While word clouds pack a lot of data into a relatively small area, it is unclear whether readers actually benefit from all of that information, or if they are only processing a few of the words shown. We sought to determine if words outside of a reader’s central vision were contributing to their interpretation by leveraging the semantic priming effect: a phenomenon in which participants will more quickly recognize a word if they have been recently "primed" with a word that is semantically related. We presented participants with word clouds containing related and unrelated words to see whether they might prime this lexical decision task more strongly than single words alone. We showed that the peripheral contents of a cloud do affect participant performance at this task, though more work is needed to understand the impacts of these differences in real-world settings.

Word clouds are frequently used to analyze and communicate text data in many domains. In order to help guide research on improving the legibility of word clouds, we have conducted a survey of their usage in Digital Humanities academia and journalism. Using a modified grounded theory approach, we sought to identify the most common purposes for which word clouds were employed and the most common visual encodings they contained. Our findings indicate that font size, color, and word placement dominate as the primary data-encoding channels, as we hypothesized. Perhaps more surprisingly, we found that asking viewers to perform analytical tasks with word clouds was relatively common, especially in DH sources. This suggests that research into the interactions of these visual encoding channels (particularly in regards to legibility) is warranted.

Discussions of data visualization or visual analytics in the (digital) humanities often focus on relevant software tools. When it comes to teaching approaches to creating and interpreting visualizations, however, we need to stress the theory and decision-making processes that feed into the resulting visual representations. This workshop will explore and discuss strategies for teaching visualization literacy - from sketching by hand to creating visualizations using computational tools. Led by instructors from the humanities, visual analytics, design, and computer science, sessions will cover different methodologies and challenges of teaching visualization to audiences with varied technical and disciplinary expertise and goals.

Visualization researchers, always looking for new data into which to gain insight, have jumped at the possibility of mapping topics from the humanities into the domain of visualization. The prospect of rich data sets, some containing language that is thousands of years old, is tempting. But, in our excitement at the variety of different data sets available, we have grouped the work of many disciplines and subdisciplines from the humanities and social sciences under the same rubric, which we simply label “data.” It is important for us to be aware of how the different ways that the sciences and humanities create knowledge can be lost during the visualization design process. That is not to say that visualization specialists neglect their stakeholders, just that something is lost in the translation to the screen. Over the past three years and through two iterations of the Vis4DH workshop, we have come to understand that in these translations of work into data we have missed that these disciplines have as much to teach us about visualization as visualization has to teach them about their subjects of study.

For the past two years, researchers from the visualization community and the digital humanities have come together at the IEEE VIS conference to discuss how both disciplines can work together to push research goals in their respective disciplines. In this paper, we present our experiences as a result of this collaboration.

Many visualizations, including word clouds, cartographic labels, and word trees, encode data within the sizes of fonts. While font size can be an intuitive dimension for the viewer, using it as an encoding can introduce factors that may bias the perception of the underlying values. Viewers might conflate the size of a word's font with a word's length, the number of letters it contains, or with the larger or smaller heights of particular characters (‘o’ versus ‘p’ versus ‘b’). We present a collection of empirical studies showing that such factors—which are irrelevant to the encoded values—can indeed influence comparative judgements of font size, though less than conventional wisdom might suggest. We highlight the largest potential biases, and describe a strategy to mitigate them.

In the digital humanities, variation in the data is both a blessing and a curse. On one hand, variation represents change that can provide us with broad-scale historical insight (e.g., “How is what people talked about in the 18th century different from what they talked about in the 17th century?”). On the other hand, some kinds of variation can actually obscure others (e.g., did people actually stop talking about a particular word in the 17th century, or did the spelling just change?). As DH research scales up to more and more documents covering ever longer spans of history, variation of both kinds is necessarily unavoidable and must be explicitly accounted for. In this paper, we discuss the challenges concerning variation that we have encountered in a multi-year, collaborative project focused on a collection of print documents from 1470-1800. We have addressed this variation through a combination of data standardization and task-driven visualization design.

As topic modeling has grown in popularity, tools for visualizing the process have become increasingly common. Though these tools support a variety of different tasks, they generally have a view or module that conveys the contents of an individual topic. These views support the important task of gist-forming: helping the user build a cohesive overall sense of the topic's semantic content that can be generalized outside the specific subset of words that are shown. There are a number of factors that affect these views, including the visual encoding used, the number of topic words included, and the quality of the topics themselves. To our knowledge, there has been no formal evaluation comparing the ways in which these factors might change users' interpretations. In a series of crowdsourced experiments, we sought to compare features of visual topic representations in their suitability for gist-forming. We found that gist-forming ability is remarkably resistant to changes in visual representation, though it deteriorates with topics of lower quality.

Topic modeling, a method of statistically extracting thematic content from a large collection of texts, is used for a wide variety of tasks within text analysis. Though there are a growing number of tools and techniques for exploring single models, comparisons between models are generally reduced to a small set of numerical metrics. These metrics may or may not reflect a model's performance on the analyst's intended task, and can therefore be insufficient to diagnose what causes differences between models. In this paper, we explore task-centric topic model comparison, considering how we can both provide detail for a more nuanced understanding of differences and address the wealth of tasks for which topic models are used. We derive comparison tasks from single-model uses of topic models, which predominantly fall into the categories of understanding topics, understanding similarity, and understanding change. Finally, we provide several visualization techniques that facilitate these tasks, including buddy plots, which combine color and position encodings to allow analysts to readily view changes in document similarity.

Exploration and discovery in a large text corpus requires investigation at multiple levels of abstraction, from a zoomed-out view of the entire corpus down to close-ups of individual passages and words. At each of these levels, there is a wealth of information that can inform inquiry - from statistical models, to metadata, to the researcher's own knowledge and expertise. Joining all this information together can be a challenge, and there are issues of scale to be combatted along the way. In this paper, we describe an approach to text analysis that addresses these challenges of scale and multiple information sources, using probabilistic topic models to structure exploration through multiple levels of inquiry in a way that fosters serendipitous discovery. In implementing this approach into a tool called Serendip, we incorporate topic model data and metadata into a highly reorderable matrix to expose corpus level trends; extend encodings of tagged text to illustrate probabilistic information at a passage level; and introduce a technique for visualizing individual word rankings, along with interaction techniques and new statistical methods to create links between different levels and information types. We describe example uses from both the humanities and visualization research that illustrate the benefits of our approach.

In this position paper, we enumerate two approaches to the evaluation of visualizations which are associated with two approaches to knowledge formation in science: reductionism, which holds that the understanding of complex phenomena is based on the understanding of simpler components; and holism, which states that complex phenomena have characteristics more than the sum of their parts and must be understood as complete, irreducible units. While we believe that each approach has benefits for evaluating visualizations, we claim that strict adherence to one perspective or the other can make it difficult to generate a full evaluative picture of visualization tools and techniques. We argue for movement between and among these perspectives in order to generate knowledge that is both grounded (i.e. its constituent parts work) and validated (i.e. the whole operates correctly). We conclude with examples of techniques which we believe represent movements of this sort from our own work, highlighting areas where we have both "built up" reductionist techniques into larger contexts, and "broken down" holistic techniques to create generalizable knowledge.

A valuable task in text visualization is to have viewers make judgments about text that has been annotated (either by hand or by some algorithm such as text clustering or entity extraction). In this work we look at the ability of viewers to make judgments about the relative quantities of tags in annotated text (specifically text tagged with one of a set of qualitatively distinct colors), and examine design choices that can improve performance at extracting statistical information from these texts. We find that viewers can efficiently and accurately estimate the proportions of tag levels over a range of situations; however accuracy can be improved through color choice and area adjustments.

Statistical topic modeling is an increasingly popular approach to text analysis. Many existing visualization tools focus on analyzing the model itself, distinct from the documents upon which it was trained. In contrast, we seek to treat the model as a lens through which to view the original documents. This would enable the reader to observe trends and build hypotheses at multiple scales—ranging from across a corpus to within a single text—and find both algorithmic data and textual examples to defend these hypotheses. Supporting this workflow requires a multi-tiered framework that affords comparisons at three levels: the entire corpus, small sets of documents, and a single document. We provide such a tool in our implementation of Serendip, a web-application that combines view-coordinated reorderable matrices, small multiples displays, and tagged text in order to allow readers to develop insight at multiple levels and carry that insight into their analysis of the others.