Towards Faithful Response Generation for Chinese Table Question Answering

Anonymous

Abstract: The response generation for TableQA aims to automatically generate a response to end-users from a SQL query and its corresponding execution result (in the form of table). It is an essential and practical task. However, there has been little work on it in recent years. We consider this may be blamed on the lack of large-scale and high-quality datasets in this area. In this paper, we present ResponseNLG, a large-scale and high-quality Chinese dataset for TableQA response generation, to advance the field in both academic and industrial communities. Further, to bridge the structural gap between the input SQL and table and establish better semantic alignments, we propose a Heterogeneous Graph Transformation approach. In this way, we establish a joint encoding space for the two heterogeneous input data and convert this task to a Graph-to-Text problem. We further introduce the Node Segment Embedding to better preserve the original graph structure upon PLMs based models.