Literal Encoding: Text is a first-class data encoding
LLiteral Encoding: Text is a first-class data encoding
Richard Brath * Uncharted Software Inc. A BSTRACT
Digital humanities are rooted in text analysis. However, most vi-sualization paradigms use only categoric, ordered or quantitativedata. Literal text must be considered a base data type to encode intovisualizations. Literal text offers functional, perceptual, cognitive,semantic and operational benefits. These are briefly illustrated with asubset of sample visualizations focused on semantic word sequences,indicating benefits over standard graphs, maps, treemaps, bar chartsand narrative layouts.
Index Terms:
Human-centered computing—Visualization—Visu-alization techniques—; Human-centered computing—Visualiza-tion—Visualization design and evaluation methods
NTRODUCTION
The role of text in visualization is poorly defined. In most referenceworks on visualization the focus is on categoric, ordered or quan-titative data as base types of data that are transformed into visualattributes then drawn in a display, e.g. [4, 12, 14, 31, 32, 56]. Textis usually understood as categoric data such as labels added on agraph or words in a tag cloud. However, categoric, ordered andquantitative encoding alone do not consider the unique role of lit-erally encoding text. This paper will show functional, perceptual,cognitive, semantic and operational benefits that literal text affordsvisualizations.Consider the portion of the world history chart in Fig. 1 [39]: it is atime-oriented visualization showing empires as streams: essentially astoryline visualization (e.g. [48] based on xkcd.com/657). However,it also has a large amount of text including the names of empiresin big text, rulers, dates, key events, short descriptive phrases, fullsentences and so on: the informational content far exceeds storylinevisualizations and the encodings are beyond categoric data.
Literal text goes beyond categories or individual words.
Categoric data in a visualization is typically transformed into avisual attribute such as hue or shape. However, perception limits thenumber of unique hues to ten or fewer [52]. While it is feasible tohave an unlimited number of unique shapes [4], in practice very fewunique shapes are provided by default in off-the-shelf software andprogramming libraries (e.g. nine in Excel, ten Tableau, seven D3.js).There are tens of thousands of unique English words, with thou-sands understood by most adult speakers. There is a larger set ofproper nouns. Then highly unique word sequences can be con-structed to form phrases, sentences, stories, poetry and so forth.Further, words are much richer than categories and can expressnuance e.g. many alternate words exist for each emotion [40]. * e-mail: [email protected] Figure 1: Text-intensive chronological visualization of world history.
Combinations of words, relationships between words and sequencesof words are typically lost in popular visualization approaches. Tagclouds typically represent individual words, e.g. [17]. Simple charttypes, such as bar charts and pie charts, often use labels whichrequire short text and tools often truncate text automatically. Somevisualization techniques do work with longer text sequences suchas WordTrees [54], Arc Diagrams [53] or some of the knowledgemaps at scimaps.org (e.g. [8]) meaning that visualization of longersequences do exist, but, these are not common.Closely related shortened strings are the underlying analytictechniques that reduce text to individual words or phrases such asword frequency analysis, entity detection, sentiment analysis, emo-tion analysis, and vector-based representations such as Word2Vec(e.g. [34]). As such, the analysis of text by separating words losesmuch semantic context: Working with text in visualization meansreconsidering text as a base type of data.
There are many historic visualization examples of text-based rep-resentations. Even within prose, highlighting techniques such aschanges in color, italics, capitalization and weight have been usedfor centuries to make letters or words perceptually standout fromsurrounding text. And there are many other examples of text heavyvisualizations such as:
Flowcharts document processes and may need text to describeeach step, such as in Fig. 2 (from [10]).
Timelines may be heavily annotated such as Chapple and Garofa-los
History of Rock N Roll (in [51]); may contain names and phrasessuch as Figure 1; detailed sentences such as Rand McNallys
His-tomap [47], or many other examples in
Cartographies of Time [43]).
Knowledge maps and mind maps require text, some of whichcontain large amounts of descriptive text such as Diderots
Tree ofKnowledge in Fig. 3 [44]. Modern procedurally-generated knowl-edge maps are text oriented as well, such as the proportional Eulerdiagrams in openknowledgemaps.org [28]. a r X i v : . [ c s . H C ] S e p igure 2: Portion of a text-intensive flowchart. In the field of information visualization, there are more than 400text visualizations at the
Text Visualization Browser (textvis.lnu.se).However, approximately one quarter of these visualizations have notext at all. Most of the remaining techniques are focused on labels, i.e.a word or two. Only a few manipulate individual characters (such asbackground per glyph [1] or font weight [36]). Some display longertext, although many of these limit visualization to manipulationof keywords in context (KWIC, e.g. [20]) and/or blocks of text;e.g. [46] does both.Use of literal text, beyond simple labels, may be uncommon ininformation visualizations for many reasons [9], such as:1.
Historic convention separates words from images due to limi-tations of print-based processes or the extra expense associatedwith engravings, lithographs, etc.2.
Display limitations in early visualizations resulted in poorquality text (low resolution 96DPI) and limited real-estate(small screen sizes), which prompted guidelines to displaydetails on demand [45].3.
Graphic design conventions , starting with mid-century mod-ernism, favor less text, more white space and more graphicssuch as icons.4.
Preattentive pattern perception precludes reading text linearly.These issues can be addressed, for example:1. Modern web browsers do not have the same restrictions asprint.2. Modern displays are much higher resolution: a 4k display has25x pixels of an early 1990’s display.3. Graphic design counter-trends bring more text manipulationinto media, starting with post-modernists in the 1980’s.4. Some visualization tasks do not require preattention (e.g. in-ventories such as roadmaps), or subsets of text can be madepreattentive with typographic formats [4, 5, 9].
Figure 3: Top part of Diderot’s
Tree of Knowledge . Figure 4: Literal text benefits (red text) in relation to the visualizationpipeline.
The rest of this paper, and the key contribution, will characterize thebenefits of literal text directly encoded in visualization and illustratesome examples focused on visualizations of strings of text.
ITERAL E NCODING
Data can be encoded literally into strings of text and representeddirectly. Literal encodings are unique to text. Instead of consideringtext as a different form of data, some researchers consider text asa visual attribute, similar to shape, size or color, e.g. [7, 12, 32, 56],whereas other researchers do not include text. Bertin, for example,does not include text as a data type or encoding, however, Bertinnarrowly defined his visual attributes as retinal variables, explicitlyfocusing on low level visual channels into which data is transformedfor thematic representations for fast perception. Note that Bertin diddiscuss text and font attributes such as bold, italic and typeface infour pages in an appendix to the original French edition of SmiologieGraphique [4] and a followon article [5], which were not includedin the later English translation.There are many benefits to literal encoding, which can be orga-nized by function, perception, cognition, semantics and operation.These benefits can be aligned to the visualization pipeline as shownin Fig. 4 (this diagram based on a simplified version from [12]).
Literal text may be effective in meeting the needs associated withthe particular use case:
A. Information organization or communication purpose.
Thegoal of a visualization is not necessarily analysis of patterns by preat-tentive perception, such as a pattern of dots in a scatterplot. Bertindefines other purposes, including communicating key informationand organizing a large amount of data. Text is common in historicvisualizations that organize large amounts of information as shownin the prior section. It is doubtful that these data-dense visualizationswhich organize information could function without text.
B. Text is the primary subject.
Some visualizations are primar-ily about text. The simple tag cloud could not exist without text.Similarly, other visualizations related to text analysis presumablywill have a strong need to explicitly represent literal text: extractedentities, topic analysis, sentiment and emotion analytics, keywordanalysis, word stemming, dialogue analysis, taxonomies, and so on.
C. Increased information content.
Many historic visualizationsare highly similar to modern interactive visualizations. For example,Figure 1 is a storyline visualization. Flowcharts like Figure 2 areessentially graphs with much larger nodes sized to fit descriptivetext. Fundamentally, visualization is a lossy medium [12], and theaddition of text into visualization can increase the informationalcontent.
D. High number of categories.
Many visual attributes do notsupport categories of high cardinality as discussed in 2.1. Text cansupport high cardinality as words, phrases and sentences; or codese.g. labeling aircraft in air traffic control, equipment in networkdiagrams, nodes in circuit diagrams, and so on.
E. Reduced design effort.
It is feasible pictographic icons canbe quick to decipher, but effective icons typically require significanteffort by expert graphic designers to create a series of icons whichwork together. . Unambiguous Decoding.
While an icon can represent a wordor phrase, they can be ambiguous, e.g. Clarus the dogcow [18]. Textliterally encodes the specific words, phrases or sentences of interest,so there is no information loss decoding back to text. Further, textualglyphs are highly learned and unambiguous in a well-designed font.
Many visualization evaluations focus on testing time and errors, suchas the rapid perception of the presence of a visual attribute. Whiletext can be enhanced with attributes such as bold or italic to makea word preattentively pop-out, to read the text still requires activeattention - which is slower than preattention.However, real-world user tasks may be much more complex,requiring full decoding and comprehension of context. If the goalof a visualization is overall more efficient completion of tasks, thenthe performance of the entire system must be considered. Activeattention of text can provide multiple perceptual benefits:
A. Reading is Automatic.
Reading labels is very fast comparedto interactions such as tooltips. The parallel letter recognition modelof words states letter information is used to recognize the words [30],letters within a word are recognized simultaneously, and the andword perception is extremely fast [15].
Automatic word recognition is a common explanation for theStroop effect which theorizes that reading is automatic and diffi-cult to voluntarily stop. The relative speed of processing model ,hypothesizes that it is faster to read the word than to name the color.Both explanations imply that reading words is fast. More generally,automaticity is the ability to perform a well practiced task with lowattentional requirements: (a) the user is unaware that the task isoccurring, (b) does not need to initiate the task, (c) does not havethe ability to stop the process, and, (d) the task has low cognitiveload [3]. The implication for visualization is that some text will beautomatically read and that the cognitive cost of reading will be low.
B. Interaction is Slow.
Interactions such as tooltips can be usedto identify items on a screen. Interaction requires additional cogni-tive effort, as the viewer must first determine their interactive goal,engage in motor skills and progressively refine those skills to achievethe target [41].Zoom and pan with level of detail for progressive appearanceof labels is an alternative approach to revealing text, often usedon interactive maps. While zoom-based labeling does occur invisualization (e.g. [26]), it is not common. Visualization does nothave the equivalent labeling heuristics developed over centuries incartography [42].
C. Legends are slow.
In any form of visualization the viewermust both perceive the information of interest and then decode it.When visually assessing categoric data encoded as discrete instancesin a visual attribute (e.g. using color to indicate categories), theviewer needs to recall the mapping between the colors and categories.With text, the viewer can directly decode the item.
D. Reduced load on short term memory.
Tasks reliant on short-term human memory can benefit from explicit text. Text representeddirectly does not require short term memory, for example whenreferring back and forth between items and legends.
Beyond attention, problem-solving tasks require more cognitiveresources to complete more complex tasks, such as determining thesegment in a Venn diagram that corresponds to a logical condition,or, tracing a path in a network. Complex reasoning can be aided bythe appropriate visual representations, as shown with force diagrams,pulley combinations and so forth (e.g. see many of the papers at
Diagrams conference).Readers construct mental representations of what they read atmultiple levels: a) surface: words and syntax; b) propositional content; c) situation model incorporating and organizing contentwith respect to real world knowledge from memory [38].
A. Proposition and problem-solving tasks.
The propositionalmodel is self contained based on the presented facts: it can be un-derstood through logical inferences, or through spatialization andvisually queried. Problem solving is improved when textual instruc-tions are directly integrated into diagrams [11]. Task performanceimproves when information is provided through both text and imageswith close spatial positioning and/or linkages [37]. More broadly,Larkin and Simons findings indicate that diagrammatic representa-tions of content aid problem-solving [29] by:1.
Locality:
Spatially organizing information together reducessearch effort.2.
Reducing cross-referencing:
Fewer steps to decode mappings.3.
Perceptual enhancement:
Visual relationships support percep-tual inferences.Extending Larkin and Simons findings to text and visualizationimplies improved performance for text directly integrated into avisualization:1.
Reduced Search (Locality).
Spatially organized displays aidfinding information of interest. For example, in the knowledgegraph (e.g. Fig. 3), the hierarchical relations and descriptionsare spatially grouped facilitating search and providing text de-tails. Identification of local relations is a common task in manyvisualizations and has been expressed by many researchers, forexample:
Tobler’s first law of geography
Everything is related toeverything else, but near things are more related than distantthings. [35] Many visualization layouts attempt to locate re-lated objects close together (e.g. graph layouts, scatterplots,treemaps and set visualization). Text in any of these layoutssupports identification and context of local related entities.
Thundt et al’s serendipity
Serendipitous discovery isthe fortuitous unexpected discovery by accident, often occur-ring during search. It is researched in library sciences andsummarized by Thudt et al [49] who note that serendipity isclosely associated with coincidence: wherein related ideasmay manifest as simultaneous occurrences that seem acausalbut meaningful. Both local proximity and related words andphrases can be recognized thereby providing different cogni-tive associations than visual associations alone.2.
Reduced Cross-Referencing . Text within a visualization ele-ment provides immediate detail as opposed to drill-down orreferring to other linked visualizations:
Tufte, Shneiderman and details
Tufte popularized mi-cro/macro readings; that is visual displays with high-densityinformation, visually read at levels ranging from high-levelmacro patterns such as trends, clustering and outliers; downto localized patterns, such as individual observations and localpeers [50]. Tufte summarizes this with to clarify, add detail.Tuftes approach is somewhat similar to Shneidermans visualinformation-seeking mantra: overview first, zoom and filter,then details on demand. [45] With Shneiderman, interactivityis used to reveal the detail information whereas Tufte plots itdirectly. Tufte plotted data on high-resolution paper, whereasShneiderman was constrained to low-res 1990s screens, neces-sitating interaction.
Bertin and categories
Bertin states that shapes can rep-resent categories of high cardinality and provides a map with59 different glyph types (p. 157 in [4]). Bertin shows globalpatterns with respect to one shape cannot be seen. The viewermust linearly scan across glyphs and can only see local patterns.Furthermore, the viewer cannot ascertain if a local pattern is igure 5: Perceptual inferences can be made via text density, linesand areas, or patterns forming columns. meaningful without comparison to the global pattern. Textcan be different. Text can be used non-categorically to encodeunique entities and concepts (e.g. countries, people, idioms,quotes, etc.) a comparison to global is not needed. Further-more, any unique text can be accessed non-linearly based onthe layout, for example, to directly compare the unique text attwo different locations of the plot area.3.
Perceptual inferences can be made across collections of labelsor phrases. For example, a density of text can indicate commonnodes across classification schemes in the graph by Haeckel(Fig. 5 left, [19]), the alignment of text along paths to formlines and areas (e.g. Axis Maps: Typographic Maps in Fig. 5middle [2]), or the repetition of Xs and Os in point and figurecharts (Fig. 5 right), e.g. [16].
Semantic content is added to text when words are combined insequences. Consider the following:• Jack and Jill went up the hill.• Humpty Dumpty sat on a wall.• The Owl and the Pussycat went to sea.Text, when processed into discrete words, such as a word cloud,loses the semantic content of the word sequence. These sentenceshave meaning beyond the collection of words. Jack, Jill and Humptyall have height and might fall. Jack, Jill, the Owl and the Cat wentsomewhere, Humpty did not.Analytic and visualization approaches need to consider applica-tions where word sequence is maintained, or assemblies of relatedwords. Search user interfaces have evolved significantly beyondsimple keywords and document metadata to include contextual ti-tles, phrases and sentences in search results (e.g. [20] Chapter 5).Studies have shown superior performance for results containing textsnippets [13].More broadly, a variety of visualization techniques have evolvedin the humanities for close reading, summarized by Jnicke et al [25].For close reading, these approaches maintain word sequence andchange visual attributes to markup the text (similar to keywordmarkup), superimpose markers and connections on top of words (e.g.like a graph), or increase space between letters and words to addmarkers such as flows or phonetics (e.g. like WordTree).
Literal text allows for textual operations which can be used to pre-process text, or interactively to operate on text to provide for addi-tional uses:•
Ordering.
Text can be ordered and sorted (alphabetically),facilitating search and lookup.•
Search and Filter.
Text can be navigated and/or filtered bycomplete or partial strings. •
Summarization.
Text can be summarized: longer texts can bereduced to shorter texts while retaining a subset of the originalmeaning.•
Comparison, similarity and translation.
Text can be com-pared and assessments made regarding similarity. Tools suchas thesauruses and dictionaries aid in understanding similar-ity. Translation extends comparative analysis to produce nearequivalent meaning in another language.•
Tone, opinion, sentiment and emotion.
Text can be evaluatedassigned quantitative values, for example, for opinion (e.g. rat-ings), tone (e.g. news tone), sentiment (e.g. score for positiveor negative), or emotion (e.g. based on text or other cues).•
Categorization, taxonomies and topic analysis.
Organizationof many texts follow classification schemes and tagging oftopics by keywords; such as general purpose classifications(e.g. Dewey, Library of Congress) and domain specific classi-fications (e.g. ACM, IEEE).•
Natural Language Processing and Machine Learning.
All ofthe above operations are rapidly evolving with NLP and ma-chine learning, such as topic extraction, emotion analysis, sen-timent detection, machine translation, automated summariza-tion and so on. For example, Zhang et al.’s recent transformermodels have achieved near human-level abstractive summariza-tion [57]. Open source NLP libraries, such as NLTK, spaCy,and compromise [6, 23, 27], provide parts of speech tagging,tokenization, entity extraction, dependency parsing, and entitylinking, which can be assembled into a wide variety of textualanalyses extracting and linking related words and phrases, suchas such as character descriptions or rhetorical devices.
ISCUSSION
There are many ways to consider these benefits in the humanities.As there are many existing visualization techniques focused oncollections of disconnected words (e.g. word clouds), the followingfocuses on applications relevant to semantic word sequences (i.e.phrases and sentences) and the corresponding benefits.
A. Sentences on Paths.
In Fig. 6, Nigel Holmes reduced keytestimony in the Iran-Contra affair into a set of statements impli-cating individuals represented as color-coded textual connectionsbetween politicians (i.e. a graph) [22]. Literal text is required forthe communications purpose (3.1:A). Magazines are non-interactivethe text is a more efficient alternative to referring to a legend (3.2:C),and aids reasoning with directed connections enabling perceptualinferences (3.3:A3).
Figure 6: Connections between individuals with testimony.igure 7: Portion of Amsterdam SmellMap with narrative paths.
Kate McLean uses linear text along geographic contour lines toform artistic geospatial narratives in Fig. 7 [33]. The literal text is theprimary subject (3.1:B), which the viewer will automatically start toread (3.2:A), and compare local information within a geolocation(3.3:A1), with added potential for residents to supplement real-worldlocal knowledge (3.3:B).Programmatic examples include news headlines with associatedstock price movement, and twitter commentary with associatedretweet counts [9]; which enable comparison between content (thetext) and the response (the behaviour of markets / sharing by users).
B. Small Blocks of Text.
NewsMap extends the visualizationlayout of a treemap with news headlines Fig. 8 [55]. Informationcontent beyond a treemap is increased with hue, brightness andtext (3.1:C). Interaction is not a suitable: direct reading is required(3.2:B). The treemap layout of text ensures that similar headlinesare spatially proximate (3.3:A1) and augments the viewers real-world understanding of news (3.3:B). Interactive search and filteraid exploration (3.5).
Figure 8: NewsMap headlines. Size, hue and brightness encode data.
One challenge with the treemap layout is the prioritization of areaaccuracy over the display of text, as such, some squares are too smallfor legibile text, or text does not fit well. Consider a dataset of 2894coroners inquests from Georgian London [21, 24] where each caseis summarized by a sentence and verdict, e.g.• Mary Roberts drowned herself. Suicide.• Mary Gardiner struck with hand. Homicide.• Ann Fitsall suffocated and burnt. Accident.• Nicholas Bone, John Dayson and James Cusack killed by abrick wall. Accident.An analysis may be interested in causes of death. From eachcase can be extracted a subject, verb and object. These can beorganized into a hierarchy, by verb (e.g. drowning, suffocation),object (e.g. hand, wall) and subject (e.g. Mary Roberts, Ann Fitsall).There are a few caveats, for example, where two or more verbs orobjects are used in a single sentence, only one is selected so thatindividuals are not double counted. The hierarchy can be visualized,for example, with a treemap (Fig. 9), which draws attention to bigboxes and bright colors (e.g. drowned) but at the cost of fitting labelsor skipping labels on small boxes.
Figure 9: Treemap showing causes of death in Georgian London.Figure 10: Activities, objects, and the names of deceased.
Instead, Fig. 10 shows the same data in a representation con-ceptually similar to a dictionary listing with font weight, color andbackground shading conveying data, live at https://codepen.io/Rbrath/full/rNewgde . All text is visible and readable on a 4Kdisplay: an enlarged portion is shown at right (i.e. organizationalpurpose 3.1:A). Broad patterns are visible, for example, large blocksof names such as the red text (homicide) on red background (genderfemale) under the verb
STRANGLED (i.e. a preattentive percep-tion of the large block, then the bold allcaps text, which is identifiedby automatic reading 3.2:A). On close inspection, the local text canbe read, for example, many local objects associated with the verb
STRUCK (e.g. adze, bar, beam, bottle, broom ) and the names ofthe victims (e.g. adze : Sarah Skyring, bar : William Blakshaw, etc)no drill-down is needed (3.3:A2). Furthermore, search can be usedto find the common forms of death for child, or filters can be used toisolate only homicides (3.5).Fig. 11, is a simple adjacency matrix indicating dialogue fromone character (vertical axis) to another character (horizontal axis),with bubble size proportional to the amount of dialogue.
Figure 11: Volume of character dialogue via an adjacency matrix.igure 12: Character dialogue via an adjacency matrix, with high-lighted repeated sets of words per character.
Fig. 12 is the same adjacency matrix where each cell contains thedialogue from one character to another, increasing content (3.1:C).Cells are expanded if the next cell to the right would otherwise beempty. The matrix contains much of the dialogue, although somecells are truncated. In addition to showing dialogue, some phraseshave been highlighted: an simple NLTK algorithm has tagged setsof words that are frequently repeated by characters (not includingsets made of only prepositions, articles and conjunctions), so forexample, the moral of that is is highlighted each time the Duchessrepeats the phrase to Alice (aiding perceptual inferences 3.3:C).Word sets are repeated words, but the order may vary. This alsohighlights Carrolls logical inversions, such as the Hatters I see whatI eat and I eat what I see. It also highlights variants of catchphrases,such as off with his/her head yelled by the Queen. Highlightingsimilar words in different orders aids analysis of the variation of thesemantic content (3.4). This use of NLP could be further tuned, forexample, to tag rhetorical devices; which in turn, could be visualizedin techniques such as an adjacency matrix, story flow, text on path,and so on.
C. Superimposition.
Text can be superimposed over top of data-driven graphical marks similar to using a highlighter, where thehighlight may represent quantities or keywords or so on.Each row in Fig. 13 is a top selling song, shown as a line of textidentifying artist, song tile, and opening lyrics. The line is super-imposed over a blue bar chart indicating number of singles soldand also superimposed over highlights of common keywords such asChristmas , love , and baby . The superimposed text per bar in-creases information beyond the standard bar chart (3.1:C); providesdetail instead of an interactive tooltip (3.2:B), allows perceptualinferences (patterns of repeated words) (3.3:A3) and reading cantrigger real world recall of a familiar tune (3.3:B).Fig. 14 shows body text superimposed over summary words ex-tracted via NLP. When reading a long text on a mobile device only aparagraph or two are visible, meaning that traditional textual cuessuch as headings and spacing may be off screen, thereby makingit more difficult to navigate around the text. Using a large propernoun and associated verb could facilitate skimming while scrollingby providing textual landmarks to characters and their actions. Thislarge text could disappear on scroll-stop. This is a summary oper-ation (3.5) to reduce text, arguably to a high number of categories(3.1:D); and reducing cognitive load by not requiring the viewer torecall off-screen headings (3.2:D).
ONCLUSION
The itemized literal benefits provide a starting point for researchersand designers to consider the alternative visualizations and the po-
Figure 13: Song title and lyrics over bar chart and repeated words.Figure 14: Body text over a two-word paragraph summary. tential benefits across different design alternatives. These literalbenefits can be considered in any kind of text-enhanced visualiza-tion, whether simple labels, or through to the examples shown here.The example analyses here focus on collections of short semanticword sequences. The examples here may be highly-specific, how-ever, the intent is to illustrate the possibilities and benefits of literaldepictions. These should encourage digital humanities researchersto search for new visualization techniques. It should also encouragevisualization researchers to validate benefits beyond prior researchin adjacent fields through experimentation directly on text-orientatedvisualizations to quantify and characterize benefits. And, it shouldencourage linguistics and NLP researchers to innovate with human-ities researchers to find and annotate word sequences, which canbe used in human-in-the-loop analytical user interfaces for use bydigital humanities scholars. A CKNOWLEDGMENTS
Images created by the author are CC-BY-4.0. Some of these im-ages and portions of the text will appear in the forthcoming book
Visualizing with Text , by Richard Brath, as part of the VisualizationSeries edited by Tamara Munzner to be published by A.K. Peters inNovember 2020. R EFERENCES [1] A. Abdul-Rahman, J. Lein, K. Coles, E. Maguire, M. Meyer, M. Wynne,C. R. Johnson, A. Trefethen, and M. Chen. Rule-based visual mappings–ith a case study on poetry visualization. In
Computer Graphics Forum ,vol. 32, pp. 381–390. Wiley Online Library, 2013.[2] S. Afzal, R. Maciejewski, Y. Jang, N. Elmqvist, and D. S. Ebert. Spatialtext visualization using automatic typographic maps.
Visualization andComputer Graphics, IEEE Transactions on , 18(12):2556–2564, 2012.[3] J. A. Bargh. The four horsemen of automaticity: Awareness, intention,efficiency, and control in social cognition. In
Handbook of SocialCognition , pp. 1–40, 1994.[4] J. Bertin.
S´emiologie Graphique . Gauthier-Villars, Paris, 1967.[5] J. Bertin. Classification typographique: Voulez-vous jouer avec mon a.
Communication & Langages , 45(1):70–75, 1980.[6] S. Bird, E. Klein, and E. Loper.
Natural language processing withPython: analyzing text with the natural language toolkit . ” O’ReillyMedia, Inc.”, 2009. See also: .[7] K. B¨orner.
Atlas of Knowledge: Anyone Can Map . MIT Press, Cam-bridge, Mass, 2015.[8] K. B¨orner, E. Hardy, B. Herr, T. Holloway, and W. B. Paley. Taxonomyvisualization in support of the semi-automatic validation and optimiza-tion of organizational schemas.
Journal of Informetrics , 1(3):214–225,2007.[9] R. Brath.
Visualizing with Text . A.K. Peters, Boca Raton, FL, 2020.[10] W. C. Brinton.
Graphic Presentation . Brinton Associates, 1939.[11] P. Chandler and J. Sweller. Cognitive load theory and the format ofinstruction.
Cognition and instruction , 8(4):293–332, 1991.[12] M. Chen and L. Floridi. An analysis of information in visualization.
Synthese , 2013.[13] C. L. Clarke, E. Agichtein, S. Dumais, and R. W. White. The influ-ence of caption features on clickthrough patterns in web search. In
Proceedings of the 30th annual international ACM SIGIR conferenceon Research and development in information retrieval , pp. 135–142.ACM, 2007.[14] W. S. Cleveland and R. McGill. Graphical perception: Theory, ex-perimentation, and application to the development of graphical meth-ods.
Journal of the American statistical association , 79(387):531–554,1984.[15] M. Davis. Aoccdrnig to a rscheearch at cmabrigde... http://mrc-cbu.cam.ac.uk/people/matt.davis/Cmabrigde , 2003.[16] V. DeVilliers and O. Taylor.
The Point and Figure Method of Anticipat-ing Stock Price Movements . Harriman House, 1933.[17] J. Feinberg. Wordle.
Beautiful Visualization: Looking at Data Throughthe Eyes of Experts , 2010.[18] S. Hackett. The history of clarus the dogcow. http://512pixels.net/dogcow/ , 2015.[19] E. Haeckel.
The Evolution of Man: A Popular Exposition of thePrincipal Points of Human Ontogeny and Phylogeny . D. Appleton andCompany, 1897.[20] M. Hearst.
Search User Interfaces . Cambridge University Press, 2009.[21] T. Hitchcock, R. Shoemaker, S. Howard, J. McLaughlin, and et al.London lives, 16901800. , 2012.[22] N. Holmes. Pointing the finger. In
Time Magazine , p. 27, 1987.[23] M. Honnibal and M. Johnson. An improved non-monotonic transitionsystem for dependency parsing. In
Proceedings of the 2015 conferenceon empirical methods in natural language processing , pp. 1373–1378,2015. See also: spacy.io .[24] S. Howard. London lives coroners inquests, 17601799. london.sharonhoward.org/CoronersInquests , 2018.[25] S. J¨anicke, G. Franzini, M. F. Cheema, and G. Scheuermann. Onclose and distant reading in digital humanities: A survey and futurechallenges. In
Eurographics Conference on Visualization (EuroVis)-STARs , pp. 83–103, 2015.[26] D. Jonker, S. Langevin, D. Giesbrecht, M. Crouch, and N. Kronenfeld.Graph mapping: Multi-scale community visualization of massive graphdata.
Information Visualization , 16(3):190–204, 2017.[27] S. Kelly. Compromise: modest natural language processing. https://github.com/spencermountain/compromise , 2019.[28] P. Kraker, C. Kittel, and A. Enkhbayar. Open knowledge maps: Cre-ating a visual interface to the worlds scientific knowledge based onnatural language processing.
Zeitschrift fr Bibliothekskultur/ Journalfor Library Culture , 4(2):98–103, 2016.[29] J. H. Larkin and H. Simon. Why a diagram is (sometimes) worth ten thousand words.
Cognitive science , 11(1):65–100, 1987.[30] K. Larson. The science of word recognition.
Advanced ReadingTechnology, Microsoft Corporation , 2004.[31] J. MacKinlay. Automating the design of graphical representations.
ACM Transactions on Graphics , 5, 2006.[32] R. Mazza.
Introduction to Information Visualization . Springer-Verlag,2009.[33] K. McLean.
Smellmap Amsterdam . Sensory Maps, 2014.[34] T. Mikolov, K. Chen, G. Corrado, and J. Dean. Efficient estimation ofword representations in vector space. arXiv preprint , 2013.[35] H. Miller. Toblers first law and spatial analysis.
Annals of the Associa-tion of American Geographers , 94:284–289, 2004.[36] M. Nacenta, U. Hinrichs, and S. Carpendale. Fatfonts: combining thesymbolic and visual aspects of numbers. In
Proceedings of the Interna-tional Working Conference on Advanced Visual Interfaces , vol. 57, pp.407–414, 2012.[37] N. H. Narayanan and M. Hegarty. Multimedia design for communi-cation of dynamic information. In
International journal of humancomputer studies , vol. 57, pp. 279–315, 2002.[38] S. Payne. Mental models in human-computer interaction. In
TheHuman-Computer Interaction Handbook , pp. 80–93. CRC Press, 2007.[39] E. Pick. Tableau de lhistoire universelle. (eastern hemisphere). davidrumsey.com/luna/servlet/s/61kab2 , 1858.[40] R. Plutchik. The nature of emotions: Human emotions have deepevolutionary roots, a fact that may explain their complexity and providetools for clinical practice.
American Scientist , 89:344–350, 2001.[41] R. W. Proctor and K.-P. L. Vu. Human information processing: anoverview for human-computer interaction.
The human-computer inter-action handbook , 89:60–79, 2007.[42] A. Robinson, J. Morrison, P. Muehrcke, A. Kimerling, and S. Guptill.
Elements of Cartography . John Wiley & Sons, New York, NY, 1995.[43] D. Rosenberg and A. Grafton.
Cartographies of Time: A History of theTimeline.
Princeton Architectural Press, 2012.[44] C. Roth. Genealogical distribution of the arts and sciences. omeka.lehigh.edu/exhibits/show/diderot/sciences/tree , 1780.[45] B. Shneiderman. The eyes have it: A task by data type taxonomy forinformation visualizations. In
Proceedings 1996 IEEE symposium onvisual languages , pp. 336–343. IEEE, 1996.[46] M. Skeppstedt, K. Kucher, M. Stede, and A. Kerren. Topics2themes:Computer-assisted argument extraction by visual analysis of importanttopics. In
Proceedings of the 3rd Workshop on Visualization as AddedValue in the Development, Use and Evaluation of Language Resourcesat LREC , vol. 18, pp. 9–16, 2018.[47] J. Sparks. The histomap. four thousand years of world history. relativepower of contemporary states, nations and empires. davidrumsey.com/luna/servlet/s/61grv9 , 1931.[48] Y. Tanahashi and K.-L. Ma. Design considerations for optimizing story-line visualizations.
IEEE Transactions on Visualization and ComputerGraphics , 18(12):2679–2688, 2012.[49] A. Thudt, U. Hinrichs, and S. Carpendale. The bohemian bookshelf:supporting serendipitous book discoveries through information visual-ization. In
Proceedings of the SIGCHI Conference on Human Factorsin Computing Systems , pp. 1461–1470. ACM, 2012.[50] E. Tufte.
The Visual Display of Quantitative Information . GraphicsPress, 1983.[51] E. Tufte.
Envisioning Information . Graphics Press, 1990.[52] C. Ware.
Information Visualization: Perception for Design . Springer-Verlag, 2000.[53] M. Wattenberg. Arc diagrams: Visualizing structure in strings. In
IEEESymposium on Information Visualization, 2002. INFOVIS 2002. , pp.110–116. IEEE, 2002.[54] M. Wattenberg and F. B. Vi´egas. The word tree, an interactive vi-sual concordance.
IEEE transactions on visualization and computergraphics , 14(6):1221–1228, 2008.[55] M. Weskamp. Projects: Newsmap. marumushi.com/projects/newsmap , 2004. accessed: 03/03/2014.[56] L. Wilkinson.
The Grammar of Graphics . Springer, 1999.[57] J. Zhang, Y. Zhao, M. Saleh, and P. J. Liu. Pegasus: Pre-training withextracted gap-sentences for abstractive summarization.