NatureKG: an ontology and knowledge graph for nature finance with a Text2Cypher application

In one sentence

Summary

This paper addresses the lack of structured tools for mapping dependencies between natural capital and financial assets, a critical gap in nature finance where institutions struggle to systematically assess environmental risks and impacts. The authors propose NatureKG, the first ontology and instantiated knowledge graph tailored to this domain, grounded in frameworks like ENCORE and SBTN. The research positions this work as a foundational step toward integrating domain-specific ontologies with AI systems to enhance transparency and scalability in sustainable finance decision support, particularly for the built environment sector. The methodology involves designing an ontology defining entities such as Actions, Drivers of Nature Loss, and Value Chains, which was instantiated into a Neo4j database containing 320 nodes and 540 relationships curated by experts. To enable natural language interaction, the authors constructed a Text2Cypher dataset and fine-tuned three open-source LLMs (Phi-3, LLaMA-3.1-8B, and Mistral-7B) using LoRA. The experimental design included three dataset split strategies (paraphrase, cypher-level, and generalization) to evaluate linguistic and structural generalization, using metrics like execution accuracy, BLEU, and clause-level F1 scores. Results indicate that Phi-3 achieved the highest execution accuracy (0.21) and Macro F1 (0.56), demonstrating that smaller, fine-tuned models can generalize effectively in low-resource, domain-specific settings, outperforming larger models like Mistral-7B. However, the study acknowledges limitations including modest initial accuracy, potential memorization from paraphrasing, and a narrow focus on the built environment. The findings validate the feasibility of LLM-assisted querying for nature finance but highlight the need for larger, domain-aligned data catalogues and refined reasoning capabilities for practical deployment in financial risk assessment.

Research problem

Financial institutions lack structured tools to systematically map dependencies between natural capital and financial assets, making it difficult to assess environmental risks, impacts, and dependencies. Existing systems do not adequately support the integration of nature-related considerations into financial decision-making, leading to gaps in risk management and sustainable investment strategies.

Core contributions

• Introduction of NatureKG, the first ontology and knowledge graph specifically tailored to nature finance, defining key entities like Actions, Drivers of Nature Loss, and Value Chains.
• Instantiation of the ontology into a Neo4j database with 320 nodes and 540 relationships, curated by domain experts and grounded in ENCORE and SBTN frameworks.
• Construction of a domain-specific Text2Cypher dataset to enable natural language querying of the knowledge graph.
• Demonstration of fine-tuning open-source LLMs (Phi-3, LLaMA-3.1-8B, Mistral-7B) for Text2Cypher tasks, establishing a baseline for LLM-assisted querying in nature finance.
• Evaluation of model performance across different dataset split strategies to assess linguistic and structural generalization in a low-resource setting.

Method

• Designed a domain ontology grounded in ENCORE and SBTN frameworks.
• Instantiated the ontology into a Neo4j graph database with expert curation.
• Constructed a Text2Cypher dataset by pairing natural language questions with Cypher queries.
• Fine-tuned LLMs using LoRA with 4-bit quantization for efficient training.
• Evaluated models using BLEU, exact match, execution accuracy, and clause-level F1 scores.
• Implemented three dataset split strategies: paraphrase, cypher-level, and generalization.

Results & findings

• Phi-3 achieved the highest execution accuracy (0.21) and Macro F1 score (0.56) on the paraphrase split.
• Phi-3 also showed strong performance on generalization splits, with execution accuracy of 0.14 and partial match scores up to 0.60.
• LLaMA-3.1-8B exhibited balanced performance with high execution rates (0.981) but lower execution accuracy compared to Phi-3.
• Mistral-7B lagged significantly, achieving near-zero execution accuracy and partial scores across all splits.
• Clause-level analysis revealed that models perform well in identifying graph structures (MATCH clause) but struggle with logical filtering (WHERE clause).
• Common error patterns include naming mismatches, introduction of non-existent fields, and confusion between node labels and properties.

Data & experiments

• NatureKG knowledge graph containing 320 nodes and 540 relationships focused on the built environment sector.
• Text2Cypher dataset with 545 samples, augmented via paraphrasing to expand training data.
• Fine-tuning of three open-source LLMs: Phi-3-medium-4k-instruct, LLaMA-3-8b-instruct, and Mistral-7b-instruct.
• Evaluation using BLEU, exact match, execution accuracy, and clause-level F1 scores across paraphrase, cypher-level, and generalization splits.
• Experiments conducted on two A100 GPUs using LoRA with 4-bit quantization.

Limitations & caveats

• Focus is limited to the built environment sector, restricting generalizability.
• Modest execution accuracy limits immediate practical deployment in high-stakes financial decisions.
• Paraphrasing augmentation may lead to memorization of cypher outputs.
• Models struggle with structural generalization and unseen schema combinations.
• Potential for hallucination and naming mismatches remains a concern.

Curator notes

Discovery status: specific_domain_round2_promoted. Source links: Exploring the Potential of Large Language Models in Supply Chain Management: A Study Using Big Data