uncertain interaction - university of cambridge · 2019-01-30 · computational interaction •...
TRANSCRIPT
UncertainInteraction
PerOlaKristenssonDepartmentofEngineeringUniversityofCambridge
Solution principles for
VisionsoftheFuture
VisionsoftheFuture
• Ubiquitoussensing– Smarthome,InternetofThings,etc.
• Pervasiveagents– Spokendialogue-basedcommandandquery
interfaces• Virtualreality– Portableoffice,training,immersivedataanalytics
• Phonewithoutaphone– Opticalsee-throughhead-mounteddisplayswithform
factorscomparabletoeverydayglasses
VisionsoftheFuture
• Ubiquitoussensing– Smarthome,InternetofThings,etc.
• Pervasiveagents– Spokendialogue-basedcommandandquery
interfaces• Virtualreality– Portableoffice,training,immersivedataanalytics
• Phonewithoutaphone– Opticalsee-throughhead-mounteddisplayswithform
factorscomparabletoeverydayglasses
Allassumefluidinterfacesbasedonfundamentallyuncertaininteraction
Computationalinteraction• Classichuman-computerinteraction(HCI)method
doesnothandleuserinterfacedesignunderuncertaintyverywell
• ClassicHCImethodisunderpinnedonelicitinguserneedsusingavarietyofprocessesandthenaniterativeprocessofdesignandevaluation,inwhichdesignisdrivenbydesigningenuityratherthanprinciples
• Thismeans:– Noautomateddesignwork– Noexplicitmodel– Datainfluenceddesignonlythroughthedesigner
• ComputationalinteractionisanemergingdisciplineinHCIwhichproposesuserinterfacedevelopmentbyallowingalgorithmstoperformwork,byexplicitmodelling,andbyallowingdatatodirectlyinfluencedesign.
Computationalinteraction• Computationalinteractionwouldtypicallyinvolve
atleastoneof:I. anexplicitmathematicalmodelofuser-system
behavior;II. awayofupdatingthatmodelwithobserveddata
fromusers;III. analgorithmicelementthat,usingthismodel,can
directlysynthesiseoradaptthedesign;IV. awayofautomatingandinstrumentingthe
modelinganddesignprocess;V. theabilitytosimulateorsynthesiseelementsofthe
expecteduser-systembehavior.
Intelligenttextentryasanexampleofdesigninginteractionunder
uncertainty
Principlesofintelligenttextentry
Kristensson,P.O.2009.Fivechallengesforintelligenttextentrymethods.AIMagazine30(4):85-94.
Principlesofintelligenttextentry
1. Letterssimplifiedtolinemarks
2. Commonwordstemscompressedintosimplelinemarksordots
3. Commonwordstemsidentifiedbywordfrequencyanalysisofthebookofpsalms
Kristensson,P.O.2009.Fivechallengesforintelligenttextentrymethods.AIMagazine30(4):85-94.
Principlesofintelligenttextentry
1. Letterssimplifiedtolinemarks
2. Commonwordstemscompressedintosimplelinemarksordots
3. Commonwordstemsidentifiedbywordfrequencyanalysisofthebookofpsalms
Kristensson,P.O.2009.Fivechallengesforintelligenttextentrymethods.AIMagazine30(4):85-94.
Principlesofintelligenttextentry
1. Letterssimplifiedtolinemarks
2. Commonwordstemscompressedintosimplelinemarksordots
3. Commonwordstemsidentifiedbywordfrequencyanalysisofthebookofpsalms
Kristensson,P.O.2009.Fivechallengesforintelligenttextentrymethods.AIMagazine30(4):85-94.
Principlesofintelligenttextentry• Inotherwords:
1. Optimisespeedbyminimsingtheamountofinformationusershavetoarticulate
2. Exploitredundanciesinnaturallanguagesbycreatingalanguagemodel
Kristensson,P.O.2009.Fivechallengesforintelligenttextentrymethods.AIMagazine30(4):85-94.
Principlesofintelligenttextentry• Inotherwords:
1. Optimisespeedbyminimsingtheamountofinformationusershavetoarticulate
2. Exploitredundanciesinnaturallanguagesbycreatingalanguagemodel
Kristensson,P.O.2009.Fivechallengesforintelligenttextentrymethods.AIMagazine30(4):85-94.
Principlesofintelligenttextentry• Inotherwords:
1. Optimisespeedbyminimsingtheamountofinformationusershavetoarticulate
2. Exploitredundanciesinnaturallanguagesbycreatingalanguagemodel
Kristensson,P.O.2009.Fivechallengesforintelligenttextentrymethods.AIMagazine30(4):85-94.
Principlesofintelligenttextentry• ...whichcanoftenbe
thoughtofasaninferenceproblem:
Kristensson,P.O.2009.Fivechallengesforintelligenttextentrymethods.AIMagazine30(4):85-94.
Whydonearlyalltextentrymethodsfail?
Mainstreammobiletextentrymethods
Mainstreammobiletextentrymethods
Graffiti
Mainstreammobiletextentrymethods
Graffiti Multi-tapandpredictivetext
Mainstreammobiletextentrymethods
Graffiti Multi-tapandpredictivetext
Touchscreenkeyboards
Mainstreammobiletextentrymethods
Graffiti Multi-tapandpredictivetext
Touchscreenkeyboards
Gesturekeyboards
Mainstreammobiletextentrymethods
Graffiti Multi-tapandpredictivetext
Touchscreenkeyboards
Gesturekeyboards Physicalthumb
keyboards
Mainstreammobiletextentrymethods
• Entryanderrorrate• Learningcurve,familiarity
andimmediateefficacy• Formfactor,preparation
timeandcomfort• Userengagement• Visualattentionand
cognitiveresources• Privacy• Singlevs.multi-character
entry
• Specificationvs.navigation
• One-handedvs.two-handed
• Taskintegration• Robustness• Deviceindependence• Computationaldemands• Manufacturingand
supportcost• Localisation• Marketacceptance
Mainstreammobiletextentrymethods
• Entryanderrorrate• Learningcurve,familiarity
andimmediateefficacy• Formfactor,preparation
timeandcomfort• Userengagement• Visualattentionand
cognitiveresources• Privacy• Singlevs.multi-character
entry
• Specificationvs.navigation
• One-handedvs.two-handed
• Taskintegration• Robustness• Deviceindependence• Computationaldemands• Manufacturingand
supportcost• Localisation• Marketacceptance
• Higheffectiveentryrate– Amongthefastestoftheirgeneration
• Highfamiliarityandhighimmediateefficacy– Eitherextremelyeasy-to-learnorverysimilartoexistingtechnology(orboth)
Mainstreammobiletextentrymethods
• Entryanderrorrate• Learningcurve,familiarity
andimmediateefficacy• Formfactor,preparation
timeandcomfort• Userengagement• Visualattentionand
cognitiveresources• Privacy• Singlevs.multi-character
entry
• Specificationvs.navigation
• One-handedvs.two-handed
• Taskintegration• Robustness• Deviceindependence• Computationaldemands• Manufacturingand
supportcost• Localisation• Marketacceptance
• Higheffectiveentryrate– Amongthefastestoftheirgeneration
• Highfamiliarityandhighimmediateefficacy– Eitherextremelyeasy-to-learnorverysimilartoexistingtechnology(orboth)
Mobiletextentry:thestateoftheart
Speed(wpm)
Yearsofmobiletextentryresearch
2006-2008 2014
Mobiletextentry:thestateoftheart
Speed(wpm)
Yearsofmobiletextentryresearch
40-60wpm(ceilingrates)
2006-2008 2014
Mobiletextentry:thestateoftheart
Speed(wpm)
Yearsofmobiletextentryresearch
40-60wpm(ceilingrates)
Mainstreamtextentrymethods(touchscreenQWERTY,thumbkeyboard,gesturekeyboard)
2006-2008 2014
Mobiletextentry:thestateoftheart
Speed(wpm)
Yearsofmobiletextentryresearch
40-60wpm(ceilingrates)
Mainstreamtextentrymethods(touchscreenQWERTY,thumbkeyboard,gesturekeyboard)
2006-2008 2014
Researchtextentrymethods(slowerthancommercialsolutions)
Mobiletextentry:thestateoftheart
Speed(wpm)
Yearsofmobiletextentryresearch
40-60wpm(ceilingrates)
Mainstreamtextentrymethods(touchscreenQWERTY,thumbkeyboard,gesturekeyboard)
2006-2008 2014
Researchtextentrymethods(slowerthancommercialsolutions) Optimised
keyboards
Mainstreammobiletextentrymethods
• Entryanderrorrate• Learningcurve,familiarity
andimmediateefficacy• Formfactor,preparation
timeandcomfort• Userengagement• Visualattentionand
cognitiveresources• Privacy• Singlevs.multi-character
entry
• Specificationvs.navigation
• One-handedvs.two-handed
• Taskintegration• Robustness• Deviceindependence• Computationaldemands• Manufacturingand
supportcost• Localisation• Marketacceptance
• Higheffectiveentryrate– Amongthefastestoftheirgeneration
• Highfamiliarityandhighimmediateefficacy– Eitherextremelyeasy-to-learnorverysimilartoexistingtechnology(orboth)
Thecross-overpoint
Time
Performance
Thecross-overpoint
Time
Performance
Familiarinterface
Thecross-overpoint
Time
Performance
Unfamiliarinterface
Familiarinterface
Thecross-overpoint
Time
Performance
Unfamiliarinterface
Familiarinterface
Thecross-overpoint
Time
Performance
Unfamiliarinterface
Familiarinterface
Cross-overpoint
Thecross-overpoint
Time
Performance
Unfamiliarinterface
Familiarinterface
Timeinvestment
Cross-overpoint
Thecross-overpoint
Time
Performance
Unfamiliarinterface
Familiarinterface
Benefit
Timeinvestment
Cross-overpoint
Thecross-overpoint
Time
Performance
Thecross-overpoint
Time
Performance
Thecross-overpoint
Time
Performance
Thecross-overpoint
Objectivebenefit
Nicosia,M.,Oulasvirta,A.andKristensson,P.O.2014.Modelingtheperceptionofuserperformance.InProceedingsofthe32ndACMConferenceonHumanFactorsinComputingSystems(CHI2014).ACMPress:1747-1756.
Thecross-overpoint
Perceivedbenefit
Nicosia,M.,Oulasvirta,A.andKristensson,P.O.2014.Modelingtheperceptionofuserperformance.InProceedingsofthe32ndACMConferenceonHumanFactorsinComputingSystems(CHI2014).ACMPress:1747-1756.
Thenarrowdesignspace
Thenarrowdesignspace
Thenarrowdesignspace
Thenarrowdesignspace
Interactionstrategies
Thenarrowdesignspace
Efficientencodings
Interactionstrategies
Thenarrowdesignspace
Optimisinglayouts
Efficientencodings
Interactionstrategies
Thenarrowdesignspace
Thenarrowdesignspace
Thenarrowdesignspace
Solutionprinciples• Fromclosedtoopen-loop
– Avoidtheneedforavisualfeedbackloop• Continuousnovice-to-experttransition
– Avoidexplicitlearning• Pathdependency
– Avoidredesigningtheinteractionlayer• Flexibility
– Enableuserstocomposeandeditinavarietyofstyleswithoutexplicitmodeswitching
• Probabilisticerrorcorrection– Usethehypothesisspacetodesignoptimalerrorcorrection
strategies• Fluidregulationofuncertainty
– Allowuserstoseamlesslyinfluencetheinferenceprocess• Efficiency
– Letusers’creativitybethebottle-neck
FromClosedtoOpenLoop
Reimagingthekeyboard
Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
Justthen,thewhite
Howgesturekeyboardswork
Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
Justthen,thewhite
Howgesturekeyboardswork
Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
Justthen,thewhite
Howgesturekeyboardswork
Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
Justthen,thewhite
Howgesturekeyboardswork
Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
Justthen,thewhite
Howgesturekeyboardswork
Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
Justthen,thewhite
Howgesturekeyboardswork
Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
Justthen,thewhite
Howgesturekeyboardswork
Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
Justthen,thewhite
Howgesturekeyboardswork
Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
Justthen,thewhite
Howgesturekeyboardswork
Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
Justthen,thewhite
Howgesturekeyboardswork
Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
Justthen,thewhite
Howgesturekeyboardswork
Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
Justthen,thewhite
Howgesturekeyboardswork
Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
Justthen,thewhite
Howgesturekeyboardswork
Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
Justthen,thewhite
Priorprobability
Howgesturekeyboardswork
Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
Justthen,thewhite
Priorprobability
Likelihood
Howgesturekeyboardswork
Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
Justthen,thewhiterabbit
Priorprobability
Likelihood
Posteriorprobability
Howgesturekeyboardswork
Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
Justthen,thewhiterabbit
Priorprobability
Likelihood
Posteriorprobability
Decodingnoisygesturesintotextusingacombinationofgesturerecognitionandlanguagemodelling
Closed-andopen-loop• Closed-loop:
– Continuousfeedback-driveninteraction– Visually-guidedmotion– Slowandprecise– Modelledwellbythe“crossinglaw”
• Averagemovementtime=a+blog2(D/W+1);aandbarelinearregressioncoefficients;DandWarethedistanceandwidthtothecrossinggoalrespectively
• Open-loop:– Notfeedback-driven– Directrecallfrommotormemory– Fastandimprecise– Nogoodmodelexits
• Gesturekeyboardinteractionisamixofclosed-andopen-loopinteraction
ContinuousNovice-to-ExpertTransition
Continuoustransitionfromnovicetoexpertbehaviour
Skillacquisition
Fallingbackandrelearning
Consistentmovementpattern
Completenovice:TracinglettertoletterClosed-loopSlowandaccurate
Completeexpert:GesturingwordshapesOpen-loopFastandinaccurate
PathDependency
Example:typingonasmartwatch
• Smallscreensizeisobviouslyaconstraint• Manynaïvesolutions:– Progressivezoomingtechniques– Reducekeyset(álatheoldtelephonekeypadtechniques)
– Variousmulti-strokestrategies• Allslow• Alldemanduserlearning(noimmediateefficacy)
Thecross-overpoint
Time
Performance
Newsmartwatchinputmethod
Familiarinterface
Benefit
Timeinvestment
Cross-overpoint
Thecross-overpoint
Time
Performance
Newsmartwatchinputmethod
Familiarinterface
Benefit
Timeinvestment
Cross-overpoint
=40hoursofdedicatedpractice
Thecross-overpoint
Time
Performance
Newsmartwatchinputmethod
Familiarinterface
Benefit
Timeinvestment
Cross-overpoint
=40hoursofdedicatedpractice
Assumetheusertypesforfiveminutesontheirsmartwatcheveryday
Thecross-overpoint
Time
Performance
Newsmartwatchinputmethod
Familiarinterface
Benefit
Timeinvestment
Cross-overpoint
=40hoursofdedicatedpractice
Userperformanceaftermonthsofuse
Assumetheusertypesforfiveminutesontheirsmartwatcheveryday
Thecross-overpoint
Time
Performance
Newsmartwatchinputmethod
Familiarinterface
Benefit
Timeinvestment
Cross-overpointreachedafter480days
=40hoursofdedicatedpractice
Userperformanceaftermonthsofuse
Assumetheusertypesforfiveminutesontheirsmartwatcheveryday
Mainstreammobiletextentrymethods
• Entryanderrorrate• Learningcurve,familiarity
andimmediateefficacy• Formfactor,preparation
timeandcomfort• Userengagement• Visualattentionand
cognitiveresources• Privacy• Singlevs.multi-character
entry
• Specificationvs.navigation
• One-handedvs.two-handed
• Taskintegration• Robustness• Deviceindependence• Computationaldemands• Manufacturingand
supportcost• Localisation• Marketacceptance
• Higheffectiveentryrate– Amongthefastestoftheirgeneration
• Highfamiliarityandhighimmediateefficacy– Eitherextremelyeasy-to-learnorverysimilartoexistingtechnology(orboth)
Mainstreammobiletextentrymethods
• Entryanderrorrate• Learningcurve,familiarity
andimmediateefficacy• Formfactor,preparation
timeandcomfort• Userengagement• Visualattentionand
cognitiveresources• Privacy• Singlevs.multi-character
entry
• Specificationvs.navigation
• One-handedvs.two-handed
• Taskintegration• Robustness• Deviceindependence• Computationaldemands• Manufacturingand
supportcost• Localisation• Marketacceptance
• Higheffectiveentryrate– Amongthefastestoftheirgeneration
• Highfamiliarityandhighimmediateefficacy– Eitherextremelyeasy-to-learnorverysimilartoexistingtechnology(orboth)
• IttakesaverylongtimetolearnQWERTY(orlearnanewlayout)
Mainstreammobiletextentrymethods
• Entryanderrorrate• Learningcurve,familiarity
andimmediateefficacy• Formfactor,preparation
timeandcomfort• Userengagement• Visualattentionand
cognitiveresources• Privacy• Singlevs.multi-character
entry
• Specificationvs.navigation
• One-handedvs.two-handed
• Taskintegration• Robustness• Deviceindependence• Computationaldemands• Manufacturingand
supportcost• Localisation• Marketacceptance
• Higheffectiveentryrate– Amongthefastestoftheirgeneration
• Highfamiliarityandhighimmediateefficacy– Eitherextremelyeasy-to-learnorverysimilartoexistingtechnology(orboth)
• IttakesaverylongtimetolearnQWERTY(orlearnanewlayout)
• UsersarefamiliarwithtouchscreenQWERTY
Mainstreammobiletextentrymethods
• Entryanderrorrate• Learningcurve,familiarity
andimmediateefficacy• Formfactor,preparation
timeandcomfort• Userengagement• Visualattentionand
cognitiveresources• Privacy• Singlevs.multi-character
entry
• Specificationvs.navigation
• One-handedvs.two-handed
• Taskintegration• Robustness• Deviceindependence• Computationaldemands• Manufacturingand
supportcost• Localisation• Marketacceptance
• Higheffectiveentryrate– Amongthefastestoftheirgeneration
• Highfamiliarityandhighimmediateefficacy– Eitherextremelyeasy-to-learnorverysimilartoexistingtechnology(orboth)
• IttakesaverylongtimetolearnQWERTY(orlearnanewlayout)
• UsersarefamiliarwithtouchscreenQWERTY
• KeepQWERTY
Mainstreammobiletextentrymethods
• Entryanderrorrate• Learningcurve,familiarity
andimmediateefficacy• Formfactor,preparation
timeandcomfort• Userengagement• Visualattentionand
cognitiveresources• Privacy• Singlevs.multi-character
entry
• Specificationvs.navigation
• One-handedvs.two-handed
• Taskintegration• Robustness• Deviceindependence• Computationaldemands• Manufacturingand
supportcost• Localisation• Marketacceptance
• Higheffectiveentryrate– Amongthefastestoftheirgeneration
• Highfamiliarityandhighimmediateefficacy– Eitherextremelyeasy-to-learnorverysimilartoexistingtechnology(orboth)
• IttakesaverylongtimetolearnQWERTY(orlearnanewlayout)
• UsersarefamiliarwithtouchscreenQWERTY
• KeepQWERTY
Touchmodelling
2DGaussianscenteredateachkey.Separatevariancesinthex-andy-dimensions.
Vertanen,K.,Memmi,H.,Emge,J.,Reyal,S.andKristensson,P.O.2015.VelociTap:investigatingfastmobiletextentryusingsentence-baseddecodingoftouchscreenkeyboardinput.InProceedingsofthe33rdACMConferenceonHumanFactorsinComputingSystems(CHI2015).ACMPress:659-668.
Languagemodelling
• Languagemodels:– 12-gramlettermodel– 4-gramwordmodelwithunknownword– Trainedonbillionsofwordsofdata
§ Twitter,blog,socialmedia,Usenet,andwebdata
– Optimizedforshortemail-likemessages– Letter+wordlanguagemodel=~4GBmemory
Vertanen,K.,Memmi,H.,Emge,J.,Reyal,S.andKristensson,P.O.2015.VelociTap:investigatingfastmobiletextentryusingsentence-baseddecodingoftouchscreenkeyboardinput.InProceedingsofthe33rdACMConferenceonHumanFactorsinComputingSystems(CHI2015).ACMPress:659-668.
Decoding
Observation1 Observation2 Observation3fgc
z
ϵ
abz
Tokenstrack:probability,LMcontext,traceback
o
ϵ
a
z
X X
d
ϵ
o za
X
d
good
god
go
Beamprunetokeeptractable
X XX
X
Entryanderrorrate
Condition
Normal Standardportraitkeyboard,60mmwide
Small Bigsmartwatch,40mmwide
Tiny Smallsmartwatch,25mmwide
Typingonatinykeyboard
Flexibility
Speechrecognitionerrorcorrection:thestandardmethod
• User:“thecatsat”
Speechrecognitionerrorcorrection:thestandardmethod
• User:“thecatsat”• System:“thebatsat”
Speechrecognitionerrorcorrection:thestandardmethod
• User:“thecatsat”• System:“thebatsat”• User:“selectbat”
Speechrecognitionerrorcorrection:thestandardmethod
• User:“thecatsat”• System:“thebatsat”• User:“selectbat”• System:“thebatsatdissectrat”
Speechrecognitionerrorcorrection:thestandardmethod
• User:“thecatsat”• System:“thebatsat”• User:“selectbat”• System:“thebatsatdissectrat”• (User:“Ihatethis…”)
Theflexiblemultimodalfusionapproach
• Userspeaks:“thecatsat”
Theflexiblemultimodalfusionapproach
• Userspeaks:“thecatsat”• System:“thebatsat”
Theflexiblemultimodalfusionapproach
• Userspeaks:“thecatsat”• System:“thebatsat”• Usergesturestheword:“cat”
Theflexiblemultimodalfusionapproach
• Userspeaks:“thecatsat”• System:“thebatsat”• Usergesturestheword:“cat”• System:“thecatsat”
Theflexiblemultimodalfusionapproach
• Userspeaks:“thecatsat”• System:“thebatsat”• Usergesturestheword:“cat”• System:“thecatsat”
• Thesystemautomaticallyidentifiestheerrorlocationandcorrectstheerror
Kristensson,P.O.andVertanen,K.2011.Asynchronousmultimodaltextentryusingspeechandgesturekeyboards.InProceedingsofthe12thAnnualConferenceoftheInternationalSpeechCommunicationAssociation(Interspeech2011).ISCA:581-584.
Outputfromatextentrymodality
Gesturekeyboard
Outputfromatextentrymodality
Gesturekeyboard
Timestep1
Outputfromatextentrymodality
Gesturekeyboard
Timestep1
Outputfromatextentrymodality
Gesturekeyboard
Timestep1 Timestep2
Outputfromatextentrymodality
Gesturekeyboard
thee0.3
the0.6
three0.1
Outputfromatextentrymodality
Gesturekeyboard
the0.3
ε0.6
thee0.1
Outputfromatextentrymodality
Gesturekeyboard
thee0.3
the0.6
three0.1
VAT0.2
Outputfromatextentrymodality
Gesturekeyboard
thee0.3
the0.6
three0.1
cat0.6
cart0.2
at0.28rat0.28
Outputfromtwotextentrymodalities
the0.56
a0.38
cat0.82
at0.06fat0.06
ε0.87
at0.09
sat0.75
nat0.19
the0.94 bat0.57
cat0.09
sat0.47
ε0.21
at0.28rat0.28
Softeningthewordconfusionnetworks:addingwild-cardtransitions
the0.56
a0.38*0.03
cat0.82
at0.06fat0.06*0.03
ε0.87
at0.09*0.03
sat0.75
nat0.19*0.03
the0.94
*0.03
bat0.57
cat0.09*0.03
sat0.47
ε0.21*0.03
at0.28rat0.28
Softeningthewordconfusionnetworks:addingepsilontransitions
the0.56
a0.38*0.03ε0.02
cat0.82
at0.06fat0.06*0.03ε0.02
ε0.87
at0.09*0.03
sat0.75
nat0.19*0.03ε0.02
the0.94
*0.03ε0.02
bat0.57
cat0.09*0.03ε0.02
sat0.47
ε0.21*0.03
at0.28rat0.28
Softeningthewordconfusionnetworks:addingwild-cardself-loops
the0.56
a0.38*0.03ε0.02
0.01*
0.01* cat0.82
at0.06fat0.06*0.03ε0.02
0.01* ε0.87
at0.09*0.03
0.01* sat0.75
nat0.19*0.03ε0.02
0.01*
the0.94
*0.03ε0.02
0.01*
0.01* bat0.57
cat0.09*0.03ε0.02
0.01* sat0.47
ε0.21*0.03
0.01*
at0.28rat0.28
Searchforthehighestjointpathinbothrecognitionmodalities
the0.56
a0.38*0.03ε0.02
0.01*
0.01* cat0.82
at0.06fat0.06*0.03ε0.02
0.01* ε0.87
at0.09*0.03
0.01* sat0.75
nat0.19*0.03ε0.02
0.01*
the0.94
*0.03ε0.02
0.01*
0.01* bat0.57
cat0.09*0.03ε0.02
0.01* sat0.47
ε0.21*0.03
0.01*
Speech-onlyflexiblerepair
Probabilisticerrorcorrection
Probabilisticerrorcorrection
• Foranyprobabilistictextentrymethod…– Capableofassigningposteriorprobability
distributionstowords• …thereexistsahypothesisspace• Thebestresultisthemaximumprobabilitypath
inthishypothesisspace– However,itneednotbetheonetheuserintended
• Byexposingpartofthehypothesisspacetousers,highefficienciescanbegainedwhenuserscorrectwords
Fluidregulationofuncertainty
Theauto-correcttrap• Auto-correctisgreatwhenitworks• However,whenauto-correctfailserrorcorrectionactivities
exhibitahighpenalty• Thesolutionistoprovideuserswithmoreagencyand
allowthemtoregulatetheircertainty
Weir,D.,Pohl,H.,Rogers,S.,Vertanen,K.andKristensson,P.O.2014.Uncertaintextentryonmobiledevices.InProceedingsofthe32ndACMConferenceonHumanFactorsinComputingSystems(CHI2014).ACMPress:2307-2316.
Pressure-sensitiveauto-correct
• LikelihoodofaGaussianwithstandarddeviationregulatedbypressure
• StandarddeviationcomputedasC/ωT,whereCisaconstantandωTisthepressurefortouchT
• TunedCsothatthepressureofatypicaltouchhadastandarddeviationofhalfakeywidth
Weir,D.,Pohl,H.,Rogers,S.,Vertanen,K.andKristensson,P.O.2014.Uncertaintextentryonmobiledevices.InProceedingsofthe32ndACMConferenceonHumanFactorsinComputingSystems(CHI2014).ACMPress:2307-2316.
Results• Enablinguserstoregulatetheircertaintybyforce
resultedina10%percentagedropinactivecorrections(fixingawordbybackspacingorretyping)
• Thisimprovedentryrateby20%
Efficiency
Eye-typing
Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
Eye-typing
Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
Eye-typing
Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
125ms
Eye-typing
Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
250ms
Eye-typing
Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
375ms
Eye-typing
Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
500ms
Eye-typing
Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
625ms
Eye-typing
Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
750ms
Eye-typing
Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
875ms
Eye-typing
Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
1000ms
Recordspeedsachievedwhenwritingbygaze
• Eye-typing– 5–10wpm(MajarantaandRäihä2002;Roughetal.2014)
• Eye-typingwithadjustable-dwell– 7-20wpm(Majarantaetal.2009;RäihäandOvaska2012;Roughetal.2014)
• Dasher– 12–26wpm(Tuiskuetal.2008;WardandMacKay2002;Roughetal.2014)
Dwell-freeeye-typing
Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
Kristensson,P.O.andVertanen,K.2012.Thepotentialofdwell-freeeye-typingforfastassistivegazecommunication.InProceedingsofthe7thACMSymposiumonEye-TrackingResearch&Applications(ETRA2012).ACMPress:241-244.
Dwell-freeeye-typing
Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
Dwell-freeeye-typing
Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
Dwell-freeeye-typing
Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
Dwell-freeeye-typing
Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
Dwell-freeeye-typing
Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
Dwell-freeeye-typing
Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
Dwell-freeeye-typing
Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
Thecat
Dwell-freeeye-typing
Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
Humanperformanceestimateofdwell-freeeye-typing
• Recorded400minutesofeye-tracedata• Participantsenteredatotalof2026phrases• Participantswerepromptedphrasesandaskedtocopy
themasquicklyandasaccuratelyaspossible• Oursystemknewwhattheuserwassupposedtowriteand
verifiedthattheuserisgazingatthelettersequencecorrespondingtothestimulus
Entryrate
Humanperformancemodel
Humanperformancemodel
Eye-typingusingadjustabledwell,finalentryrate(mean=20wpm)
Humanperformancemodel
Eye-typingusingadjustabledwell,finalentryrate(mean=20wpm)
Humanperformancemodel
Eye-typingusingadjustabledwell,finalentryrate(mean=20wpm)
230%
Entryrate,first10-15minutes
Entryrate,first10-15minutes
Eye-typingusingadjustabledwell,entryrateinthefirstsession(mean=6.9wpm)
Entryrate,first10-15minutes
Eye-typingusingadjustabledwell,entryrateinthefirstsession(mean=6.9wpm)
520%
Astep-changeingazecommunication
• Existinggazecommunicationsolutions– Limitedtocirca20wpm
• Dwell-freeeye-typing– Empiricallymeasuredhumanperformancepotential:46wpmaverage
• Releasedasaproduct:Tobii-DynavoxI-Series+
Conclusions• Atextentrymethodlikelytobeadoptedbyusersis
probablysimilartoexistingsolutionsandatleastasfast• Itisstillpossibletomakeprogressbyusingafewsolution
principles:– Fromclosedtoopen-loop– Continuousnovice-to-experttransition– Pathdependency– Flexibility– Probabilisticerrorcorrection– Fluidregulationofuncertainty– Efficiency
• Ingeneral,thesecanbeviewedassolutionprinciplesforuncertaininteraction
Kristensson,P.O.2015.Next-generationtextentry.IEEEComputer48(7):84-87.