Workload Evaluation to Create Data Visualization Using ChatGPT

TL;DR: This paper will evaluate the workload for creating data visualization using ChatGPT 3.5 with high temporal and mental demand, mainly due to the vocabulary used and the completeness of the user instructions.

Abstract: The value of good data visualization has already been shown in several scenarios. Still, it is not always easy to obtain it, as it depends on factors such as the dataset, the amount of data, task types, the user profile, the type of interaction, etc. To mitigate the challenges addressed, automated or semi-automated systems have been proposed, emphasizing rule-based/heuristic approaches and machine-learning models. However, many of these applications require specialized knowledge and present results (data visualizations) that are not flexible for customization. Papers have highlighted the ease of tools like ChatGPT in creating various tasks, including creating data charts. This facility, in addition to the intelligent computational model involved, is also due to the expressiveness used in the requests to execute the tasks by the users since these tools use Natural Language Interfaces. Despite adopting these tools overgrowing in different scenarios of society, studies on the best way to use them, integrate them into existing processes, or evaluative studies on their effectiveness or efficiency are still incipient. Thus, this paper will evaluate the workload for creating data visualization using ChatGPT 3.5. For assessment, the NASA Task Load Index (Nasa TLX) methodology was applied, and users with experience creating data visualization created two proposed scenarios. The preliminary results showed high temporal and mental demand, mainly due to the vocabulary used and the completeness of the user instructions. The average time to create and perform InfoVis tasks in two proposed evaluation scenarios was 33 and 44 minutes, and 14 queries were applied on average for both scenarios. The direct consequence was that the users have redone the requests and improved the instructions at each new iteration, and all users completed the proposed tasks.