AutoChart: A Dataset for Chart-to-Text Generation Task

Proceedings of Recent Advances in Natural Language Processing, pages 1636–1644 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_183 1636 AutoChart: A Dataset for Chart-to-Text Generation Task Jiawen Zhu1, Jinye Ran3, Roy Ka-wei Lee1, Kenny Choo1, and Zhi Li2 1Singapore University of Technology and Design 2University of Saskatchewan 3China Merchants Bank {jiawen_zhu,roy_lee,kenny_choo}@sutd.edu.sg z.li@usask.ca, rjy777@163.com Abstract The analytical description of charts is an ex- citing and important research area with many applications in academia and industry. Yet, this challenging task has received limited at- tention from the computational linguistics re- search community. This paper proposes Au- toChart, a large dataset for the analytical de- scription of charts, which aims to encourage more research into this important area. Specif- ically, we offer a novel framework that gen- erates the charts and their analytical descrip- tion automatically. We conducted extensive human and machine evaluations on the gener- ated charts and descriptions and demonstrate that the generated texts are informative, coher- ent, and relevant to the corresponding charts. 1 Introduction Natural language generation (NLG) is one of the core research areas in artificial intelligence (Gatt and Krahmer, 2018). Recent NLG studies have explored data-to-text generation, where exciting ap- plications such as automated news reporting (Lep- pänen et al., 2017) were developed to generate text from non-linguistic data automatically. In this pa- per, we explore the chart-to-text generation prob- lem, where analytical textual descriptions are auto- matically generated for a given graphical chart. Chart-to-text generation has many exciting aca- demic and commercial applications. For instance, preliminary analyses can be generated on charts to aid users in authoring analytical documents. On the accessibility front, automatically generated chart analyses can also support accessibility since text descriptions can be fed into speech-to-text modules and help visually impaired individuals to under- stand charts. Chart-to-text generation could also be applied to aid academic writing. Text descriptions of visual elements such as diagrams, charts, and graphs, are among the core academic assignments This bar graph shows the number of visits to South Korea and Singapore by overseas residents, respectively, from 2014 to 2019. In 2014, there was a huge gap in the number of visits to these two countries. The number of visits to South Korea is about 20 million, whereas the number of visits to Singapore is over 80 million. There is a continuous decrease in the number of visits to Singapore, with the largest decrease in 2015 to about 60 million. In 2019, the number of visits becomes about 45 million. By contrast, the number of visits to South Korea has been on the rise since 2014 but seems to have plateaued in 2017. Figure 1: Example of IELTS AWT1. in linguistics (Molle and Prior, 2008). For example, the IELTS Academic Writing Task 1 (AWT1) is an assessment task that elicits written responses on a visual-verbal relationship. The AWT1 requires test takers to “describe, summarise, or explain the information in a graph, table, chart, or diagram.” Figure 1 shows an example of the AWT1. Chart- to-text generation offers the potential to generate large-scale chart …

Content truncated to 500 words