Snowy: Recommending Utterances for Conversational Visual Analysis

Abstract

Figure 1: Examples of utterance recommendations in Snowy. (A) To assist with the "cold start" problem during data analysis, Snowy infers potentially interesting patterns from the underlying dataset and suggests analytic inquiries one may want to begin exploring the data with. (B) Upon executing a NL utterance, Snowy suggests follow-up utterances to drill down into specific data subsets or adjust the current view. (C) As marks are selected on the view through direct manipulation, Snowy recommends deictic utterances to perform popular calculations using the selected marks.

XNLI: Explaining and Diagnosing NLI-Based Visual Data Analysis

“…NL4DV [5] is a general NLI toolkit that integrates a set of NLP techniques (e.g., dependency parsing) to interpret NL data queries and recommends visualizations to generate results. Snowy [23] and QRec-NLI [24] recommend utterances in NL-based visual analysis, leveraging NL to provide analytical guidance. In recent years, NLIs have been applied to various scenarios and tasks, such as flow data exploration [25], visualization authoring [26], and comparative analysis [27].…”

“…The visualization results of NL4DV are specified with Vega-Lite grammar (Figure 4A), which is popular in the field of NLI [5], [23], [24], [28], [29], [30], [32], [52]. We leverage provenance to depict the three aspects of the visualization process and introduce the Provenance Generator, which consists of three major components to construct the visualization provenance from the Vega-Lite grammar, select representative sample data for demonstration, and apply visual cues to highlight the changes in provenance.…”

“…After user adjustments through widgets, the system considers the final visualizations as the desired results and generates query examples to help the users in query phrasing. The query examples are generated by a template-based method, which has been shown to be effective and controllable in specifying known analysis intents [23].…”

MEDLEY: Intent-based Recommendations to Support Dashboard Composition

“…For instance, P3, said "While this interface is powerful and easy to use, you have to know what kind of questions you need to ask in order to select those attributes and intents, and it would be really helpful to bring in something like Ask Data or Power BI Q&A [examples of commercial NLIs] to make those selections for you." Along these lines, a promising opportunity for future research is to extend NLIs for visualization (e.g., [15,25,29,39,42]) to go beyond supporting the creation of individual views and instead support dashboard generation. Such NLIs could enable people to use high-level utterances for stating their dashboard goals and reduce the tedium or challenges associated with explicitly selecting attributes and intents.…”

A Design Space for Surfacing Content Recommendations in Visual Analytic Platforms

“…This did identify a small number of additional papers from beyond the IEEE and ACM archives. However, we acknowledge that relevant articles may still fall out of our search (e.g., recommendations for next-step actions in analytical workflows [67,75]). Nevertheless, we believe that the general nature of our proposed design space means that it can also be applied in principle to characterize many approaches described in papers missed by our search.…”