Proceedings of the Workshop on Interactive Language Learning, Visualization, and Interfaces, pages 22–29,Baltimore, Maryland, USA, June 27, 2014. c�2014 Association for Computational Linguistics

Dynamic Wordclouds and Vennclouds for Exploratory Data Analysis

Glen CoppersmithHuman Language Technology Center of Excellence

Johns Hopkins Universitycoppersmith@jhu.edu

Erin KellyDepartment of Defense

elkelly8@gmail.com

Abstract

The wordcloud is a ubiquitous visualiza-tion of human language, though it fallsshort when used for exploratory data anal-ysis. To address some of these shortcom-ings, we give the viewer explicit controlover the creation of the wordcloud, allow-ing them to interact with it in real time–a dynamic wordcloud. This allows itera-tive adaptation of the visualization to thedata and inference task at hand. We nextpresent a principled approach to visualiza-tion which highlights the similarities anddifferences between two sets of documents– a Venncloud. We make all the visual-ization code (primarily JavaScript) freelyavailable.

1 Introduction

A cornerstone of exploratory data analysis is visu-alization. Tremendous academic effort and engi-neering expertise has created and refined a myriadof visualizations available to the data explorer, yetthere still exists a paucity of options for visualizinglanguage data. While visualizing human languageis a broad subject, we apply Polya’s dictum, andexamine a pair of simpler questions for which westill lack an answer:

• (1) what is in this corpus of documents?

• (2) what is the relationship between thesetwo corpora of documents?

We assert that addressing these two questions isa step towards creating visualizations of humanlanguage more suitable for exploratory data anal-ysis. In order to create a meaningful visualiza-tion, one must understand the inference questionthe visualization is meant to inform (i.e., the rea-son for which (1) is being asked), so the appro-priate aspects of the data can be highlighted with

the aesthetics of the visualization. Different infer-ence questions require different aspects to be high-lighted, so we aim to create a maximally-flexible,yet simple and intuitive method to enable a userto explore the relevant aspects of their data, andadapt the visualization to their task at hand.

The primary contributions of this paper are:

• A visualization of language data tailored forexploratory data analysis, designed to exam-ine a single corpus (the dynamic wordcloud)and to compare two corpora (the Venncloud);

• The framing and analysis of the problem interms of the existing psychophysical litera-ture;

• Distributable JavaScript code, designed to besimple to use, adapt, and extend.

We base our visualizations on the wordcloud,which we deconstruct and analyze in §3 and §4.We then discuss the literature on wordclouds andrelevant psychophysical findings in §5, takingguidance from the practical and theoretical foun-dations explored there. We then draw heavily onsimilarities to more common and well understoodvisualizations to create a more useful version ofthe wordcloud. Question (1) is addressed in §7,and with only a small further expansion describedin §8, an approach to (2) becomes evident.

2 Motivating Inference Tasks

Exploratory data analysis on human language en-compasses a diverse set of language and infer-ence tasks, so we select the following subset fortheir variety. One task in line with question (1)is getting the general subject of a corpus, high-lighting content-bearing words. One might wantto examine a collection of social media missives,too numerous to read individually, perhaps to de-tect emerging news (Petrovic et al., 2013). Sepa-rately, author identification (or idiolect analysis)

22

attempts attribution of documents (e.g., Shake-speare’s plays or the Federalist papers) by com-paring the author’s writing style, focusing onstylistic and contentless words – for a review see(Juola, 2006). Further, some linguistic psycho-metric analysis depends on the relative distribu-tion of pronouns and other seemingly contentlesswords (Coppersmith et al., 2014a; Chung and Pen-nebaker, 2007).

Each of these questions involves some analy-sis of unigram statistics, but exactly what analy-sis differs significantly, thus no single wordcloudcan display all of them. Any static wordcloud isa single point in a distribution of possible word-clouds – one way of calculating statistics from theunderlying language and mapping those calcula-tions to the visual representation. Many such com-binations and mappings are available, and the opti-mal wordcloud, like the optimal plot, is a functionof the data and the inference task at hand. Thus,we enable the wordcloud viewer to adjust the rela-tionship between the aspects of the data and theaesthetics of the display, which allows them toview different points in the distribution of possi-ble wordclouds. The dynamic wordcloud was im-plicitly called for in (Rayson and Garside, 2000)since human expertise (specifically knowledge ofbroader contexts and common sense) is neededto separate meaningful and non-meaningful differ-ences in wordclouds. We enable this dynamic in-teraction between human and visualization in real-time with a simple user interface, requiring only amodicum more engineering than the creation of astatic wordcloud, though the depth of extra infor-mation conveyed is significant.

3 Wordcloud Aesthetics

We refer to each visual component of the visual-ization as an aesthetic (ala (Wickham, 2009)) –each aesthetic can convey some information to theviewer. For context, the aesthetics of a scatterplotinclude the x and y position, color, and size ofeach point. Some are best suited for ordinal data(e.g., font size), while others for categorical data(e.g., font color).

Ordinal data can be encoded by font size,the most prominent and noticeable to the viewer(Bateman et al., 2008). Likewise, the opacity(transparency) of the word is a prominent and or-dinal aesthetic. The order in which words are dis-played can convey a significant amount of infor-

mation as well, but using order in this fashion gen-erally constrains the use of x and y position.

Categorical data can be encoded by the colorof each word – both the foreground of the worditself and the background space that surrounds it(though that bandwidth is severely limited by hu-man perception). Likewise for font weight (bold-ness) and font decoration (italics and underlines).While font face itself could encode a categoricalvariable, making comparisons of all the other as-pects across font faces is likely to be at best unin-formative and at worst misleading.

4 Data Aspects

As the wordcloud has visual aesthetics that we cancontrol (§3), the data we need to model has aspectsthat we want to represent with those aesthetics.This aspect-to-aesthetic mapping is what makes auseful and informative visualization, and needs tobe flexible enough to allow it be used for a rangeof inference tasks.

For clarity, we define a word (w) as a uniqueset of characters and a word token (w) as a sin-gle usage of a word in a document. We can ob-serve multiple word tokens (w) of the same word(w) in a single document (d). For any document dwe represent the term frequency of w as tf

d

(w).Similarly, the inverse document frequency of was idf(w). A combination of tf and idf is oftenused to determine important words in a documentor corpus. We focus on tf and idf here, but this isjust an example of an ordinal value associated witha word, there are many other such word-ordinalpairings that are worth exploring (e.g., weights ina classifier).

The dynamic range (“scaling” in (Wickham,2009)) also needs to be considered, since the datahas a natural dynamic range – where meaningfuldifferences can be observed (unsurprisingly, thedefinition of meaningful depends on the inferencetask). Likewise, each aesthetic has a range of val-ues for which the users can perceive and differen-tiate (e.g., words in a font size too small are illeg-ible, those too large prevent other words from be-ing displayed; not all differences are perceptible).Mapping the relevant dynamic range of the datato the dynamic range of the visualization is at theheart of a good visualization, but to do this algo-rithmically for all possible inference tasks remainsa challenge. We, instead, enable the user to adjustthe dynamic range of the visualization explicitly.

23

5 Prior Art

Wordclouds have a mixed history, stemming fromJim Flanagan’s “Search Referral Zeitgeist”, usedto display aggregate information about websiteslinking to his, to its adoption as a visual gim-mick, to the paradoxical claim that ‘wordcloudswork in practice, but not in theory’ (see (Viegasand Wattenberg, 2008) for more). A numberof wordcloud-generators exist on the web (e.g.,(Feinberg, 2013; Davies, 2013)), though thesetend towards creating art rather than informativevisualizations. The two cited do allow the userlimited interaction with some of the visual aesthet-ics, though not of sufficient scope or response timefor general exploratory data analysis.

Enumerating all possible inference tasks involv-ing the visualization of natural language is impos-sible, but the prior art does provide empirical datafor some relevant tasks. This further stresses theimportance of allowing the user to interact withthe visualization, since optimizing the visualiza-tion a priori for all inference tasks simultaneouslyis not possible, much like creating a single plot forall numerical inference tasks is not possible.

5.1 Psychophysical AnalysesThe quintessential studies on how a wordcloudis interpreted by humans can be found in (Ri-vadeneira et al., 2007) and (Bateman et al.,2008). They both investigated various measures ofimpression-forming and recall to determine whichaesthetics conveyed information most effectively– font size chief among them.

Rivandeneira et al. (Rivadeneira et al., 2007)also found that word-order was important for im-pression forming (displaying words from most fre-quent to least frequent was most effective here),while displaying words alphabetically was bestwhen searching for a known word. They alsofound that users prefer a search box when search-ing for something specific and known, and a word-cloud for exploratory tasks and things unknown.

Bateman et al. (Bateman et al., 2008) examinedthe relative utility of other aesthetics to convey in-formation, finding that font-weight (boldness) andintensity (opacity) are effective, but not as goodas font-size. Aesthetics such as color, number ofcharacters or the area covered by the word wereless effective.

Significant research has gone in to the place-ment of words in the wordcloud (e.g., (Seifert et

al., 2008)), though seemingly little informationcan be conveyed by these layouts (Schrammel etal., 2009). Indeed, (Rivadeneira et al., 2007) in-dicates that words directly adjacent to the largestword in the wordcloud had slightly worse recallthan those not-directly-adjacent – in essence, get-ting the most important words in the center maybe counterproductive. Thus we eschew these algo-rithms in favor of more interpretable (but perhapsless aesthetically pleasing) linear ordered layouts.

5.2 Wordclouds as a tool

Illustrative investigations of the wordcloud as atool for exploratory data analysis are few, but en-couraging.

In relation to question (1), even static word-clouds can be useful for this task. Users per-forming an open-ended search task preferred us-ing a wordcloud to a search box (Sinclair andCardew-Hall, 2008), possibly because the word-cloud prevented them from having to hypothesizewhat might be in the collection before searchingfor it. Similarly, wordclouds can be used as afollow-up display of search results from a queryperformed via a standard text search box (Knautzet al., 2010), providing the user a crude summaryof the results. In both of these cases, a simplestatic wordcloud is able to provide some usefulinformation to the user, though less research hasbeen done to determine the optimal compositionof the wordcloud. What’s more, the need for adynamic interactive wordcloud was made explicit(Knautz et al., 2010), given the way the users iter-atively refined their queries and wordclouds.

Question (2) has also been examined. One ap-proach is to make a set of wordclouds with softconstraints that the same word appears in roughlythe same position across multiple clouds to fa-cilitate comparisons (Castella and Sutton, 2013).Each of these clouds in a wordstorm visualizes adifferent collection of documents (e.g., subdivi-sions via metadata of a larger corpus).

Similarly addressing our second question, Par-allel Tag Clouds (Collins et al., 2009) allow thecomparison of multiple sets of documents (or dif-ferent partitions of a corpus). This investigationprovides a theoretically-justified approach to find-ing ‘the right’ static wordcloud (for a single in-ference task), though this does depend on somelanguage-specific resources (e.g., stopword listsand stemming). Interestingly, they opt for ex-

24

plicit removal of words and outliers that the userdoes not wish to have displayed (an exclusionlist), rather than adjusting calculations of the en-tire cloud to remove them in a principled and fairmanner.

5.3 Wordclouds and Metadata

Wordclouds have previously been extended toconvey additional information, though these adap-tations have been optimized generally for artisticpurposes rather than exploratory data analysis.

Wordclouds can been used to display how lan-guage interacts with a temporal dimension in (Du-binko et al., 2007; Cui et al., 2010; Lee et al.,2010). Dubinko and colleagues created a tag cloudvariant that displays trends in tag usage over time,coupled with images that have that tag (Dubinko etal., 2007). An information-theoretic approach todisplaying information changing in time gives riseto a theoretically grounded approach for display-ing pointwise tag clouds, and highlighting thosepieces that have changed significantly as com-pared to a previous time period (Cui et al., 2010).This can be viewed as measuring the change inoverall language usage over time. In contrast, us-ing spark lines on each individual word or tag canconvey temporal trends for individual words (Leeet al., 2010).

Meanwhile, combining tag clouds with geospa-tial data yields a visualization where words can bedisplayed on a map of the world in locations theyare frequently tagged in, labeling famous land-marks, for example (Slingsby et al., 2007).

6 Desiderata

In light of the diverse inference tasks (§2) andprior art (§5), the following desiderata emerge forthe visualization. These desiderata are explicitchoices, not all of which are ideal for all infer-ence tasks. Thus, chief among them is the first:flexibility to allow maximum extensions and mod-ifications as needed.

Flexible and adjustable in real time: Any sin-gle static wordcloud is guaranteed to be subopti-mal for at least some inference tasks, so allowingthe user to adjust the aspect-to-aesthetic mappingof the wordcloud in real time enables adaptationof the visualization to the data and inference taskat hand. The statistics described in §4 are relevantto every language collection (and most inferencetasks), yet there are a number of other ordinal val-

ues to associate a word (e.g., the weight assignedto it by a classifier). Thus, tf and idf are meantto be illustrative examples though the visualizationcode should generalize well to others.

Though removal of the most frequent words(stopwords) is useful in many natural languageprocessing tasks, there are many ways to definewhich words fall under this category. Unsurpris-ingly, the optimal selection of these words can alsodepend upon the task at hand (e.g., psychiatric v.thematic analysis as in §2), so maximum flexibilityand minimum latency are desirable.

Interpretable: An explicit legend is needed tointerpret the differences in visual aesthetics andwhat these differences mean with respect to theunderlying data aspects.

Language-Agnostic: We need methods for ex-ploratory data analysis that work well regard-less of the language(s) being investigated. Thisis crucial for multilingual corpora, yet decidedlynontrivial. These techniques must be maximallylanguage-agnostic, relying on only the most rudi-mentary understanding of the linguistic structureof the data (e.g., spaces separate words in English,but not in Chinese), so they can be extended tomany languages easily.

This precludes the use of a fixed set of stopwords for each language examined, since a new setof stopwords would be required for each languageexplored. Alternatively, the set of stopwords canbe dealt with automatically, either by granting theuser the ability to filter out words in the extremesof the distributions (tf and df alike) through theuse of a weight which penalizes these ubiquitousor too-rare words. Similarly precluded is the useof stemming to deal with the many surface formsof a given root word (e.g., type, typing, typed).

7 Dynamic Wordclouds

We address Question (1) and a number of ourdesiderata with the addition of explicitly labeledcontrols to the static wordcloud display, which al-lows the user to control the mapping from dataaspects to the visualization aesthetics. We sup-plement these controls with an explicit explana-tion of how each aesthetic is affected by eachaspect, so the user can easily read the relevantmappings, rather than trying to interpret the loca-tion of the sliders. An example of which is that“Larger words are those that frequently occur inthe query”, when the aspect tf is mapped to the

25

[X] entities commoncloud

tffilter

idffilter

sizecontrols

opacitycontrols

sortby

wordclouddescription

do notredraw

highlightkeywords

orioles at to for inbaltimore a game iand yankees of goon is this vs w/ o'sare it win with be i'mmy you that others newhave just 2 3 up 1 let's first so alltonight fans but out season we now get soxan time md series al baseball will about me notlike as 5 playoff lets from over was they if what 4 good nogreat year day one love back amp fan magic can east playoffs beat gtred sports rt today do going how last more place team games night blue

file:///Users/gcoppersmith/Desktop/dwc_dev/orioles.html

1 of 2 2/24/14, 4:36 PM

[X] entities commoncloud

tffilter

idffilter

sizecontrols

opacitycontrols

sortby

wordclouddescription

do notredraw

highlightkeywords

13 espn fox leads md ny report starusa buck chris glove mark move press14 gold hamilton jim third birds gonzalez jo september white homers rangers tigers 15 fallyankee sox yard al happen pick blue cc extra m oldcard closer hits tie tied face omg ravens says yahoo citysaying second starting stay ass manager players real lt threebout football left sunday sweep goes hey gets wild yanks adam jobnews times won't magic place innings 2013 base gt hr pitching doesfucking os rain 10 friday outs suck coming makes also check wearinghaha show through god los losing may always being keep 9 hate gotta looksmaking posted away free im life we're weekend call give little bar bring didn'tdoing least look stop thing top where giants long photo record johnson leaguemost thank tv ? please proud walk wish 2012 any end every hell post samebecause many pitch said its new runs everyone excited field finally bullpen let redyes believe fun made these boys hit into those against mlb race i'll teams amazing sayw years another people 1st both postseason wait before 8 looking loss him lead man shitfeel inning damn oh 7 baby i've again should bad then thanks two wins nice much that'swinning yankees than gonna playing tickets yeah awesome beat why division lol wowcongrats lost ready lose us big could never even fuck our want since ball ever happy start ambetter make run hope think only over he were getting has sports really his u work 6 next wonchecked chicago did them some after well best tomorrow watch take their when had too orioles there angames world need who 0 stadium clinch your been down love don't way by east or know off play would 4more o's it's no can't rt playoffs right as still how can back here playoff 1 5 year 3 night watching team do lastone amp got day series home see come great today from like 2 will fan going what pic they about time good lets firstseason if get was tonight we not fans let's all me baseball out now up just so but have be win you that with i'm it are myothers this is washington on vs of w/ go i and

file:///Users/gcoppersmith/Desktop/dwc_dev/orioles.html

1 of 1 2/24/14, 4:35 PM

[X] entities commoncloud

tffilter

idffilter

sizecontrols

opacitycontrols

sortby

wordclouddescription

do notredraw

highlightkeywords

0 1 1st 2 2012 3 4 5 6 7 ? a about after again against al all amp an and are as

at back baltimore baseball be beat been best better

big birds blue buck but by can can't card checked chicago come could

day did do don't down east espn even ever fan fans first for from fuck gamegames get go going gonna good got great gt had has have he here his hit home how i i'mif in inning is it it's johnson just keep know last lead let's lets like lol lose love

magic make mark md me mlb more my need new news next

night no not now o's of off on one only or orioles others our

out over photo pic place play playoff playoffs post postseason rangersravens really red report right rt run season see series should since so some

sox sports start still take team than that their them there they think this tied time

times to today tomorrow tonight two up vs w/ was watch watching way we well were

what when white who why wild will win with won world would yahooyankees yanks year years you your

file:///Users/gcoppersmith/Desktop/dwc_dev/orioles.html

1 of 1 2/24/14, 4:37 PMFigure 1: Three example settings of the dynamic wordcloud for the same set of tweets containing “Orioles”. Left: sizereflects tf , sorted by tf ; Center: size reflects idf , sorted by idf ; Right: size reflects tf*idf , sorted alphabetically.

aesthetic font-size (and this description is tied tothe appropriate sliders so it updates as the slid-ers are changed). The manipulation of the visu-alization in real time allows us to take advantageof the human’s adept visual change-detection tohighlight and convey the differences between set-tings (or a range of settings), even subtle ones.

The data aspects from §4 are precomputed andmapped to the aesthetics from §3 in a JavaScriptvisualization displayed in a standard web browser.This visualization enables the user to manipulatethe aspect-to-aesthetic mapping via an intuitiveset of sliders and buttons, responsive in real time.The sliders are roughly segmented into three cat-egories: those that control which words are dis-played, those that control how size is calculated,and those that control how opacity is calculated.The buttons control the order in which words ap-pear.

One set of sliders controls which words aredisplayed by examining the frequency and rar-ity of the words. We define the range ⌧

Freq

=

[tmin

Freq

, tmax

Freq

] as the range of tf values for words tobe displayed (i.e., tf(w) 2 ⌧

Freq

). The viewer isgranted a range slider to manipulate both tmin

Freq

andtmax

Freq

to eliminate words from the extremes of thedistribution. Similarly for df and ⌧

Rarity

. Thosewords that fall outside ⌧

Freq

or ⌧Rarity

are not dis-played. Importantly, tf is computed from the cur-rent corpus displayed while df is computed over amuch larger collection (in our running examples,all the works of Shakespeare or all the tweets forthe last 6 months). Those with high df or hightf are often stopwords, those with low tf and lowdf are often rare, sometimes too rare to get goodestimates of tf or idf (e.g., names).

A second set of sliders controls the mapping be-tween aspects and aesthetics for each individualword. Each aesthetic has a weight for the impor-tance of rarity (�

Rarity

) and the importance of fre-quency (�

Freq

), corresponding to the current val-ues of their respective slider (each in the range[0, 1]). For size, we compute a weight attributedto each data aspect:

!Freq

(w) = (1� �Freq

) + �Freq

tf(w)

and similarly for Rarity.In both cases, the aesthetic’s value is computed

via an equation similar to the following:

a(w) = !Freq

(w)!Rarity

(w)�Range

b

where a(w) is either font size or opacity, and bis some base value of the aesthetic (scaled by adynamic range slider, �

Range

) and the weights forfrequency and rarity of the word. In this manner,the weights are multiplicative, so interactions be-tween the variables (e.g., tf*idf ) are apparent.

Though unigram statistics are informative, see-ing the unigrams in context is also important formany inference tasks. To enable this, we use reser-voir sampling (Vitter, 1985) to maintain a repre-sentative sample of the observed occurrences ofeach word in context, which the user can view byclicking on the word in the wordcloud display.

Examples of the dynamic wordcloud in varioussettings can be found in Figure 1, using a set oftweets containing “Orioles”. The left wordcloudhas tf mapped to size, the center with idf mappedto size, and the right with both high tf and highidf mapped to size. We only manipulate the sizeaesthetic, since the opacity aesthetic is sometimeshard to interpret in print. To fit the wordclouds

26

to the small format, various values for ⌧Freq

and⌧Rarity

are employed, and order is varied – theleft is ordered in descending order in terms of fre-quency, the center is ordered in descending orderin terms of rarity, and the right is in alphabeticalorder.

8 Vennclouds

Question (2) – “how are these corpora related” re-quires only a single change to the dynamic sin-gle wordcloud described in §7. We refer to twocorpora, left and right, which we abbreviate Land R (perhaps a set of tweets containing “Ori-oles” for left and those containing “Nationals” forright as in Figure 2). For the right documents, letR = {d1, ..., dnR} so |R| = n

R

and let TR

be thetotal number of tokens in all the documents in R

TR

=

X

d2R

|Td

|

We separate the wordcloud display into three re-gions, one devoted to words most closely associ-ated with R, one devoted to words most closelyassociated with L, and one for words that shouldbe associated with both. “Association” here can bedefined in a number of ways, but for the nonce wedefine it as the probability of occurrence in thatcorpus – essentially term frequency, normalizedby corpus length. Normalizing by length is re-quired to prevent bias incurred when the corporaare different sizes (T

L

6= TR

). Specifically, wedefine the number of times w occurs in left (tf ) as

tfL

(w) =

X

di2L

T (w, di

)

and this quantity normalized by the number of to-kens in L,

tfL

(w) = tfL

(w)/TL

and this quantity as it relates to the term frequencyof this w in both corpora

tfL|R(w) =

tfL

(w)

tfL

(w) + tfR

(w)

Each word is only displayed once in the Ven-ncloud (see Figure 2, so if a word (w) only occursin R, it is always present in the right region, andlikewise for L and left. If w is in both L and R,we examine the proportion of documents in eachthat w is in and use this to determine in which re-gion it should be displayed. In order to deal with

Figure 2: Three example Vennclouds, with tweets contain-ing “Orioles” on the left, “Nationals” on the right, and com-mon words in the middle. From top to bottom we allow pro-gressively larger common clouds. The large common wordsmake sense – both teams played a Chicago team and madethe playoffs in the time covered by these corpora.

the cases where w occurs in approximately similarproportions of left and right documents, we havea center region (in the center in Figure 2). Wedefine a threshold (⌧

Common

) to concretely define“approximately similar”. Specifically,

• if tfR

(w) = 0, w is displayed in left.

• if tfL

(w) = 0, w is displayed in right.

• if tfR

(w) > 0 and tfL

(w) > 0,

– if tfR|L(w) > tf

L|R(w) + ⌧Common

, wis displayed in right.

– if tfL|R(w) > tf

R|L(w) + ⌧Common

, wis displayed in left.

– Otherwise, w is displayed in center.

The user is given a slider to control ⌧Common

, al-lowing them to determine what value of “approx-imately similar” best fits the data and their task athand.

9 Anecdotal Evaluation

We have not yet done a proper psychophysicalevaluation of the utility of dynamic wordclouds

27

[X][X] entitiesentities common cloudcommon cloud tf filtertf filter idf filteridf filter size controlssize controls opacity controlsopacity controls sort bysort by legendlegend do not redraw do not redraw highlight keywords highlight keywords

@ @nationals am baseball best bryce but

cardinals cards clinchcome cubs dc did

getting go going got here home if

know louis lt marlins me

nationals nats nl not or others

park philadelphia phillies pic

see st stadium strasburg teddy

today too u w/ was washington watching well werth

work

0 2 4 5 6 7 8 9 ? a about

after again all

amp and another any are as awesome baby back

bar be because beenbefore being believe better big boys

bring bullpen by call can can't checked chicago

could damn day division do don't down east end

even ever every everyone excited fan fans feel

field finally first for free

from fuck game games get

give gonna good great had has have he hell him his

hope how i i'll i'm im

in inning into is it it's its just

last league let let's lets life like littlelol long look looking lose losing loss love made

make man many

more most much my need

next night no now of off oh on one only

our out pitch playplaying playoff playoffs please postseason proud

ready really right rt run runs saidsay season series should since so some

start still take team than

thank thanks that that's their them then there these they thing

think this those time to tomorrow tonight two

up vs wait want watch way we were what when where

who why will win winning wish with won worldwould yeah year yes you your

1 1st 2012 3 @orioles against al an atbaltimore beat birds

blue buck card espn

gt hit

johnson keep lead magic mark md

mlb new news o's

orioles over photo place

post rangers ravensred report

sox sports tied

white wild yahoo yankees yanks years

Words displayed occur at least 22 and at most 6590 times in the query.Words displayed occur in fewer than 5 and more than 10000 documents in the whole corpus.Larger words frequently occur in the query and rarely occur in the corpus (TF*IDF). [TF:0.92,IDF:1]Darker words are more frequent in the query (TF). [TF:1]Words are sorted alphabetically.

[X]go o's @ oriole park at camden yards for boston redsox vs baltimore orioles w/ 163 others

i'm at oriole park at camden yards for chicago whitesox vs baltimore orioles baltimore md

#orioles 5 things the boston red sox can learn from thebaltimore orioles oakland a's bleacher#sportsroadhouse

[X]nats white sox south carolina 100 to win 450

#orioles welcome to red sox nation orioles' andnationals' fans #sportsroadhouse

god damn nats just tried to boo in a dc bar almost gotchased out since my sox are out i just want the cardsto do well #problems #mlb

file:///Users/gcoppersmith/src/dwc_dev/nats_v_orioles.html

1 of 1 4/24/14, 9:53 AM

Figure 3: Screenshot of a Venncloud, with controls. The sliders are accessible from the buttons across the top, displaying asa floating window above the wordcloud itself (replacing the current display of the legend). Also note the examples in the lowerleft and right corners, accessed by clicking on a word of interest (in this case “Sox”).

and Vennclouds for various tasks as compared totheir static counterparts (and other visualizations).In part, this is because such an evaluation requiresselection of inference tasks to be examined, pre-cisely what we do not claim to be able to do. Weleave for future work the creation and evaluationof a representative sample of such inference tasks.

We strongly believe that the plural of anecdoteis not data – so these anecdotes are intended asillustrations of use, rather than some data regard-ing utility. The dynamic wordclouds and Ven-nclouds were used on data from across the spec-trum, from tweets to Shakespeare and politicalspeeches to health-related conversations in devel-oping nations. In Shakespeare, character and placenames can easily be highlighted with one set ofslider settings (high tf*idf ), while comparisonsof stopwords are made apparent with another (hightf , no idf ). Emerging from the debates betweenMitt Romney and Barack Obama are the commonthemes that they discuss using similar (economics)and dissimilar language (Obama talks about the“affordable care act” and Romney calls it “Oba-macare”). These wordclouds were also used to dosome introspection on the output of classifiers insentiment analysis (Mitchell et al., 2013) and men-tal health research (Coppersmith et al., 2014b) toexpose the linguistic signals that give rise to suc-cessful (and unsuccessful) classification.

10 Conclusions and Future Directions

Exploratory data analysis tools for human lan-guage data and inference tasks have long laggedbehind their numerical counterparts, and here we

investigate another step towards filling that need.Rather than determining the optimal wordcloud,we enable the wordcloud viewer to adapt the visu-alization to the data and inference task at hand. Wesuspect that the pendulum of control has swungtoo far, and that there is a subset of the possi-ble control configurations that produce useful andinformative wordclouds. Work is underway tocollect feedback via instrumented dynamic word-clouds and Vennclouds as they are used for variousinference tasks to address this.

Previous research, logic, and intuition wereused to create this step, though it requires fur-ther improvement and validation. We provideanecdotes about the usefulness of these dynamicwordclouds, but those anecdotes do not providesufficient evidence that this method is somehowmore efficient (in terms of human time) than ex-isting methods. To make such claims, a controlledhuman-factors study is required, investigating (fora particular inference task) how this method af-fects the job of an exploratory data analyst. Inthe meantime, we hope making the code freelyavailable1 will better enable our fellow researchersto perform principled exploratory data analysis ofhuman language content quickly and encourage adeeper understanding of data, within and acrossdisciplines.

Acknowledgments

We would like to thank Carey Priebe for in-sightful discussions on exploratory data analysis,

1from https://github.com/Coppersmith/vennclouds

28

Aleksander Yelskiy, Jacqueline Aguilar, KristyHollingshead for their analysis, comments, andimprovements on early versions, and AinsleyR. Coppersmith for permitting this research toprogress in her early months.

ReferencesScott Bateman, Carl Gutwin, and Miguel Nacenta.

2008. Seeing things in the clouds: the effect of vi-sual features on tag cloud selections. In Proceedingsof the nineteenth ACM conference on Hypertext andhypermedia, pages 193–202. ACM.

Quim Castella and Charles A. Sutton. 2013. Wordstorms: Multiples of word clouds for visual compar-ison of documents. CoRR, abs/1301.0503.

Cindy Chung and James W Pennebaker. 2007. Thepsychological functions of function words. Socialcommunication, pages 343–359.

Christopher Collins, Fernanda B Viegas, and MartinWattenberg. 2009. Parallel tag clouds to exploreand analyze faceted text corpora. In Visual Analyt-ics Science and Technology, 2009. VAST 2009. IEEESymposium on, pages 91–98. IEEE.

Glen Coppersmith, Mark Dredze, and Craig Harman.2014a. Quantifying mental health signals in twitter.In Proceedings of ACL Workshop on ComputationalLinguistics and Clinical Psychology. Association forComputational Linguistics.

Glen Coppersmith, Craig Harman, and Mark Dredze.2014b. Measuring post traumatic stress disorder inTwitter. In Proceedings of the International AAAIConference on Weblogs and Social Media (ICWSM).

Weiwei Cui, Yingcai Wu, Shixia Liu, Furu Wei,Michelle X Zhou, and Huamin Qu. 2010. Contextpreserving dynamic word cloud visualization. InPacific Visualization Symposium (PacificVis), 2010IEEE, pages 121–128. IEEE.

Jason Davies. 2013. Wordcloud generator using d3,April.

Micah Dubinko, Ravi Kumar, Joseph Magnani, Jas-mine Novak, Prabhakar Raghavan, and AndrewTomkins. 2007. Visualizing tags over time. ACMTransactions on the Web (TWEB), 1(2):7.

Jason Feinberg. 2013. Wordle, April.

Patrick Juola. 2006. Authorship attribution. Founda-tions and Trends in information Retrieval, 1(3):233–334.

Kathrin Knautz, Simone Soubusta, and Wolfgang GStock. 2010. Tag clusters as information retrievalinterfaces. In System Sciences (HICSS), 2010 43rdHawaii International Conference on, pages 1–10.IEEE.

Bongshin Lee, Nathalie Henry Riche, Amy K Karl-son, and Sheelagh Carpendale. 2010. Spark-clouds: Visualizing trends in tag clouds. Visualiza-tion and Computer Graphics, IEEE Transactions on,16(6):1182–1189.

Margaret Mitchell, Jacqueline Aguilar, Theresa Wil-son, and Benjamin Van Durme. 2013. Open domaintargeted sentiment. In Proceedings of the 2013 Con-ference on Empirical Methods in Natural LanguageProcessing, pages 1643–1654. Association for Com-putational Linguistics.

Sasa Petrovic, Miles Osborne, Richard McCreadie,Craig Macdonald, Iadh Ounis, and Luke Shrimpton.2013. Can twitter replace newswire for breakingnews. In Seventh International AAAI Conference onWeblogs and Social Media.

Paul Rayson and Roger Garside. 2000. Comparingcorpora using frequency profiling. In Proceedingsof the workshop on Comparing Corpora, pages 1–6.Association for Computational Linguistics.

AW Rivadeneira, Daniel M Gruen, Michael J Muller,and David R Millen. 2007. Getting our head in theclouds: toward evaluation studies of tagclouds. InProceedings of the SIGCHI conference on Humanfactors in computing systems, pages 995–998. ACM.

Johann Schrammel, Michael Leitner, and ManfredTscheligi. 2009. Semantically structured tagclouds: an empirical evaluation of clustered presen-tation approaches. In Proceedings of the 27th inter-national conference on Human factors in computingsystems, pages 2037–2040. ACM.

Christin Seifert, Barbara Kump, Wolfgang Kienreich,Gisela Granitzer, and Michael Granitzer. 2008. Onthe beauty and usability of tag clouds. In Informa-tion Visualisation, 2008. IV’08. 12th InternationalConference, pages 17–25. IEEE.

James Sinclair and Michael Cardew-Hall. 2008. Thefolksonomy tag cloud: when is it useful? Journal ofInformation Science, 34(1):15–29.

Aidan Slingsby, Jason Dykes, Jo Wood, and KeithClarke. 2007. Interactive tag maps and tagclouds for the multiscale exploration of large spatio-temporal datasets. In Information Visualization,2007. IV’07. 11th International Conference, pages497–504. IEEE.

Fernanda B Viegas and Martin Wattenberg. 2008.Timelines tag clouds and the case for vernacular vi-sualization. interactions, 15(4):49–52.

Jeffrey S Vitter. 1985. Random sampling with a reser-voir. ACM Transactions on Mathematical Software(TOMS), 11(1):37–57.

Hadley Wickham. 2009. ggplot2: elegant graphics fordata analysis. Springer Publishing Company, Incor-porated.

29