Row-based hierarchical graph network for multi-hop question answering over textual and tabular data

Multi-hop Question Answering over heterogeneous data is a challenging task in Natural Language Processing(NLP), which aims to find the answer among heterogeneous data sources and reasoning chains. When facing complex reasoning scenarios, most existing QA systems can only focus on some specific types of data. To solve this issue, we propose a new approach based on Row Hierarchical Graph Network(RHGN), which can accomplish multi-hop QA over both textual and tabular data. Specifically, RHGN consists of two phases: the row selection phase is designed to find the table row that most likely contains the answer, and the row reading comprehension phase that aims to locate the final answer in the answer row. In the row selection phase, we utilize a retriever to search all the supporting evidence related to the question, and a pre-training language model is employed to select the appropriate answer row. In the succeeding stage of row reading comprehension, we propose a row-based hierarchical graph network to capture the structural information, and a gated mechanism is used to perform graph reasoning. Eventually, the optimum final answer can be obtained by three interrelated sub-tasks. The experimental results demonstrate the effectiveness of RHGN and it achieves superior performance on the HybridQA dataset.