The Scientific Vision
Current Large Language Models (LLMs) excel in natural language but struggle with hallucinations, bias, and lower accuracy in specialized tasks. To solve these scientific challenges, our research focuses on the mutual integration of LLMs and Knowledge Graphs (KGs). KGs provide verifiable facts to ground LLMs in reality, while LLMs leverage their reading capabilities to automatically extract and update KG structures from unstructured text. The Action tackles six core research questions to build AI systems that are reliable, explainable, and linguistically diverse.
LLM Augmentation Strategies
We explore general solutions to inject structured KG data into LLMs, utilize KGs as external plugins during inference, and formally verify outputs to reduce confabulations.
Domain-Specific Adaptation
The research produces targeted data and algorithmic resources to measure and improve LLM performance in specialized, high-stakes fields like biomedicine, finance, and science.
Automated KG Construction
We develop new methodologies that utilize LLM-based information extraction to efficiently generate, interlink, and refine massive, highly accurate semantic networks from textual datasets.
Cross-Lingual Alignment
To overcome language barriers and support low-resource languages, we investigate multilingual entity linking and advanced embeddings to align structured knowledge across different cultures.
Ethical AI and Bias Mitigation
To overcome language barriers and support low-resource languages, we investigate multilingual entity linking and advanced embeddings to align structured knowledge across different cultures.
Benchmarking and Validation
We establish rigorous standardized metrics, new benchmarks, and real-world validation protocols to quantitatively evaluate the factual accuracy, efficiency, and ethical integrity of LLM-KG integrations
