Observational and Experimental Investigation of Typing ... using Virtual Keyboards on Mobile Devices ... virtual keyboard; ... on virtual keyboards that are operated by touching the screen.
Observational and Experimental Investigation of TypingBehaviour using Virtual Keyboards on Mobile DevicesNiels HenzeUniversity of OldenburgOldenburg, Germanyniels.email@example.comEnrico RukzioUniversity of Duisburg-Essen /Lancaster Universityenrico.firstname.lastname@example.orgSusanne BollUniversity of OldenburgOldenburg, Germanysusanne.email@example.comABSTRACTWith the rise of current smartphones, virtual keyboards fortouchscreens became the dominant mobile text entry tech-nique. We developed a typing game that records how userstouch on the standard Android keyboard to investigate userstyping behaviour. 47,770,625 keystrokes from 72,945 instal-lations have been collected by publishing the game. By visu-alizing the touch distribution we identified a systematic skewand derived a function that compensates this skew by shift-ing touch events. By updating the game we conduct an ex-periment that investigates the effect of shifting touch events,changing the keys labels, and visualizing the touched posi-tion. Results based on 6,603,659 keystrokes and 13,013 in-stallations show that visualizing the touched positions usinga simple dot decreases the error rate of the Android keyboardby 18.3% but also decreases the speed by 5.2% with no posi-tive effect on learnability. The Android keyboard outperformsthe control condition but the constructed shift function furtherimproves the performance by 2.2% and decreases the errorrate by 9.1%. We argue that the shift function can improveexisting keyboards at no costs.Author Keywordstouchscreen; virtual keyboard; mobile phone; public study.ACM Classification KeywordsH.5.2 Interfaces and Presentation: User Interfaces - Input de-vices and strategies.General TermsDesign, Human Factors, Experimentation.INTRODUCTIONSince the introduction of the iPhone, mobile phones withtouchscreens began to dominate the smartphone market. To-day, all major phone makers have touchscreen devices in theirportfolio. In contrast to earlier devices, todays smartphonesare operated by touching the screen with the fingers and onlya few devices have a physical keyboard. Instead, users relyon virtual keyboards that are operated by touching the screen.Permission to make digital or hard copies of all or part of this work forpersonal or classroom use is granted without fee provided that copies arenot made or distributed for profit or commercial advantage and that copiesbear this notice and the full citation on the first page. To copy otherwise, orrepublish, to post on servers or to redistribute to lists, requires prior specificpermission and/or a fee.CHI12, May 510, 2012, Austin, Texas, USA.Copyright 2012 ACM 978-1-4503-1015-4/12/05...$10.00.While touchscreens and virtual keyboards have been studiedfor years, understanding users touch behaviour remains chal-lenging. Previous work usually studies the effect of singleaspects, such as key size or keyboard layout, on the usersperformance. Due to limited resources corresponding userstudies are often conducted with a homogenous sample and asingle device. Such studies usually try to seek a balance be-tween internal validity (the extent to which variance is due tothe test conditions) and external validity (the extent to whichresults are generalizable). Experimenters control most vari-ables and conduct studies with a small number of participantsin a lab (high internal and low external validity). Many resultsfrom related work are therefore based on the performance ofmale right-handed students from a technical discipline thatlive in the same region i.e. no equal gender split and mainlyparticipants from the authors institution.In contrast to previous work, our aim is to observe and manip-ulate the touch behaviour of a diverse sample, a large numberof devices, and various contexts. To collect the required largeamount of keystrokes on a virtual keyboard we developed amobile typing game. To attract a large number of participantsthe game has been published to the Android Market. Ourapproach thus allows studying a large number of users withvarying backgrounds in a large number of realistic contextswith their own devices (low internal validity due to a highvariance but high external validity). This allowed analysingthe typing performance of users whose behaviour would havebeen significantly altered in a very controlled setting. Be-cause external factors cannot be ruled out and we have littlecontrol over the participants the study has a low internal va-lidity as there was no possibility to control any contextualfactors. The flip side is that the diversity of the environmentprovides a higher external validity than common lab studies.After discussing related work, we describe the game that wedeveloped to collect the data. We provide an overview aboutthe data we collected after publishing the game to the AndroidMarket. Following this, an analysis of the touch distributionis provided that shows how touch contacts are skewed relativeto the keyboards keys centre. Afterwards, three approachesto influence the users typing behaviour are proposed. Wereport how we evaluated these approaches in an experimentby publishing an update of the game to the Android Mar-ket. We show that our adapted shift function improves theperformance, elevating the position of the keys labels is notbeneficial, and informing the user about the touched positiondecreases the error rate but also decreases the speed. We closethe paper with a conclusion and an outlook on future work.RELATED WORKIn the last decade and in particular since the emergence of theiPhone we observed a shift from using keypads and stylus-based interaction to finger-based interactions with touch-screens on mobile devices. One important aspect to be con-sidered in the design of virtual keyboards is that the outputresolution of such a touchscreen is much higher than the in-put resolution of a human thumb or finger. This leads tothe fat-finger-problem due to the difficulty to select smalltargets with a much larger finger and the aspect that the fin-ger occludes the target as well. Current smartphones addressthis aspect e.g. through a visual confirmation of what hasbeen touched or trough callouts that show the region currentlytouched in order to perform fine granular selections.One strand of research focuses on interaction techniques thatallow the selection of small targets with a finger withoutchanging the size of the target while achieving an acceptableerror rate. In Shift  this has been achieved through call-outs showing a copy of the area occluded by the finger in anon-occluded area and the possibility to move a pointer in thecallout via finger movement to select the desired target. InTapTap  the occluded area is also shown in a callout buthere a zoomed in copy of the occluded area is shown and theuser has to touch the desired target in the callout with a secondtouch. In Escape  the small targets are visually changedand indicate a direction in which the user has to drag its fin-ger after touching it in order to select it. Those interactiontechniques are not well suited for text input as additional in-teractions are required, which requires more time and a highermental effort, when compared with a simple touch.Further research focused on the optimal size of targets whileconsidering the trade-off between finger size and user inter-faces design. For almost perfect accuracy targets need to belarger than 20 x 20 mm . This means that current touch-screen phones would be able to display only around 8 targetswhile showing no other information. According to the iOSHuman Interface Guidelines  the optimal size of a tapableUI element on the iPhone is 6.74 x 6.74 mm which is a com-promise between an acceptable error rate and the availablescreen size. A significant body of research investigates theinfluence of target size and context on time needed for select-ing a target and the error rate [12, 19]. Considered contextualaspects were e.g. the actual task (e.g. inspired by Fitts lawor text input), device- and display-size and -type, thumb size, activity , touch feedback  or one-handed or two-handed interaction. The outcome is often a suggestion regard-ing an optimal target size and location under consideration ofthe given context and an assumption regarding acceptable er-ror rate, task load or user satisfaction.Relatively little research analysed how the actual location of atarget on the screen or a devices orientation affects effective-ness and efficiency. Early research, focusing on fixed touch-screens mounted on a table, showed that users touch slightlybelow the actual target if the screen is tilted away from theuser and that they touch above the target if its tilted towardsthe user . Other research showed that the location of tar-gets on the screen has an effect on effectiveness, efficiencyand user satisfaction. Himberg et al. developed an adaptivenumerical on-screen keyboard that observes where the user istouching the display in relationship to the displayed key .This information is used to adapt the shape of the virtual keysto improve the error rate. Similar work by  uses geometricpattern matching to reduce the error rate for stylus-based textentry.  developed an anchored keyboard adaptation and asimulation suggests that it reduces the errors rate. Holz andBaudisch investigated how crosshairs are targeted and presenta model that can reduce the error offset .Karlson showed that regions which are easily to reach withthe thumb when considering one-handed interaction achievethe best task performance and lowest perceived difficulty .Karlson concludes that frequently used buttons should beplaced in those regions. Perry and Hourcade showed againthat targets within easy reach of the thumb can be reachedquicker but the accuracy is best when the targets are lo-cated on the left, right and top edges of the screen .Park et al. analysed the success rate, error rate and conve-nience of 25 regions of a touchscreen when using one-handedthumb input . The authors also analysed the offset be-tween indicated target and actual touch events. They ob-served location-specific offsets and discuss the idea of ad-justing the location of the touch recognition area to improvethe overall performance. Those findings have been extendedby Henze et al. who analysed those offsets using a verylarge data set and showed that a corresponding compensa-tion function can reduce the error rate significantly . Re-cent work of Ruchenko et al. tried to improve the perfor-mance of virtual keyboards through data collected in a game. They showed the positive effects of providing feed-back about where users touched so user could adapt theirbehaviour. Unfortunately, only 6 persons participated in thelaboratory study and the potential advantages of key-targetresizing were only shown in a simulation and were not tested.Our paper is the first that analyse the offset between the dis-played keyboard keys and the actual hit locations based on avery large data set collected in a realistic context. This allowsus, in contrast to previous research that was performed in lab-oratory settings or which is based on a small number of touchevents, to calculate those offset vectors very precisely. Fur-thermore are we the first to show that the application of corre-sponding compensation functions improves performance andreduce error rate significantly for a very large number of userswho typed in various contexts using various devices.DESIGN OF THE GAMETo collect a large number of keystrokes on a virtual keyboardfrom a number of different devices and diverse participantswe decided to collect data using a mobile typing game. Dur-ing the design of Type It! we had to find a balance betweenproviding players with a game that is worth playing and a testapplication that collects meaningful data.Game playThe game play focuses on collecting basic keystrokes thatform independent words. Words are presented to the playerand the task is to type these words. The game is structured inFigure 1. Screenshots of the games three stages: stars, water, and fire.three stages called stars, water, and fire. Each stage containsfour levels and each level consists of multiple words that mustbe typed. As shown in Figure 1 the keyboard is displayed inthe lower half of the screen and the words are shown in theupper part of the screen. While playing, words are presentedin white circles with a fixed size. A circular progress bararound the circles shows the remaining time until the wordmust have been typed. The bar is coloured from red to greento also highlight the remaining time. While the time to typea word expires, the progress bar gets shorter. The availabletime to type a word depends on the level and the number ofcharacters. Depending on the level, multiple words are pre-sented simultaneously and can be typed in any order.A words characters must be typed to complete it. While typ-ing, the characters appear in a textbox just above the key-board. The player must confirm the words by either tappingthe space bar or the enter key. If a word has been typed cor-rectly the words background becomes green, the progress baraccelerates, and a rattle sound is played. If the progress bargets empty the word disappears. To make a game out of thebasic task the player must complete a word in a certain time-frame. The timeframe is reduced from word to word whilethe player proceeds through a level and also depends on thewords number of characters. Players receive a penalty pointif a word has not been completed in the given timeframe. Thegame is lost when the player collected three penalty points inone level. Players receive scores when they complete a word.The faster a word is typed the higher the score.To increase the studys internal validity, the same keyboardis used for all devices. We used the source code of the stan-dard Android 2.2 (Froyo) keyboard as basis. The Androidkeyboard is designed to scale across different devices, screensizes, and resolutions. We adapted the keyboard by removingkeys that are not required to play the game and added code tomeasure the players typing behaviour. An interesting aspectof the Android keyboard is that the position of touch eventsis internally shifted upwards by 10 density-independent pix-els (dp). Dp is an abstract unit based on the physical densityof the screen. These units are relative to a 160 dots per inchscreen and designed so that 160dp is one inch. According tothe Android Developer Guide the ratio of dp-to-pixel changeswith the screen density, but not necessarily in direct propor-level max. characters source1 4 MacKenzie et al.2 5 MacKenzie et al.3 1 one character4 7 MacKenzie et al.5 1 one character6 2 two characters7 3 three characters8 2 two random charactersTable 1. The text sources used for the first eight levels of the game.tion1. In addition, if there is free space to the left (as for thea) or to right of a key (as for the l) this area is part of thekeys interactive region. Touching, for example, on the freespace left of the a is still considered as typing on the a.We made the game visually appealing to motivate intensiveusage. Each stage has a different animated background shownin Figure 1. The total score is shown above the keyboard nextto the text box. Furthermore a player receives badges whensuccessfully completing a level or achieving other goals. Toincrease the long term motivation we implemented a globaland a local high score lists shown in Figure 2. Players canshare their score via twitter if they achieve a high score.Levels and text sourcesTo increase the external validity and the players fun weuse words with different length and from different sourcesthroughout the levels. Table 1 provides an overview about thefirst eight levels and the used words. For most levels we ran-domly select words from the phrase set provided by  witha fixed maximum number of characters. In addition, for somelevels we use all words with two or three characters from theOfficial Tournament and Club Word List for Scrabble2 andwords consisting of one or two random characters. We alsovary the available time to complete the words and the numberof words. In general, the game gets more challenging fromlevel to level. While the first levels are very easy we assumethat the very last level is impossible to finish successfully.Measures and consentWe collect various data about the used devices and the per-formance of the players. An identifier for each installation isderived from a devices Android ID to anonymize the data.Furthermore, we collect the users locale (e.g. en GB ores ES), the devices name (e.g. GT-I9000 for the Sam-sung Galaxy S), the time zone as well as the width and theheight of the virtual keyboard in pixels. During the game werecord the presented words and the position where the playertaps the virtual keyboard. We record the position where theplayers finger initially hit the screen and the position wherethe finger lifts-off the screen. We do not record intermediatemovement as this would have led to very large data sets tobe logged and transmitted to our server later on. For the tapswe record the position before the Android keyboard shifts the1Android Developer Guide - More Resource Types:http://developer.android.com/guide/topics/resources/more-resources.html2Scrabble tournament and club word list: http://www.isc.ro/en/commands/lists.htmlhttp://developer.android.com/guide/topics/resources/more-resources.htmlhttp://developer.android.com/guide/topics/resources/more-resources.htmlhttp://www.isc.ro/en/commands/lists.htmlhttp://www.isc.ro/en/commands/lists.htmlFigure 2. Type It! in the Android market (left), a modal dialog that in-forms the player about the study (centre), and the high score list (right).touch events by the 10dp mentioned above. For all events welog the time elapsed since the start of the level.The properties of the used device are transmitted to our serverwhen a game is started and the data collected while playing istransmitted after a level is finished. The data is stored inter-nally on the phone and retransmitted after the next level if thetransmission failed. We do not store data that allows iden-tifying individual players or installations. We inform play-ers about the fact that data is collected to act ethically and toconform to corresponding legislation in many countries. Themodal dialog shown in Figure 2 tells players that they areabout to participate in a study when the game is started for thefirst time. In addition, the description in the Android Marketbriefly outlines our intention, what we record, and what weare trying to achieve with the collected data.PUBLISHING IN THE MARKETWe published Type It!3 in the Android Market on April 29,2011. Figure 2 shows the appearance of the game in theAndroid Market. Till July 31, 2011 the game got installed89,262 times according to Googles Developer Console. Intotal the game received 880 ratings with an average of 3.98on the five point scale (the higher the better). On our serverwe collected data from 80,424 installations but only 72,945installations provided meaningful data (see below). We pro-vide an overview about the data in the following.DemographicsWe collected data from devices with 581 different names.Most of these names appeared, however, only a few timesand are rather exotic such as the Pulse Mini MG Mod. For104 names we collected data from more than a hundred in-stallations. As the mobile network operators give differentnames to the same device type there are in fact much less dif-ferent devices than the 581 names suggest. The Galaxy Tab,for example, appears with at least six different names. Af-ter harmonizing the names for common devices, the 15 mostcommon devices represent 44.13% of all installations.The collected locales and time zones show that there is a biastowards western countries among the players. The most com-mon locales are English speaking US (en US, 65.94%) and3Type It! in the Android Market: http://tiny.cc/type_itFigure 3. Number of collected keystrokes. The graph must be inter-preted as on y installations more than x keystrokes have been recorded.E.g. 40.000 installations contributed more than 200 keystrokes.English speaking Great Britain (en GB, 10.44%). This isfollowed by Germany (de DE, 1.84%), Spanish speakingUS (es US, 1.68%), and France (fr FR, 1.37%). Theother 191 locales together result in 18.72% including 64 fur-ther English locales representing 4.77% of all installations.The recorded time zones show a similar picture. The onlynon US American or European time zone among the ten mostcommon ones is Asia/Calcutta.Collected dataWhile we received data from 80,424 installations, not allof them provided meaningful data. We only use data from72,945 installations because installations provided inconsis-tent data or we did not record a single played level. In to-tal 952,487 levels have been played and on average 13.06levels (SD=58.88) have been played on each installation.On 45.94% of the installations less than 5 levels have beenplayed. There are, however, a few very intensive players and21 installations, for example, contributed more than 1,000levels each. The number of keystrokes per installation is ana-logue. In total 47,770,625 keystrokes have been recorded andon average 654.89 keystrokes (SD=4,149.46) have been pro-duced on each installation. Figure 3 provides an overviewabout the number of keystrokes per installation.OBSERVED TOUCH BEHAVIOURWe computed the distribution of the positions where the play-ers fingers lift-off the screen for each key using our entiredataset. To compute the distribution, we could either assigna touched position to the key that fits the presented word orto the key that is recognized by the keyboard. As the touchesare normally distributed and the error rates are low we as-sign the position to the key recognized by the keyboard. Fig-ure 4 shows the touch distribution for the two most commondevices: the Optimus One, a device with a 3.2 inch screenby LG, and the Ascend, a device with a 3.5 inch screen byHuawei. In the following we analyse the horizontal and thevertical skew in the distribution of taps.Vertical offsetFor each key, we analysed the distribution of touch events bycomputing the distance between the centre of the touch distri-bution and the centre of the visual area of the respective key.http://tiny.cc/type_itFigure 4. Touch distribution of the Optimus One based on 2,407,164keystrokes (top) and the Ascend based on 4,589,967 keystrokes (bottom).Green regions cover 50% of all taps. Red and green regions combinedcover 80% of the taps. Black dots show the distributions centre andblack ellipses one standard deviation.On average the centre of the touch distribution is 10.60 pix-els (SD=8.79 pixels) below the keys centre for the OptimusOne and 8.02 pixels (SD=8.34 pixels) below the keys centrefor the Ascend. The 95% confidence intervals for the centreof the touch distributions are 0.01 pixels wide. Taking thescreens physical size into account this means that on the Op-timus One players hit on average 2.24 mm below the keyscentre and on the Ascend 1.85 mm below the keys centre.The deviation from the centre varies for the different charac-ter keys but the difference between the two devices is consis-tent for all keys. Using the distance between the keys centresand the centre of the touch distribution for the 26 characterkeys as the sample, an unpaired t-test shows that the distanceis significantly different (pBesides the general limitations of our approach the resultsare mainly limited by the fact that comparing different de-vices means conducting a quasi-experiment. Thus, we cannotknow if the observed differences are because of the devicesitself or because of other factors. E.g. some devices might bepreferred by a person group that tends to have smaller handsthan the average (e.g. kids), users with different backgroundmight hold the device differently, and the devices form fac-tor might also have an impact. As our knowledge about theplayers is very limited we cannot factor out those aspects.Furthermore, the analysis is based on data collected using thestandard Android keyboard that shifts the users input. Eventhough players still tap below the keys centre and the ma-jority is certainly not aware of this shift, we cannot know theinfluence of this approach.The analysis of touch behaviour revealed a systematic skewin the distribution of taps. By shifting the users input by10dp towards the top of the screen the Android keyboard al-ready tries to compensate this systematic skew. To our knowl-edge, however, no published work actually showed if thisrather simple compensation function improves the users per-formance. Another potential approach to influence the userstouch behaviour is to change the design of the keys. Assum-ing that the users try to hit the keys labels, shifting the po-sition of the labels towards the upper part of the keys mightinfluence the users to also move their taps upwards. Finally, itseems reasonable to assume that users are not aware that theirtouch distribution is distorted. Visualizing the position wherethe device recognized a touch event, users might be able toadjust their behaviour according to the provided feedback.INFLUENCING TOUCH BEHAVIOURTo analyse the three approaches to influence the users touchbehaviour proposed in the previous section we design accord-ing implementations which are discussed in the following.Shifting touch positionsTo analyse how shifting the users input influences the touchbehaviour we use three different ways to shift the users taps.The first technique, that we call no shift, does not shift thetouch events and simply uses the touch events raw position.As the second technique, that we call native shift we usethe standard Android keyboard that shifts the touch events by10 density independent pixels towards the upper part of thescreen.For the third technique, called adapted shift, we derived acompensation function from the data described in the previ-ous section. The technique follows the assumption that it isbest to shift the users input in a way that moves the centresof the touch distributions to the centres of the keys. Figure 6,exemplarily visualizes the compensation function computedfor the Galaxy S. For each key, we use the centres of the dis-tribution of touch events as support points (highlighted by awhite dot in Figure 6). For each key a vector to the keys cen-tre is derived (shown in blue). Using linear interpolation theshift vectors for the corners of all keys are computed based onthe support points. To avoid shifting taps off the keyboard therespective component of the vectors located at the keyboardsFigure 6. Touch distribution and computed compensation functionfor the Galaxy S based on 3,242 installations that provided 1,831,489keystrokes. The white dots show the origin of the shift vectors, the yel-low dots their destination, and the blue lines show the actual vectors.border are set to zero. E.g. the y-component of vectors lo-cated at the upper boundary of the keyboard is set to zero toavoid shifting taps above the keyboard. Finally, shift vectorsfor further positions are derived using linear interpolation toproduce an array of 21x5 shift vectors. All touch events onthe keyboard are shifted according to the surrounding vectors.The touch distribution differs for different devices. Therefore,a compensation function is computed for each of the 50 mostcommon devices in the dataset. For the remaining devicesa function is computed using data from all devices with thisparticular resolution. However, all resolutions with less than150,000 keystrokes in the dataset are rejected which leads to58 different functions. The compensation functions are in-tegrated in the Android keyboard and applied before touchevents are processed by the keyboard.Shifting key labelsThis approach is based on the assumption that users are influ-enced by the locations of the keys labels and, at least to somedegree, try to hit the label. The labels are either shifted tothe upper part of the keys (elevated labels) or not (defaultlabels). To maintain the labels font size and leave roomfor larger upper case characters the labels are only slightlyshifted. The centre of the labels is shifted from the centre ofthe key to the upper third of the key depending on the devicesresolution. E.g. for the Optimus One the labels are shifted by11 pixels towards the top of the screen, the same amount ofpixels that players touches are skewed towards the bottom ofthe screen. As the labels shift only depends on the devicesresolution, it only roughly approximates the skew we foundin the previous section. The horizontal position is not altered.Figure 7 shows the difference between the default and the el-evated labels.Showing touched positions using dotsThe third approach tries to inform the user about the touchedposition in an unobtrusive way. A small red dot appears assoon as the users finger touches the screen. The dot followsthe touch events until the finger lift-off again and remains atFigure 7. The two different positions of the keys labels. The green la-bels show the default Android keyboard and the white labels show theelevated labels.Figure 8. Keyboard that shows a red dot at the position where the usertouches the screen after typing an f.this position. Thereby, the dot always shows the last touchedposition. We refrain from showing former touch positions orthe movement of the finger to keep the visualization as simpleas possible to increase the understandability and minimize theobtrusiveness. Figure 8 shows the keyboard after the user justtapped on the f.EXPERIMENTWe conducted an experiment to analyse the effect of the pro-posed approaches. The game was published as an update tothe Android Market after integrating the three approaches. Inthe following we describe the design of the experiment, fol-lowed by the results, and a discussion of our findings.DesignThe experiment follows an independent measures design toallow players adapting to the respective keyboard. The threeapproaches described in the previous section are used as theindependent variables. Table 3 provides an overview aboutthe independent variables and their degrees of freedom. Intotal, the design results in 12 conditions. Installations arerandomly assigned to one of the conditions when the gameis started for the first time or when the updated version isstarted for the first time. By randomly assigning installationsto one condition we assume that players are almost evenlydistributed across the 12 conditions.Three measures are used to assess the players behaviour.We assess the players speed by counting the number ofkeystrokes and measuring the time to complete a level to de-rive the number of keystrokes per second. In addition, wedetermine the players performance by counting the num-ber of keystrokes that contribute to the given words. E.g.independet variable degrees of freedomtouch no shiftnative shiftadapted shiftlabel default labelselevated labelsdot no dotdotTable 3. The three independent variables and their degrees of freedom.keystrokes that produced characters that are deleted after-wards are excluded. This is used to compute the number ofcorrect keystrokes per second. Finally, we assess the errorrate by dividing the number of keystrokes that lead to incor-rect characters or compensate errors by the total number ofkeystrokes.ResultsWe deployed an update of the game to the Android Mar-ket on July 31, 2011 and collected data until September 19,2011. We use the data provided by installations that updatedor newly installed the game. In total we received data from26,586 installations. For the analysis we only consider dataprovided by devices that can use one of the computed com-pensation functions. Thereby, we consider only data fromthe 50 most common devices and from the 8 most commonresolutions. Furthermore, we removed the first played levelfrom the data provided by each installation and data from allinstallations that contributed only a single level or providedinconsistent data, to reduce the noise in the data.We use data from 13,013 installations that contributed6,603,659 keystrokes by playing 120,662 levels for the fol-lowing analysis. On average an installation contributed 9.27levels (SD=25.54) and 507.47 keystrokes (SD=1,556.69).The average number of installations per condition is 1084.42(SD=36.24), the average number of played levels per con-dition is 10,055 (SD=938.46), and the average number ofkeystrokes per condition is 550,304 (SD=53,537). Becausethe device types and players locales are consistent with ourfirst observation we refrain from a detailed description.In the following we analyse the effect of the independentvariables using a three-way independent analysis of vari-ance (ANOVA). The differences between the individual con-ditions are analysed using Fishers Least Significant Differ-ence (LSD) post-hoc test. The levels of significance whencomparing the conditions are shown in Figure 12. The termcontrol condition is used for the condition with no shift, de-fault labels, and no dot - i.e. the Android keyboard withoutshifting the touch events by 10dp.Speed: To assess the effect of the independent variables onthe players speed we compare the number of keystrokes perseconds. The ANOVA show that the three main effects aswell as all interaction effects are significant (pFigure 9. Average speed assessed through the number keystrokes persecond (error bars show standard error).keystrokes per second than all other conditions. The differ-ence between these three conditions is, however, not signifi-cant. Players are 2.5% faster with one of the three conditionscompared to the control condition and players are 6.4% fasterwith the fastest condition (adapted shift with default labelsand no dot) compared to the slowest condition (no shift withelevated labels and no dot).Performance: To assess the players performance we in-vestigate the independent variables effect on the correctkeystrokes per second (i.e. those keystrokes that are not er-rors and do not compensate an error). The ANOVA showthat the two main effects touch and label (pFigure 12. Significance levels for comparison of the individual conditionsand the three depended variables using Fishers LSD.Figure 13. Normalized error rate for installations in the dots and theno dots conditions for the first 11 played levels. Error bars show stan-dard error.using data provided by 13,013 installations, 120,662 playedlevels and 6,603,659 touch events.Shifting touch positions: Using the default labels and nodot the shift function provided by the Android keyboard re-sults in 2.4% higher speed, 2.7% higher performance, anda 2.2% higher error rate compared to the control condition.The advantage for the adapted shift function is even higher.Using the adapted shift results in 2.6% higher speed, 5.0%higher performance, and a 7.7% lower error rate compared tothe control condition. The results show that the native shiftfunction improves the users typing. The adapted shift func-tion can, however, further improve the users performance by2.2% and decrease the error rate by 9.1% when compared tothe standard Android keyboard.Shifting key labels: For all conditions with a shift functionelevating the labels position decreases the speed, decreasesthe performance, and increases the error rate for the Androidkeyboard. The only condition that improved by elevating thelabels is without shift and with dot. Still, the overall resultsstrongly suggest that elevating the labels position to the up-per part of the key does not improve the users typing.Showing touched positions using dots: It is found that vi-sualizing where the users finger lift-off the keyboard usinga dot decreases the error rate for all conditions. Adding thedot to the control condition, for example, decreases the errorrate by 16.9% and adding the dot to the Android keyboard de-creases the error rate by 18.3%. For all conditions with the na-tive or the adapted shift function, however, the dot decreasesthe speed up to 5.2% and also decreases the performance. Toreduce the error rate users must pay attention to the dot. Weassume that this attention spend on the dot reduces the speedand as a consequence also the performance.Limitations: In line with previous work that investigates typ-ing on virtual keyboards one limitation of the study lies in thetask that is used to collect the data. While other work asksparticipants to copy text, often in a highly controlled envi-ronment using a single device, our task is part of a typinggame. Using a public game, however, has the advantage thatthe study has a high external validity. It can be assumed thatusers have more diverse backgrounds and use the game inmore diverse contexts than what can be achieved in commonlab studies. Furthermore, the study is not limited to one spe-cific device but considers a broad range of devices. This isespecially important because our observation of the touch be-haviour showed that the device influences how users type.As we collected data from players that installed the game attheir own will we have very limited control over the partici-pants. We cannot control who plays the game and when theystop playing. E.g. the probability that players with a low per-formance stop playing early might be higher than for playerswith a high performance. The used approach, however, en-ables to attract a truly large sample from all over the world.Instead of conducting a study with a small homogeneous sam-ple, as it is often common in other studies, the large sampleand the participants diversity increases the studys validity.CONCLUSIONWe investigated the touch behaviour on mobile devices vir-tual keyboards. To observe the behaviour of a large numberof users we developed a typing game. Using this game wecollected 47,770,625 keystrokes that have been produced by72,945 installations playing 952,487 levels. Analysing thedata shows that players touch systematically below the centreof the keys (e.g. 12.72px or 1.39mm for the Galaxy S). Fur-thermore, we found no strong correlation between the touchskew and the devices size. It can be concluded that otherfactors, such as the devices form factor or the devices usergroups, have a stronger effect on the typing behaviour. Weused the collected data to identify three approaches that mightinfluence the typing behaviour positively. We developed afunction that compensates the systematic skews found in thetouch distribution. Furthermore, we shift the keys labels tothe upper part of the keys and show the position where theusers finger lift-off the screen using a simple dot.To compare the three approaches we conducted an experi-ment by publishing an update of the game. For the experi-ment with three independent variables we collected data from13,013 installations that contributed 6,603,659 keystrokes byplaying 120,662 levels. The results disprove our assumptionsthat elevating the position of the keys labels is beneficial.Showing the users where they touch using a dot clearly im-proves the error rate. The error rate for the standard Androidkeyboard, for example, is decreased by 18.3%. The dot, how-ever, also decreases the typing speed and has no positive ef-fect on learnability. The usefulness of this feedback there-fore depends if the error rate or speed is more important. Weshowed that the simple shift function provided by the stan-dard Android keyboard improves the users speed but pro-vides no relevant improvement for the error rate. The adaptedshift function that we derived from our observation, however,further improves the performance by 2.2% and decreases theerror rate by 9.1% compared to the Android keyboard. Be-cause this shift function can be used as a drop-in replacementfor the Android keyboards shift function we assume that itcan improve the typing on current smartphones at no costs.The conducted studies have a low internal validity but, com-pared to common lab studies, a very high external validity.Our approach has the advantage that the results are based ona very large number of users, the participants are likely rep-resentative for Android users, and the data has been collectedin real life contexts from users using their own devices. Weassume that the very large number of observed users is bynature representative for smartphone users. There are, how-ever, also disadvantages of the used approach. Studies witha high external validity, such as ours, enable to observe theusers real behaviour. In contrast, studies with a high internalvalidity that investigate isolated aspects might be better suitedto find the causes for the observed behaviour.The collected data provides a rich source of information. Wewould like to share the data with other to enable further anal-ysis. In particular, analysing the data provided by differentdevices could provide interesting insights. Furthermore, thedata could be used for analysing touch sequences to dynam-ically adapt the keyboard while typing words. We are inter-ested in using the game to investigate further aspects that in-fluences the users typing behaviour. E.g. testing further vi-sual keyboard designs and shift functions. We also aim atinvestigating more radical keyboard designs and if other wid-gets can be investigated using mobile games.Acknowledgments: Parts of this work were conductedwithin the context of the Emmy Noether research group Mo-bile Interaction with Pervasive User Interfaces funded by theGerman Research Foundation (DFG).REFERENCES1. Apple Inc. iOS Human Interface Guidelines, 2010.2. Balakrishnan, V., and Yeow, P. A study of the effect ofthumb sizes on mobile phone texting satisfaction.Journal of Usability Studies 3 (2008).3. Gunawardana, A., Paek, T., and Meek, C. Usabilityguided key-target resizing for soft keyboards. In Proc.IUI (2010).4. Henze, N., Rukzio, E., and Boll, S. 100,000,000 taps:Analysis and improvement of touch performance in thelarge. In Proc. MobileHCI (2011).5. Himberg, J., Hakkila, J., Kangas, P., and Mantyjarvi, J.On-line personalization of a touch screen basedkeyboard. In Proc. IUI (2003).6. Holz, C., and Baudisch, P. Understanding touch. InProc. CHI (2011).7. Karlson, A. Interface Design for Single-Handed Use ofSmall Devices. In Proc. UIST (2008).8. Kristensson, P.-O., and Zhai, S. Relaxing stylus typingprecision by geometric pattern matching. In Proc. IUI(2005).9. Lee, S., and Zhai, S. The performance of touch screensoft buttons. In Proc. CHI (2009).10. Lewis, J. R. Literature review of touch screen researchfrom 1980 to 1992. In IBM Technical Report 54.694(1993).11. MacKenzie, I., and Soukoreff, R. Phrase sets forevaluating text entry techniques. In Adjunct Proc. CHI(2003).12. Parhi, P., Karlson, A., and Bederson, B. Target sizestudy for one-handed thumb use on small touchscreendevices. In Proc. MobileHCI (2006).13. Park, Y., Han, S., Park, J., and Cho, Y. Touch key designfor target selection on a mobile phone. In Proc.MobileHCI (2008).14. Perry, K., and Hourcade, J. Evaluating one handedthumb tapping on mobile touchscreen devices. In Proc.GI (2008).15. Roudaut, A., Huot, S., and Lecolinet, E. TapTap andMagStick: improving one-handed target acquisition onsmall touch-screens. In Proc. AVI (2008).16. Rudchenko, D., Paek, T., and Badger, E. Text textrevolution: A game that improves text entry on mobiletouchscreen keyboards. In Proc. Pervasive (2011).17. Schildbach, B., and Rukzio, E. Investigating selectionand reading performance on a mobile phone whilewalking. In Proc. MobileHCI (2010).18. Sears, A. Improving touchscreen keyboards: Designissues and a comparison with other devices. Interactingwith computers 3 (1991).19. Sears, A., and Zha, Y. Data entry for mobile devicesusing soft keyboards: Understanding the effects ofkeyboard size and user tasks. International Journal ofHuman-Computer Interaction 16 (2003).20. Vogel, D., and Baudisch, P. Shift: a technique foroperating pen-based interfaces using touch. In Proc.CHI (2007).21. Yatani, K., Partridge, K., Bern, M., and Newman, M.Escape: a target selection technique using visually-cuedgestures. In Proc. CHI (2008).IntroductionRelated WorkDesign of the GameGame playLevels and text sourcesMeasures and consentPublishing in the MarketDemographicsCollected dataObserved Touch BehaviourVertical offsetHorizontal offsetDiscussionInfluencing Touch BehaviourShifting touch positionsShifting key labelsShowing touched positions using dotsExperimentDesignResultsDiscussionConclusionREFERENCES