ChartSumm: A large scale benchmark for Chart to Text Summarization

Abstract

Information visualization such as bar- and line-charts are quite popular for understanding large tabular data. But, interpreting information solely with different visualization techniques can also be difficult due to different reasons like visual impairment or the requirement of prior domain knowledge to understand the chart. Automatic "chart to text summarization" can be promising and effective tool for providing accessibility as well as precised insights of chart data in natural language. In spite of having a good potential, there have not been a lot of works on chart to text summarization making it a low resource task. Scarcity of large scale datasets for chart to text summarization is one of the reason behind this. The human written descriptions in the available dataset also contains information beyond the knowledge of the chart making it difficult for us to have an unbiased evaluation. In our thesis, we propose ChartSumm a large scale dataset for chart to text summarization consisting of 84,363 charts along with their metadata and descriptions. We also propose two test sets: test-e and test-h for evaluating the performance of the trained models available in this domain. Our experiment shows that a T5 model trained on our dataset has achieved BLEU score of 75.72 in test-e set and 64.78 in test-h set. From our analysis we can conclude that large language models like T5 and BART can generate short precised deception from given chart metadata.