From Literature to Knowledge Graphs: Automated Extraction and Retrieval of Life Cycle Assessment Data with Large Language Models

The large and steadily growing volume of scientific literature poses challenges for accessing and utilizing data due to its unstructured nature. Life cycle assessment (LCA), in particular, relies on high-quality life cycle inventory (LCI) data to quantify the environmental impacts of products or services across their life cycle, yet manual extraction of such data is time-consuming and labor-intensive. This study presents a framework that integrates large language models (LLMs) with knowledge graph (KG) to extract and manage LCA data from published studies. A retrieval-augmented generation (RAG) pipeline is developed to mine three core data types from full-text articles: LCI data, life cycle impact assessment (LCIA) results, and LCA modeling assumptions. The extracted information is normalized and mapped to an ontology-driven, LCA-oriented knowledge graph (LCAKG) implemented in Neo4j. To support user interaction, an LLM-based question-answering system translates natural language queries into executable graph queries, allowing users to retrieve rich information without prior knowledge of KG schemas. The framework is evaluated using a case study of LCA studies on chemical production. The results demonstrate high semantic accuracy in data extraction, with F1-scores ranging from 73.54% to 93.34%. Query performance is significantly improved by combining similarity search with text-to-cypher reasoning, increasing the F1-score from 56.98% (baseline) to 75.18%. The proposed framework enhances the accessibility and interoperability of LCA domain data and provides a scalable foundation for large-scale knowledge synthesis to support sustainability research.