Leveraging Dual Encoder Models for Complex Question Answering over Knowledge Bases

Knowledge-based question answering is a hot topic in Natural Language Processing (NLP), especially in addressing complex questions. Existing methods, which transform complex questions into query graphs, often struggle with low-quality graphs. To improve this, we propose a dual-encoder model for generating and ranking query graphs. We incorporate beam search and a scoring function for high-quality graph generation, and use a dual-encoder model with attention mechanism for graph ranking. By extracting semantic structures from complex questions, we further refine the ranking process. Our experiments on benchmark datasets show competitive results, suggesting practical applications in complex question answering.