We characterize and demonstrate how the principles of direct manipulation can improve interaction with large language models. This includes: continuous representation of generated objects of interest; reuse of prompt syntax in a toolbar of commands; manipulable outputs to compose or control the effect of prompts; and undo mechanisms. This idea is exemplified in DirectGPT, a user interface layer on top of ChatGPT that works by transforming direct manipulation actions to engineered prompts. A study shows participants were 50% faster and relied on 50% fewer and 72% shorter prompts to edit text, code, and vector images compared to baseline ChatGPT. Our work contributes a validated approach to integrate LLMs into traditional software using direct manipulation. Data, code, and demo available at https://osf.io/3wt6s.
Documents often have paragraphs packed with numbers that are difficult to extract, compare, and interpret. To help readers make sense of data in text, we introduce the concept of Charagraphs: dynamically generated interactive charts and annotations for in-situ visualization, comparison, and manipulation of numeric data included within text. Three Charagraph characteristics are defined: leveraging related textual information about data; integrating textual and graphical representations; and interacting at different contexts. We contribute a document viewer to select in-text data; generate and customize Charagraphs; merge and refine a Charagraph using other in-text data; and identify, filter, compare, and sort data synchronized between text and visualization. Results of a study show participants can easily create Charagraphs for diverse examples of data-rich text, and when answering questions about data in text, participants were more correct compared to only reading text.
Extracting underlying data from rasterized charts is tedious and inaccurate; values might be partially occluded or hard to distinguish, and the quality of the image limits the precision of the data being recovered. To address these issues, we introduce a semi-automatic system leveraging vector charts to extract the underlying data easily and accurately. The system is designed to make the most of vector information by relying on a drag-and-drop interface combined with selection, filtering, and previsualization features. A user study showed that participants spent less than 4 minutes to accurately recover data from charts published at CHI with diverse styles, thousands of data points, a combination of different encodings, and elements partially or completely occluded. Compared to other approaches relying on raster images, our tool successfully recovered all data, even when hidden, with a 78% lower relative error.
1 February 2019 in International Journal of Human-Computer Studies