Abstract: To meet users'interactive exploration needs for correlation issues among multiple charts when direct access to the underlying database is not possible,a multi-chart joint question answering method was proposed.In this method,the joint interpretation of data was achieved from multiple charts through two core stages:data preparation and answer generation.In the data preparation stage,chart data was extracted and reconstructed into tabular data,and each cell was provided with a text description,offering a unified data format for subsequent models.Furthermore,to enhance model accuracy and response efficiency,a text fact filtering method was introduced,which could screen out texts relevant to user questions from a large amount of tabular text descriptions,providing precise data support for subsequent answer generation.In the answer generation stage,multi-modal fusion technology was employed to cross-fuse information from these two different modalities,enabling more accurate answers.

Key words: multi-chart joint question-answering, multi-modal fusion, text description, fact filtering

摘要： 为了满足在无法直接访问底层数据库的情况下,用户对于多张图表间关联性问题的交互式探索需求,提出了一种多图表联合问答方法。该方法通过两个核心阶段——数据准备和答案生成,实现了对多张图表数据的联合解读。在数据准备阶段,通过将图表数据提取重构为表格数据并对其每个单元格进行文本描述,为后续模型提供统一的数据格式。此外,为提高模型的准确性和回答效率,提出文本事实筛选方法,该方法能够在大量的表格文本描述中筛选出与用户问题相关的文本,为后续的答案生成提供精准的数据支持。在答案生成阶段,采用多模态融合技术,将这两种不同模态的信息进行交叉融合,以获取更精确的回答。

关键词: 多图表联合问答, 多模态融合, 表格文本描述, 事实筛选