MRKG-DSP: Enhanced Multimodal Reasoning with Knowledge Graphs, Dynamic Semantic Pruning and LLM Integration

This paper presents an innovative approach to Knowledge Graph (KG) reasoning that integrates and Large Language Models (LLMs) into dynamically pruned message-passing networks. Conventional knowledge graph reasoning methods frequently encounter challenges related to scalability and noise, especially in multi-hop reasoning tasks that engage extensive and intricate graphs. To solve these issues, our approach introduces a novel pruning mechanism that selectively removes irrelevant nodes based on both structural and semantic information. By incorporating Semantic Random Walks with Node2Vec, we enhance node em beddings with contextual semantics derived from pre-trained word embeddings, leading to more informed pruning decisions. Additionally, we leverage the capabilities of LLMs to integrate unstructured textual data into the reasoning process, enabling a more comprehensive understanding of the query context. The proposed methodology not only improves the efficiency of KG reasoning by reducing computational complexity but also enhances accuracy in answering complex queries. Experimental results on benchmark QA datasets, such as CSQA, demonstrate that our approach significantly improves accuracy, achieving 75.10% on test data and 79.69% on development data, while exceeding existing techniques in reasoning effectiveness.