KnowledgeVIS: Interpreting Language Models by Comparing Fill-in-the-Blank Prompts.

Recent growth in the popularity of large language models has led to their increased usage for summarizing, predicting, and generating text, making it vital to help researchers and engineers understand how and why they work. We present KnowledgeVIS, a human-in-the-loop visual analytics system for interpreting language models using fill-in-the-blank sentences as prompts. By comparing predictions between sentences, KnowledgeVIS reveals learned associations that intuitively connect what language models learn during training to natural language tasks downstream, helping users create and test multiple prompt variations, analyze predicted words using a novel semantic clustering technique, and discover insights using interactive visualizations. Collectively, these visualizations help users identify the likelihood and uniqueness of individual predictions, compare sets of predictions between prompts, and summarize patterns and relationships between predictions across all prompts. We demonstrate the capabilities of KnowledgeVIS with feedback from six NLP experts as well as three different use cases: (1) probing biomedical knowledge in two domain-adapted models; and (2) evaluating harmful identity stereotypes and (3) discovering facts and relationships between three general-purpose models.

Preliminary Guidelines For Combining Data Integration and Visual Data Analysis.

Data integration is often performed to consolidate information from multiple disparate data sources during visual data analysis. However, integration operations are usually separate from visual analytics operations such as encode and filter in both interface design and empirical research. We conducted a preliminary user study to investigate whether and how data integration should be incorporated directly into the visual analytics process. We used two interface alternatives featuring contrasting approaches to the data preparation and analysis workflow: manual file-based ex-situ integration as a separate step from visual analytics operations; and automatic UI-based in-situ integration merged with visual analytics operations. Participants were asked to complete specific and free-form tasks with each interface, browsing for patterns, generating insights, and summarizing relationships between attributes distributed across multiple files. Analyzing participants' interactions and feedback, we found both task completion time and total interactions to be similar across interfaces and tasks, as well as unique integration strategies between interfaces and emergent behaviors related to satisficing and cognitive bias. Participants' time spent and interactions emergent strategies revealed that in-situ integration enabled users to spend more time on analysis tasks compared with ex-situ integration. Participants' integration strategies and analytical behaviors revealed differences in interface usage for generating and tracking hypotheses and insights , yet their emergent behaviors suggested that in-situ integration could negatively affect the ability to generate and track hypotheses and insights. With these results, we synthesized preliminary guidelines for designing future visual analytics interfaces that can support integrating attributes throughout an active analysis process.

Left, Right, and Gender: Exploring Interaction Traces to Mitigate Human Biases.

Human biases impact the way people analyze data and make decisions. Recent work has shown that some visualization designs can better support cognitive processes and mitigate cognitive biases (i.e., errors that occur due to the use of mental "shortcuts"). In this work, we explore how visualizing a user's interaction history (i.e., which data points and attributes a user has interacted with) can be used to mitigate potential biases that drive decision making by promoting conscious reflection of one's analysis process. Given an interactive scatterplot-based visualization tool, we showed interaction history in real-time while exploring data (by coloring points in the scatterplot that the user has interacted with), and in a summative format after a decision has been made (by comparing the distribution of user interactions to the underlying distribution of the data). We conducted a series of in-lab experiments and a crowd-sourced experiment to evaluate the effectiveness of interaction history interventions toward mitigating bias. We contextualized this work in a political scenario in which participants were instructed to choose a committee of 10 fictitious politicians to review a recent bill passed in the U.S. state of Georgia banning abortion after 6 weeks, where things like gender bias or political party bias may drive one's analysis process. We demonstrate the generalizability of this approach by evaluating a second decision making scenario related to movies. Our results are inconclusive for the effectiveness of interaction history (henceforth referred to as interaction traces) toward mitigating biased decision making. However, we find some mixed support that interaction traces, particularly in a summative format, can increase awareness of potential unconscious biases.

Lumos: Increasing Awareness of Analytic Behavior during Visual Data Analysis.

Visual data analysis tools provide people with the agency and flexibility to explore data using a variety of interactive functionalities. However, this flexibility may introduce potential consequences in situations where users unknowingly overemphasize or underemphasize specific subsets of the data or attribute space they are analyzing. For example, users may overemphasize specific attributes and/or their values (e.g., Gender is always encoded on the X axis), underemphasize others (e.g., Religion is never encoded), ignore a subset of the data (e.g., older people are filtered out), etc. In response, we present Lumos, a visual data analysis tool that captures and shows the interaction history with data to increase awareness of such analytic behaviors. Using in-situ (at the place of interaction) and ex-situ (in an external view) visualization techniques, Lumos provides real-time feedback to users for them to reflect on their activities. For example, Lumos highlights datapoints that have been previously examined in the same visualization (in-situ) and also overlays them on the underlying data distribution (i.e., baseline distribution) in a separate visualization (ex-situ). Through a user study with 24 participants, we investigate how Lumos helps users' data exploration and decision-making processes. We found that Lumos increases users' awareness of visual data analysis practices in real-time, promoting reflection upon and acknowledgement of their intentions and potentially influencing subsequent interactions.