Retrieval Augmented Generation for Tables in Textual Documents based on Graph Representation Learning

In the field of water resources, documents contain numerous tables with important information such as water intake and precipitation. To improve the efficiency of water resources management, this paper proposes a table information Retrieval Augmented Generation method based on graph representation learning. The proposed method mainly includes the following two aspects: addressing the difficulty of retrieval caused by nested tables and asymmetric rows and columns in the field of water resources, a table information retrieval method based on graph representation learning is proposed; addressing the high professionalism and complex semantics of tables in the field of water resources, a table question-answering method based on large language model is proposed. Additionally, a water resources table information retrieval dataset is constructed, and the proposed method is comprehensively evaluated based on this dataset. Experimental results show that the method demonstrates high accuracy and adaptability in parsing and retrieving information from complex table structures.