CRII: RI: Synergistic Learning of Language Models and Knowledge Graphs for Trustworthy Question Answering

This project advances artificial intelligence (AI) by developing more reliable and trustworthy question-answering systems that combine large language models with structured knowledge graphs. Current AI systems, while powerful in generating human-like responses, often produce answers that are plausible but factually incorrect, which can have severe consequences in critical domains like healthcare and legal services. By integrating the strengths of large language models with the structured, verifiable information from knowledge graphs, this research creates AI systems that people can trust for high-stakes decisions. The project develops novel computational methods that ensure AI outputs are grounded in verified knowledge, reducing errors and improving reliability. The research outcomes will benefit society by enabling more dependable AI assistants for healthcare diagnosis support, legal consultation, and other domains where accuracy is paramount. The project also contributes to computer science education through new courses on knowledge graphs and advanced language models at the University of California, Merced, helping prepare the next generation of AI researchers and practitioners. Open-source software tools and datasets developed through this research will be made freely available to the broader research community, accelerating progress in trustworthy AI development.

The project develops innovative approaches for combining large language models with knowledge graphs through four key technical advances. First, it creates new methods for synergistic knowledge representation that effectively capture information from both unstructured text and structured knowledge sources. Second, it designs AI agents that automatically explore knowledge paths to augment language model inputs with relevant verified information. Third, it develops constrained decoding techniques that ensure language model outputs remain consistent with knowledge graph facts. Fourth, it enables bi-directional reasoning between language models and knowledge graphs through carefully designed learning algorithms. The research methodology includes rigorous evaluation on diverse benchmarks across general domain question answering, healthcare services, and legal consultation. The project employs an efficient knowledge path exploration algorithm using intelligent pruning, which significantly reduces computational overhead while preserving model performance. The developed methods will be made available through comprehensive open-source implementations, accompanied by detailed documentation of computational requirements and efficiency metrics to facilitate adoption by other researchers and practitioners.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

Award Number: 2451683
Principal Investigator: Yiwei Wang
Funds Obligated: $175,000
State: CA