Lei Chen

Knowledge Base Group

Overview

Knowledge construction, reasoning and intelligent data systems

The Knowledge Base Group studies how to construct, maintain and use structured knowledge for intelligent data systems. A knowledge base organizes facts, entities and their relationships so they can be retrieved, updated and used in reasoning. The group connects knowledge graphs with data annotation, graph learning, retrieval-augmented generation and large language models.

The group develops methods for keeping knowledge current, grounding answers in external evidence, and improving the reliability and efficiency of knowledge-driven models.

Research areas

  • Knowledge base construction and maintenance. Building service knowledge bases and using temporal semantic rules to infer missing facts, detect outdated knowledge and keep evolving information current. [KartGPS, WeChat service KB]
  • Semantic data understanding and annotation. Identifying the meaning of table columns and structuring raw logs through knowledge graphs, language models and annotation methods. [RECA, KGLink, LLMLog, Crowdsourced annotation]
  • Knowledge graph representation and temporal learning. Learning representations of evolving graphs and facts with additional qualifiers, updating embeddings incrementally and predicting relationships over time. [AIR, SIT, Hyper-relational KGs, Temporal link prediction]
  • Knowledge-enhanced retrieval and generation. Retrieving structured evidence for multi-hop question answering, and integrating graph and multimodal knowledge into generation and dynamic text classification. [Triple-Fact Retriever, KERAG, AGRAG, MGRAG, GORAG]
  • Reliable knowledge and language model evaluation. Evaluating factual question answering, taxonomy knowledge and numerical reasoning, and examining security risks in retrieval-augmented generation. [CRAG, TaxoGlimpse, NumericBench, Secure RAG survey]
  • Knowledge-guided model selection and efficient learning. Using graph and model knowledge to select suitable GNNs, and studying training-data selection and KV-cache management to reduce learning and inference costs. [GraphAgent, GRACE, KV-cache survey]