Natural Language to Visualization by Neural Machine Translation

nvBench: A Large-Scale Synthesized Dataset for Cross-Domain Natural Language to Visualization Task

“…The intuition is based on the semantic connection between SQL queries and VIS queries: SQL queries specify what data is needed and SQL queries additionally need to specify how to visualize the data. The quality of nvBench has been validated by experts and crowd-workers, and a deep learning-based model, namely ncNet [20], training using nvBench, also verifies the power of nvBench.…”

“…As shown in Figure 1, there has been a surge of works on developing techniques to support translating natural languages to visualizations [6,8,9,11,17,20,23,26,27,29,35].…”

“…Thanks to the large number of (NL, VIS) pairs in nvBench, developers can use these pairs to train an end-to-end neural network for the NL2VIS task. ncNet [20] is a Transformer-based model for translating NL queries into visualizations. It takes nvBench as the training corpus to solve the NL2VIS task in an end-to-end way.…”

“…For example, Figure 4 shows a (NL, VIS) pair in the nvBench. The VIS query can be presented as the tree format as introduced in [18] or as the Vega-Zero (a Vega-Lite like language proposed in [20]). Figure 5 showcases a (NL, VIS) pair with a more {"data": {"url": ".news_report/journalist.csv"}, "transform": [{ "lookup": "journalist_ID", "from": { "data": {"url": ".news_report/news_report.csv"}, "key": "journalist_ID", "fields": ["Work_Type"]} }],…”

“…Creating meaningful visualizations from data through natural language is a promising interaction paradigm, especially for novices, and is an essential step towards democratizing data visualization [12-14, 19, 25, 31]. Both mainstream commercial vendors (e.g., Tableau's Ask Data [2], Microsoft Power BI [3], ThoughtSpot [4], and Amazon's QuickSight [1]) and academic researchers [6,8,9,11,[15][16][17]20,23,24,26,27,29,35] have started to explore the techniques to support NL2VIS for decades.…”