dataset was directly reported by the companies; the rest had to be carefully estimated using public energy data and smart math, but the system kept track of which numbers were "reported" and which were "estimated" so no one gets confused.

The paper also tested if this new data could be used to predict future emissions. They fed the data into a model that looked at building details and even satellite photos of the buildings. The model learned that combining the text data with satellite images gave the best predictions, suggesting that what a building looks like from space holds clues about how much pollution it creates. However, the authors are careful to note that this isn't a magic wand. The system works best where public data exists (like in Australia and parts of Europe) and struggles where data is missing. They explicitly state that their estimates are not a replacement for official, audited reports but are a transparent way to fill in the gaps so we can see the bigger picture.

In short, Scope3Trace doesn't solve the mystery of corporate pollution overnight, but it provides the best map we've ever had to navigate the fog. It proves that by combining smart AI with strict rules and clear evidence, we can turn a chaotic mountain of PDFs into a clear, trustworthy dataset. This allows researchers, investors, and policymakers to finally see the "ghost" of Scope 3 emissions, understand where it comes from, and start making decisions based on real evidence rather than guesswork.

Technical Summary: Scope3Trace

Problem Statement

Scope 3 greenhouse gas (GHG) emissions constitute the majority of corporate carbon footprints but remain difficult to analyze at scale. Current challenges include sparse disclosures, heterogeneous report formats (scanned PDFs, mixed layouts), and a lack of unified structured datasets. Existing approaches typically rely on Large Language Models (LLMs) for extraction but often suffer from two critical limitations:

Lack of Evidence Grounding: Extracted values frequently lack explicit links to the source text, making verification difficult and increasing the risk of hallucination.
Reliance on Manual Verification: Many pipelines depend on costly, time-consuming manual expert annotation to ensure reliability, hindering scalability.
Abstracted Estimation: Alternative estimation-based methods often infer emissions from industry factors and spending data, abstracting away from organization- or asset-level characteristics and offering limited transparency regarding reporting boundaries.

Consequently, Scope 3 data remains difficult to interpret, validate, and operationalize for decision-making.

Methodology: The Scope3Trace Framework

The authors propose Scope3Trace, an evidence-grounded information extraction framework designed to process heterogeneous ESG and sustainability reports. The system employs a modular, multi-agent pipeline orchestrated by three specialized agents: Search, Scope, and Verify.

1. Pipeline Architecture

The workflow proceeds through four distinct phases, combining rule-based determinism with bounded LLM semantic understanding:

Data Acquisition & Parsing: The Search Agent collects public PDF reports and converts them into structured text using OCR. It performs page-level localization to identify sections containing Scope 1–3 disclosures, replacing generic page localization with semantic understanding.
Table Reconstruction & Extraction: The system reconstructs tabular disclosures using LLM-assisted table interpretation. A hybrid rule–LLM design is employed: LLMs handle semantic tasks (localization, table structure recognition), while rule-based components handle deterministic tasks such as numeric parsing, unit normalization, category mapping, and evidence verification against the OCR text.
Asset Inventory & Alignment: The Scope Agent compiles a building asset inventory by linking public registries (e.g., NABERS, NGER) with entities mentioned in reports. It uses semantic models to resolve ambiguities (e.g., synonymous building names) and aligns asset identifiers with organizational disclosures.
Verification & Diagnostics: The Verify Agent performs cross-level alignment. It validates the relevance of asset data to organizational disclosures using semantic similarity. Crucially, it distinguishes between directly reported values (extracted verbatim) and estimated values (derived via deterministic formulas using activity data and emission factors).

2. Dual-Level Data Representation

The framework constructs a dual-level dataset:

Organization-Level: Records correspond to an organization and reporting year, containing total Scope 1–3 emissions, category-level breakdowns, reporting boundaries, and audit status. Every value is paired with a text snippet from the original report (evidence grounding).
Building-Level: Records correspond to individual buildings, documenting physical attributes, geospatial coordinates, and operational indicators. This layer integrates multimodal data (satellite imagery) and semantic metadata.

3. Evidence Grounding Strategy

A core design principle is that every extracted value is linked to its source text span. The system does not rely on speculative inference; instead, it uses LLMs for bounded semantic tasks while delegating logic (parsing, normalization, evidence checking) to rule-based components. This ensures that derived values are transparently labeled as "modeled" or "estimated" rather than "reported."

Key Contributions

Evidence-Grounded Extraction Framework: A novel pipeline integrating PDF collection, OCR parsing, LLM-assisted localization/table reconstruction, and hybrid rule–LLM extraction. It explicitly links extracted emissions values to supporting textual evidence.
Dual-Level, Multimodal Dataset: The construction of Scope3Trace, a dataset comprising organization-level disclosures and building-level asset records. It is among the first efforts to create a multimodal, evidence-grounded dataset specifically targeting Scope 3 emissions, covering heterogeneous reporting formats.
Traceable and Verifiable Extraction: The system enables the extraction of organization- and building-level Scope emissions with high accuracy, providing transparency regarding data provenance (reported vs. estimated) and reporting boundaries.

Results

The system was evaluated on a manually annotated gold set covering Scope 1/2/3 totals, categories, and page-level evidence.

Performance: Scope3Trace achieved near-perfect performance, significantly outperforming rule-based baselines, standalone LLMs (both chunked and single-call), retrieval-augmented (RAG) pipelines, and other multi-agent frameworks (e.g., AutoGen, CrewAI).
- Scope3 F1: 0.99 (vs. 0.88 for the best baseline).
- Category F1: 0.97 (vs. 0.78 for the best baseline).
- Exact Match (EM): 0.96.
Ablation Studies: Removing key components, particularly table reconstruction and evidence grounding, led to substantial performance drops. Page localization primarily affected efficiency rather than accuracy.
Model Robustness: The system's performance remained stable across different LLM backbones (GPT-4o, Gemini-2.0, DeepSeek-V3), suggesting the task's semantic portion is robust to model choice when deterministic logic is offloaded to rules.
Downstream Utility: In a building-level Scope 3 prediction task using multimodal inputs (metadata + satellite imagery), a multimodal MLP achieved an $R^2$ of 0.46 for total Scope 3 and 0.54 for category-level prediction, outperforming single-modality baselines.

Significance and Claims

The paper claims that Scope3Trace addresses the critical gap in analyzing Scope 3 emissions by providing a transparent, interpretable, and coverage-aware approach.

Transparency: By distinguishing between reported and estimated values and providing explicit evidence links, the framework allows users to understand the assumptions and limitations of the data.
Scalability: The hybrid approach reduces reliance on manual annotation, enabling the processing of heterogeneous reports at scale.
Operationalization: The dataset facilitates the examination of how reported emissions relate to underlying assets and activity proxies, highlighting structural limitations in current reporting practices (e.g., the scarcity of building-level Scope 3 disclosures).

The authors emphasize that the dataset is intended for research and analysis to support interpretable emissions modeling, not to replace audited disclosures or support compliance reporting. The work highlights both the opportunities for AI-driven climate analysis and the structural data gaps that persist in real-world sustainability reporting.