Semantic interaction for visual text analytics

Abstract

Visual analytics emphasizes sensemaking of large, complex datasets through interactively exploring visualizations generated by statistical models. For example, dimensionality reduction methods use various similarity metrics to visualize textual document collections in a spatial metaphor, where similarities between documents are approximately represented through their relative spatial distances to each other in a 2D layout. This metaphor is designed to mimic analysts' mental models of the document collection and support their analytic processes, such as clustering similar documents together. However, in current methods, users must interact with such visualizations using controls external to the visual metaphor, such as sliders, menus, or text fields, to directly control underlying model parameters that they do not understand and that do not relate to their analytic process occurring within the visual metaphor. In this paper, we present the opportunity for a new design space for visual analytic interaction, called semantic interaction, which seeks to enable analysts to spatially interact with such models directly within the visual metaphor using interactions that derive from their analytic process, such as searching, highlighting, annotating, and repositioning documents. Further, we demonstrate how semantic interactions can be implemented using machine learning techniques in a visual analytic tool, called ForceSPIRE, for interactive analysis of textual data within a spatial visualization. Analysts can express their expert domain knowledge about the documents by simply moving them, which guides the underlying model to improve the overall layout, taking the user's feedback into account.

Semantic Interaction for Sensemaking: Inferring Analytical Reasoning for Model Steering

“…1) is a visual analytic prototype for spatial text analysis using semantic interaction [1]. ForceSPIRE presents users with a spatialization of the dataset, representing documents as minimized boxes, which can be expanded to full-detail documents via double-clicking.…”

“…ForceSPIRE also includes an "Entity Viewer", from which entities can be directly created, removed, as well as have their weights directly modified. ForceSPIRE supports the following semantic interactions (described in more detail in [1]): Document movement allows users to manipulate the spatial layout directly in the spatialization by placing documents in locations based on the user's domain knowledge. These movements can be both exploratory and expressive [28]/{v2pi/}, differentiated by how they adjust the underlying model.…”

“…"Ooops, oh well, [ForceSPIRE] will fix that later", he said (instead of asking how to undo), implying that the small number of erroneous interactions will be outweighed by the sum of the meaningful interactions over time. One approach in another version of ForceSPIRE is to store previous weighting schemes, so that when an undo occurs, the previous weights are restored, and the model updates the spatialization accordingly [1]. However, a "true undo" of a semantic interaction would be one from which the model learns about the user's analytical reasoning.…”

“…Semantic interaction [1,4] is an approach that enables such coupling, where analytic interactions designed for synthesis in visualizations are also designed to steer the underlying computation responsible for foraging of relevant information. Semantic interaction focuses on enabling direct manipulation of spatializations, which are two-dimensional views of highdimensional data such that similarity between information is represented by relative distances between data points (e.g., a cluster represents a collection of similar information) [5].…”

“…1) is a visual analytic prototype incorporating semantic interaction for analysis of text document collections represented in a spatialization [1]. Semantic interactions in ForceSPIRE include repositioning documents, highlighting text, searching, and annotating documents.…”