Go to ACL ARR 2025 May homepageAI for Climate Finance: Agentic Retrieval and Multi-Step Reasoning for Early Warning System Investments
Abstract: Tracking financial investments in climate adaptation is a complex and expertise-intensive task, particularly for Early Warning Systems (EWS), which lack standardized financial reporting across multilateral development banks (MDBs) and funds which are the main funders of these EWS projects. Analysts regularly encounter diverse PDF files containing tables and images with inconsistent formatting, rows, and columns, making it difficult and time-consuming to analyze reports and extract proper financial information.
To address this challenge, we introduce an agent-based Retrieval-Augmented Generation (RAG) system that orchestrates contextual retrieval with internal chain-of-thought (COT) reasoning to extract relevant financial data, classify investments, and ensure compliance with funding guidelines.
Our study focuses on a real-world application: tracking EWS investments funded by the Climate Risk and Early Warning Systems (CREWS) Fund.
We evaluate our agent-based RAG pipeline on 25 MDB project documents from the CREWS Fund,
comparing it against five model candidates—(1) a Zero-Shot Classifier (Baseline), (2) a Few-Shot “Zero Rule” Classifier, (3) a fine-tuned transformer-based classifier, and (4) a Few-Shot-V2 CoT+ICL classifier—across both multi-label classification and budget allocation tasks.
Our agent-based RAG achieves 87\% accuracy, 89\% precision, and 83\% recall, significantly outperforming these benchmarks.
We also benchmark it against the Gemini 2.0 Flash AI Assistant, setting the stage for a comparative study of Glass-Box Agents versus Black-Box Assistants to quantify the benefits of an agentic pipeline in transparency, explainability, and performance.
Finally, we release a benchmark dataset and expert-annotated corpus to catalyze further research in AI-driven climate finance tracking.
Paper Type: Long
Research Area: Information Extraction
Research Area Keywords: Information Extraction, Financial NLP, automated financial tracking
Contribution Types: NLP engineering experiment
Languages Studied: English
Submission Number: 2902
Loading