Enhancing Large Model Document Question Answering Through Retrieval Augmentation
Jingwen Zeng, Rongrong Zheng, C. Wang, Wenting Xue, Xiaoyang Yu, Tao Zhang · 2024
Document question answering(DocQA) requires models to provide comprehensive answers, given a series of document content and questions. In recent years, large neural models have been widely applied in fields such as natural language processing and computer vision, yielding significant results. However, these applications usually face a challenge that the large language model often produce fake information, owing to model illusions. To tackle this problem, the recent proposed document question answering has proven to be effective. Current methods primarily rely on vector for passage retrieval, but the effectiveness is often limited, which restrict the model's performance. To address this issue effectively, we propose a dual-path retrieval along with precise ranking framework to enhance existing knowledge retrieval. To better evaluate this task, we also constructed a manually annotated test set for validation. Experimental results demonstrate the significant advantages of our model.