Multi-modal Knowledge Graph Representation & Reasoning & Completion: Represent the entities and relations in the vector spaces to conduct structured knowledge reasoning and new knowledge discovery. Our research mainly focuses on incorporate diverse modalities (image, text, numeric, audio, video) of the entities to enhance the representation learning and further benifit the MMKG reasoning and completion. Besides, we introduce more popular and effective representation learning techniques to solve the problems in this field and better achieve this goal:

• Imbalanced Modality Distribution: Adversarial Training (AdaMF-MAT, NATIVE)
• Fine-grained Modality Fusion: Modality Tokenization and Hierarchical Modeling (MyGO)
• Modality Entanglement and Disentanglement: Mixture-of-Experts Networks (MoMoK)

Knowledge-enhanced Large Language Models: : Incorporating structural knowledge into large language models to achieve the harmonious combination of KGs and LLMs. Our research consists of two points: employing LLMs to enhance structured reasoning, and utilizing structured knowledge (both parameterized and non-parameterized) to improve LLMs' performance in certain tasks:

• Incorporate Parameterized Structural Information into LLMs: Knowledge Prefix Adapter (KoPA)
• Systematic Evaluation of LLM's Structural Understanding Ability: Structural Knowledge Prompting Benchmark with Reasoning and Understanding Tasks (SUBARU)
• Select Better Context Knowledge: Knowledgeable Preference Alignment (KnowPAT)

Application of KGs and LLMs in Specific Domains: : Utilizing KGs and LLMs to solve real-world problems in specific domains. Our research mainly focuces on the telecom network service and E-commerce recommendation system. We have developed the following algorithms and frameworks in our collaborative PROJECTS with companies:

• Telecom Network Analysis and Fault Diagnosis: KG-enhanced Domain BERT Pre-training (TeleBERT)
• Intelligent Telecom Customer Service: Knowledgeable Preference Alignment (KnowPAT)
• E-commerce Recommendation/QA/Review Prediction: Transferable Text-informed KG Pre-training (MuDoK)