ReportParse: A Unified NLP Tool for Extracting Document Structure and Semantics of Corporate Sustainability Reporting

Gaku Morio, Soh Young In, Jungah Yoon, Harri Rowlands, Christopher D. Manning, Zhihong Zhang

Abstract

As corporate sustainability reporting becomes a standard business practice, the volume and complexity of these unstructured PDF documents have created a significant data bottleneck for researchers and analysts. This paper presents ReportParse, an open-source Python-based pipeline designed to automate the extraction of structured data and semantic meaning from massive PDF corpora. By combining advanced layout analysis with specialized NLP models, ReportParse facilitates the systematic tracking of climate commitments and the identification of potential "greenwashing." Our evaluation, involving the analysis of 2,480 reports from 2014–2020, demonstrates the tool's ability to replicate historical sentiment trends and highlight the "say-do" gap in corporate climate policy engagement.

Introduction

The proliferation of annual sustainability reports has transformed climate change from a peripheral concern to a central business metric. However, these documents—often spanning hundreds of pages with diverse layouts—lack standardization, making large-scale comparative analysis nearly impossible without significant manual labor. This "data bottleneck" hinders the robustness of sustainability research and financial analysis. ReportParse addresses this challenge by providing a standardized, reproducible framework for document parsing and semantic annotation.

Methodology

The ReportParse pipeline treats document parsing as a two-stage process, separating physical structure from semantic content.

The ReportParse workflow integrates layout analysis and specialized NLP models to process PDF reports.

1. The Reader (Structure Analysis)

The first module identifies the physical layout of the PDF. It utilizes deepdoctection to recognize structural elements such as titles, text blocks, and lists. These blocks are then segmented into individual sentences using spaCy. For scenarios requiring higher processing speeds, the tool also supports PyMuPDF for traditional text extraction.

2. The Annotator (Semantic Labeling)

Extracted text is passed to specialized NLP models integrated via the Hugging Face ecosystem. ReportParse is pre-configured to support domain-specific models like ClimateBERT and ESG-BERT. These models can identify environmental claims, climate sentiments, and specific commitments to renewable energy or carbon reduction.

The tool provides two primary interfaces: a Command Line Interface (CLI) for high-volume batch processing and a Gradio-based Web Interface for real-time visualization of bounding boxes and label distributions.

Results and Insights

The utility of ReportParse was demonstrated through a pilot study of 2,480 ESG and sustainability reports from the energy and basic materials sectors (2014–2020).

The "Say-Do" Gap

A key finding of the study was the discrepancy between corporate rhetoric and policy engagement. While there has been a significant rise in companies expressing support for IPCC and UN climate policies, there has not been a corresponding increase in engagement with specific energy transition policies, such as carbon capture.

Analysis of 2,480 reports shows that while supportive stances for UN/IPCC policies are rising (a), engagement with specific energy transition policies remains stagnant (b).

The "Parser Effect"

The authors observed that the choice of PDF "Reader" (e.g., PyMuPDF vs. deepdoctection) significantly impacts analysis outcomes. This highlights that document structure analysis is a critical variable in sustainability research, further justifying the need for a standardized tool like ReportParse to ensure reproducibility.

Human-AI Collaboration

Feedback from NPO analysts suggests that ReportParse serves as a powerful "force multiplier." While not replacing human judgment, it drastically reduces the time required to parse thousands of pages, allowing analysts to focus on high-level evidence tracking and qualitative assessment.

Discussion and Limitations

While ReportParse offers significant efficiency gains, it is subject to error propagation; failures in layout analysis can lead to "garbage in, garbage out" scenarios for the NLP models. Furthermore, the authors acknowledge ethical risks, such as companies "gaming" the system or investors over-relying on potentially misclassified model outputs. Future work will focus on improving model interpretability and developing metrics to quantify the overall quality of sustainability reporting.

Conclusion

ReportParse provides a much-needed open-source infrastructure for the systematic analysis of corporate climate disclosures. By bridging the gap between unstructured PDFs and actionable data, it empowers researchers to hold corporations accountable for their environmental commitments.

Summary updated

BibTeX
@inproceedings{morio2024reportparse,
  author = {Gaku Morio and Soh Young In and Jungah Yoon and Harri Rowlands and Christopher D. Manning and Zhihong Zhang},
  title = {{ReportParse: A Unified NLP Tool for Extracting Document Structure and Semantics of Corporate Sustainability Reporting}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
}
Figure