Text Entry Performance of VirHKey in Keyboard Use Benoît MARTIN University Paul Verlaine - Metz Ile du Saulcy 57045 Metz Cedex 1 France [email protected]

Poika ISOKOSKI University of Tampere Department of Computer Sciences 33014 University of Tampere Finland [email protected]

ABSTRACT The Virtual Hyperbolic Keyboard (VirHKey) is a text entry method that was originally designed to be used with a stylus. We constructed and tested a keyboard-based version of VirHKey. Average text entry rate ranged from 5 wpm during the first 20-minute session to 21 wpm during the 20th session. The average character error rate was under 2%. VirHKey is unlikely to be the most efficient method on any input device. Its value lies in the ability to transfer text entry skill between input devices. KEY WORDS Focus and context, hyperbolic geometry, independence, text entry, reduced keyboard.

device-

1. Introduction Systems like handwriting recognizers, miniature keypads, and soft keyboards are used when desktop keyboards are not convenient. The Virtual Hyperbolic Keyboard (VirHKey) is one of the systems that have been proposed for non-desktop text entry situations. A stylus-based implementation of VirHKey was proposed by Martin et al. [1]. In this paper we describe a keyboard-based version of VirHKey. Making text entry methods compatible with several input devices has been suggested in previous work [2][3], but these proposals have not been popular in practice. One of the reasons for the present study was to see if VirHKey performed well enough with a 6-key keyboard to finally make device-independent text entry a practical solution.

2. Previous work In this section we describe and discuss relevant previous work. The systems are characterized by the a-priori Keystrokes Per Character (KSPC) measure. KSPC is a number that tells how many key presses or other input acts are needed for an average character. Empirically measured text entry rate (in words per minute (wpm) and character error rate are also given when available. 2.1 Multi-device systems

MDITIM is a system that uses continuous stylus gestures in the four cardinal directions without explicit segmentation [3]. Input devices that produce four different events can be used. Wobbrock et al. [4] calculated MDITIM KSPC to be 3.06 using the frequencies from Soukoreff and MacKenzie [5]. With a four-key keyboard, Isokoski and Raisamo [3] reported a writing speed of ~4.9 wpm and ~3% errors after ten 30minute training sessions with a touchpad. EdgeWrite was designed for low accuracy demands in stylus-based text entry [6]. The stroke recognition is based on the order in which the gesture visits the corners of a square. Any input device that can produce five different events can be used. In its four-key version, a key is associated to a corner and the segmentation is obtained after a time stamp [4]. The KSPC for EdgeWrite is 3.79 using the frequencies reported in Wikipedia for the French language. Wobbrock et al. [2] reported an average writing speed of 15.95 wpm and less than 2% of uncorrected errors after ten 5-minute sessions. Quikwriting uses continuous stylus gestures over areas arranged in a 3×3 grid [7]. The gestures are loops that start and end in the center crossing one to three peripheral areas on the way. In keyboard use, its KSPC is 2. After twenty 15-minute sessions using stylus and gamepad, Isokoski and Raisamo reported text entry rate of 6.1 wpm during the first five minutes of using the nine-key keyboard [8]. For the next 5 minutes without the layout diagram, the speed increased to 8 wpm. Glyph is based on a decomposition of the characters in six primitives, according to an analogy with Roman characters [9]. Version 2, was designed to make glyph easier, more flexible, and faster. In six-key keyboard use, the KSPC is 2. Poitier and Belatar [9] reported an average writing speed of 7 wpm after minutes of practice. The systems mentioned above were static. Dasher [10] is a dynamic system: the user steers through an animated space populated by characters. More likely characters have a larger area making them easier to select. According to Ward et al. [10], Dasher text entry rates

reached on average ~20 wpm after one hour of training. With mouse as the input device text entry rate of 34 wpm was reported for an expert. Any input device with either continuous position input or one buttons can be used with Dasher. 2.2 Reduced Keyboards Reduced keyboards have fewer keys than there are letters in the alphabet. A well-known reduced keyboard is the telephone keypad with multi-tap and disambiguating modes of input. With multi-tap, we computed a KSPC of 2.21 using the frequencies reported in Wikipedia for the French language. James and Reischel reported novice text entry rate of 7.98 wpm and 7.93 wpm for experts [11]. Theoretical and laboratory studies suggest that disambiguation systems are helpful [12]. The KSPC of a telephone keypad using a dictionary-based disambiguation and English language is 1.0072 [13]. James and Reischel reported 9.09 wpm for novice T9 users and 20.36 wpm for experts [11]. In addition to the telephone keypad, and the whole continuum of key configurations that can be used with disambiguation systems, there are other products and proposals. The three-key keyboard has two keys to move a selector left and right in a list of character and a third key for selecting [13][14]. As an example, the “particularly promising” layout proposed by MacKenzie [14] has a KSPC of 10.63 using the frequencies from [5]. MacKenzie reports an average writing speed of 9.1 wpm and 2.11% of uncorrected errors after ten 5-minute sessions [14]. The five-key keyboard used in some early two-way pagers is another example [4][13]. It has four directional keys for moving on a soft keyboard and one key for selecting. The KSPC with the layout of the Glenayre AccessLink II two-way pager is 3.13 [4]. Bellman and MacKenzie report text entry speeds around 10 wpm after two hours of practice [15]. The 3-key and the 5-key methods can be greatly enhanced by key-repeat. A short repeat time allows high performance but can introduce difficulties to control the cursor. In chord keyboards a character is associated with a combination of keys that are pressed simultaneously. The number of keys on a chord keyboard varies. CyKey [16] is a 9 key system, the GKOS [17] uses only 6 keys and the Twiddler has 12. The multi-tap version of twiddled has a KSPC of 2.0432. The KSPC for the chorded version is 1.4764. In an experiment by Lyons et al [18] novices entered 8.2 wpm and experts 19.8 wpm. In chord mode novices entered 4.3 wpm and experts 26.2 wpm. 2.3 Summary

QuikWriting

2

6.1 - 8

novice with layout 5’ without layout Glyph 2 2 7 few minutes Dasher NC 20 - 34 1 hr - expert MultiTap 2.21 8 / 7.9 expert T9 1.007 9.1 – 20.4 novice – expert Three-key 10.63 9.8 50’ Five-Key 3.13 10 2 hrs Multitap Twiddler 2.043 8.2 – 19.8 novice - expert Chord Twiddler 1.476 4.3 - 26.2 novice - expert Table 1 : Summary of systems related to VirHKey.

3. The VirHKey system 3.1 The Hyperbolic Tool VirHKey is based on earlier work on a generic focus and context visualization tool that uses hyperbolic geometry. Further details about this tool can be found in the paper by Chelghoum et al. [19].

Fig. 1 : The hyperbolic pentagrid as a color browser. As seen in Figure 1, the visualization is a kind of fish-eye view into the hyperbolic plane. The two discs in Figure 1 show an example with color data. Both colored discs show the same data. In the rightmost, disc has been focused on the light green color. 3.2 The character layout A predictable layout for the characters was designed by breaking the alphabet in clusters at the vowels: ABCD, EFGH, IJKLMN, OPQRST, UVWXYZ. Each cluster populates a region in the pentagrid as shown in Figure 3 (the leftmost disc). The average number of navigation steps to bring the desired character into the center in this layout is 1.73 (calculated using the English bi-gram frequencies in [5]). By arranging the characters according to their frequency rather than the alphabetic order, the number could be improved to 1.46. While this is a 16% improvement, we kept the alphabetical ordering to improve novice performance. 3.3 Interaction

Method MDITIM EdgeWrite

KSPC 3.06 3.79

WPM 4.9 15.9

Training 5 hrs 50 mns

Users interact with the six central pentagons by selecting which of the five surrounding pentagons to move into the center. The five possible movements are shown in Figure

2. After a character has been centered, the user selects it and the display returns to its initial state. During the interaction, VirHKey displays its state. However, over time users learn the sequences and no longer need to rely on visual search to know where to move next. π/5..3π/5

not needed for optimal use. So, in error free usage, the user first presses one of the five keys, then uses only the keys 1, 2 and 3 to descend to the desired polygon, and then presses the ; key. Figure 3 shows the entry of a “P” (O 1 ;) with the relative mode. In the experiment the key mapping was displayed just like in Figure 3. Of course to use six-key keyboard, the (A, E, I, O and U) and (1, 2, 3, X and ?) must share the same keys.

a stroke

3π/5..π zoomO zoomU zoomI

9π/5..π/5

zoomA zoomE π..7π/5 7π/5..9π/5

Both keyboard modes have the same KSPC of 2.83 (using the frequencies reported in Wikipedia for the French language). This KSPC is higher than that of Quikwriting (nine-key keyboard) and lower than that of the five-key keyboard. The relative mode distributes the consecutive key presses on different keys making it likely to be the faster in expert use. The relative mode seems more complicated because of the changing key mapping, and possibly, difficult to learn.

Fig. 2 : The five strokes of the pen-based VirHKey.

In the absolute mapping, a key is always associated with a navigation direction as shown in Figure 2. Thus, for example, the character ‘Q’ is entered by keying O I ;. In this mode, consecutive presses of the same key are highly probable. They happen in 40% of characters of the French language. In a situation where each finger is associated with only one key, it is faster to press a sequence keys than repeatedly pressing one key. This is why we introduced the relative mode.

There are many possible key mappings and key layouts for one-handed or two-handed text entry. Figure 4 displays some of the designs that we have considered. For one-handed entry, we can use a physical keyboard based on the layout of the tiling (Figure 4a). The five zoom keys are placed around a pentagon while the 1, 2 and 3 keys are placed in a row under the forefinger, the middle finger and the ring finger. This is important to speed up the text entry with the relative key mapping. A large validation key is positioned directly under the thumb. It is also possible mimic the Twiddler design as in Figure 4b. The five zoom keys are fixed on the same side to be accessible with only three fingers. The 1, 2 and 3 keys are placed in the same column under the forefinger, the middle finger and the ring finger to facilitate using the relative key mapping. The validation key is positioned directly under the thumb.

In the relative mode, the role of the keys depends on the state of the VirHKey. For the first key press the selection is done using five keys (A, E, I, O and U) as in the absolute mode. After that, the topology of the pentagrid offers three lower level polygons to proceed to. We call these (1, 2 and 3). The fourth direction back to the previous state (X) and the last possible direction (?) are

For two-handed entry, we envision to separate the keys between the dominant and the non-dominant hands using the relative key mapping. Figure 4c shows a keyboard in which the keys A, E, I, O and U are distributed on the two hands of a right-handed user: A and E for the nondominant hand and I, O and U for the dominant hand. In addition, the 1, 2 and 3 keys are placed under the

In the original stylus interaction the stylus was moved through sequences of the five directions, and a character was entered by lifting the stylus. In keyboard interaction the selection is done using five keys (A, E, I, O and U) and character entry using a sixth key (;). These keys can be used in two ways: absolute and relative.

Key ‘O’ ⎯→ zoomO

Key ‘1’ ⎯→ zoomE

Fig. 3 : Visual feedback for the “P” character in keyboard mode with relative mapping.

(a) U1

(b)

(c)

O2 I 3

A X

I

E ?

A X

(d)

1

O2

U 3

E ?

Fig. 4 : One-hand and two-hands keyboard for the VirHKey in six-key keyboard use. forefinger, the middle finger and the ring finger of the dominant hand. Thus, the non-dominant hand is used only for the first move. To optimize the speed, the validation key could be large and easily reachable with both hands. With this mapping, user presses the same key twice in a row in only 2% of the characters. Figure 4d shows the last design based on the GKOS idea [17]. The keyboard of figure 4c is on the back of a PDA to allow two-handed text entry in a mobile situation.

several blocks of 10 phrases with about 25 characters in each phrase. The 10 phrases were randomly selected from 70 French phrases. Phrases were not repeated within a block but could appear in several blocks during a session. The phrases contained only uppercase alphabet and spaces.

4. Experiment To test the feasibility of the keyboard-based VirHKey, we ran an experiment with the two-handed relative mode. 4.1 Participants and apparatus Due to the high cost in time and effort, longitudinal studies usually involve fewer participants than experiments involving only one or few sessions per participant. Five male students volunteered to participate in the experiment. All were experienced computer users.

a e ?

o u

i 1

2

3

Fig. 6 : The task display. The task display is shown in Figure 6. The participants were asked to transcribe the phrase presented in the window. A prominent “click” was heard when an error occurred. Upon an error participants were asked to continue with the next character. Overall, they were asked to aim for both speed and accuracy. 4.3 Results and Discussion

Fig. 5 : The VirHKey key layout for the experiment. To simulate the keyboards of the Figures 4c and 4d, we distributed the six keys on a standard AZERTY keyboard as shown in Figure 5. The purpose was to simulate a sixkey keyboard, not to propose using VirHKey with a desktop keyboard. Our goal was to explore the feasibility of the relative keyboard mode, not to produce reliable data for performance in mobility or in mobile situation. 4.2 Procedure User performance was recorded in 20 sessions of about 20 minutes. One session consisted of the transcription of

During the test participants entered about 28000 characters. The first character of each phrase was excluded from the following analysis because participants stopped to read the new phrase before entering it. The average text entry rate started from 4.38 wpm and reached 21.31 wpm during the 20th session. The performance over all sessions is shown in Figure 7. We can observe steady improvement in text entry rate throughout the experiment. The mean text entry rate took the usual shape following the power law of learning (y = 4,8733x0,4762, R² = 0.98). Participant 3 was faster than the

other participants without any reason; he had no specific training or skill. He averaged 34.22 wpm on session 20. 35 30

Speed (wpm)

25 20 15 10 1 3 5

5 0 1

3

5

7

2 4 Mean

9 11 13 Session

15

17

19

21

or higher frequency of transcription led to shorter time. However, further analysis by combining KSPC, frequency and sequence of keys is interesting. For characters with KSPC of 2 (A, E, I, O, U) there were no differences in entry time. For characters with KSPC of 3 (B, C, D, F, G, H, J, L, L, P, Q, R, V, W, X), we can note that characters with twice the same key seem faster (J, Q, X). The other characters that use the two hands (non dominant hand first then dominant hand: B, C, D, F, G, H) seem to be faster. For characters with KSPC of 4, characters with twice the same key seem slower (Y, Z) but they are among the less frequent in the French Language. 2.5

Fig. 7 : Text entry rate by participant and by session.

By subtracting twice the number of errors from the entered characters we can estimate the upper bound of text entry rate after correction. Over all participants in session 20 the result was 19.32 wpm. For participant 3 in session 20 the result was 29.76 wpm. 4.0 % 1 3 5

3.5 %

Error Rate

3.0 %

2 4 Mean

2.5 %

Time (s)

An error was recorded when the user entered character that was differed from the corresponding character in the presented phrase. Corrections were not allowed. The average error rate varied from 0.92% (session 7) to 1.59% (session 16). The effect of session on error rate was not statistically significant and we noticed no significant differences between users. Participant 3 that was the fastest did not always have the highest error rate. In session 20, his error rate was 2.64%.

2 1.5 1 0.5 0 A B C D E F G H I J K L MN O P Q R S T U V WX Y Z Character

Fig. 9: Average time per character for sessions 17-20. It is difficult to compare these results to others because the procedures of tests are often different. However, VirHKey text entry rate was close to what has been reported for T9 [11] but lower than with Twiddler [18]. At the end of the 20th session each participant responded to a qualitative questionnaire. The questions included an adapted version of the usability satisfaction scale of the System Usability Scale (SUS) [20]. VirHKey received 71.88. Participants were satisfied with their performance and believed to be able to further improve their skill. However, participants were technology students possibly with positive attitudes towards technology. Thus, the results may not reflect the opinions of all potential users.

2.0 %

5. Enhanced character set

1.5 % 1.0 % 0.5 % 0.0 % 1

3

5

7

9 11 13 15 17 19 21 Session

Fig. 8 : Error rate by session and by participant. Session 21 in Figures 7 and 8 was completed with the VirHKey visualization hidden from the participants. On average, the text entry rate was 17.25 wpm with 2.24% errors. This shows a fall of 19% in speed and 58% increase in the error rate. With above-mentioned error correction estimation, we calculated a 21% fall in text entry rate (15.27 wpm). Participant 3 was an exception, his text entry rate increased to 34.34 wpm while his error rate decreased to 2.37%. We took these results as evidence showing that after some training the relative keyboard mode can be used without visual feedback. The average time spent on each character during sessions 17-20 is shown in Figure 9. Generally, fewer key presses

For real-world writing tasks support for a wider character set needs to be integrated. In a complete keyboard, we have to deal with numbers, punctuation, etc. As in QuikWriting or EdgeWrite, one solution is to add modes through several discs. For example, we could use five discs: two for letters (one for minuscule and one for uppercase), one for numbers and related characters, two for punctuation and accentuated characters (one for minuscule and one for uppercase). Figure 10 shows the different commands to switch between discs: • Ɣ is the interaction for the punctuation keyboard. •

α is the interaction for the letters keyboard.

•

0 is the interaction for the numerical keyboard.

User can switch to minuscule characters (∨) and to uppercase ones (∧). The uppercase mode can also be locked (∧∧). This can be done by a specific interaction or, for example, by doing twice the uppercase interaction. The commands Ɣ, α, 0, ∨, ∧ and ∧∧ can use empty

pentagons of the grid. So that user can generate them has writing letters. Another solution with mobile devices is to use keys or wheels often available on PDAs.

∨ ∧

α1

0

0 Ɣ

α Ɣ

α Ɣ

1

α

0 0

∨

Ɣ

∧ 1

if maj is locked Fig. 10. The different discs for French language.

5. Conclusion and future work We were surprised by how well VirHKey performed in the experiment. Results are close to the gesture version of VirHKey [1]. With the same procedure, the six-key version was, on average, 7% slower in speed with 75% fewer errors. Learning to use the relative keyboard mode is not as big an obstacle as we imagined. To validate these results, we plan to test the VirHKey in mobile use by simulating the keyboard of the Figure 4d. The pentagrid we used is only one of the possible hyperbolic tilings. Error correction facilities and support for a wider character set need now to be integrated and tested. A special strength of the hyperbolic visualization is that the number of tiles is unlimited. VirHKey is not limited to text entry, but can be used as a kind of generic command tool based on icons. Finally, other pointing devices such as joysticks can probably be used making VirHKey as versatile as other multi-device systems [2][3].

Acknowledgements We thank Maurice Margenstern for fruitful discussions, and the participants for their efforts.

References [1] Martin, B. (2005) VirHKey: a VIRtual Hyperbolic KEYboard with gesture interaction and visual feedback

for mobile devices. Proc. MobileHCI’05, ACM Press, 99106. [2] Isokoski, P. and Raisamo, R. (2000) Device independent text input: A rationale and an example. Proc. of AVI2000, ACM Press, 76-83. [3] Isokoski, P. and Raisamo, R. (2000) Device independent text input: A rationale and an example. Proc. of AVI’00, ACM Press, 76-83. [4] Wobbrock, J. O., Myers, B. A. and Rothrock, B. (2006) Few-key Text Entry Revisited: Mnemonic Gestures on Four Keys. Proc. of CHI 2006, ACM Press, 489-492. [5] Soukoreff, R. W., and MacKenzie, I. S. (1995) Theoretical upper and lower bounds on typing speed using a stylus and soft keyboard. Behaviour and Information Technology, 14, 6, 370-379. [6] Wobbrock, J. O., Myers, B. A., and Kembel, J. A. (2003) EdgeWrite: A Stylus-Based Text Entry Method Designed for High Accuracy and Stability of Motion. Proc. of UIST 2003, ACM Press, 61-70. [7] Perlin, K. (1998) Quikwriting: Continuous stylusbased text entry. Proc. of UIST '98, ACM Press, 215-216. [8] Isokoski, P. and Raisamo, R. (2004) Quikwriting as a Multi-device Text Entry Method. Proc. of NordiCHI 2004, ACM Press, 105-108. [9] Poirier F. and Belatar M. (2006). Glyph 2 : une saisie de texte avec deux appuis de touches par caractère Principes et comparaisons. Proc. of IHM 2006, 159-162. [10] Ward, D. J., Blackwell, A. F. and MacKay, D. J. C. (2000) Dasher - a data entry interface using continuous gestures and language models. Proc. of UIST 2000, ACM Press, 129-137. [11] James, C. L. and Reischel, K. M. (2001). Text input for mobile devices: comparing model prediction to actual performance. Proc. of CHI ‘01, 365-371. [12] MacKenzie, I. S. and Soukoreff, R. W. (2002) Text Entry for Mobile Computing: Models and Methods, theory and Practice. HCI, 17, 2/3, 147-198. [13] MacKenzie, I.S. (2002) KSPC (keystrokes per character) as a characteristic of text entry techniques. Proc. Mobile HCI ’02, Springer-Verlag, 195-210. [14] MacKenzie, I.S. (2002) Mobile text entry using three keys. Proc. of NordiCHI ’02, ACM Press, 27-34. [15] Bellman, T. and MacKenzie, I. S. (1998) A probabilistic character layout strategy for mobile text entry. Proc. of GI 1998, 168-176. [16] CyKey, http://www.bellaire.demon.co.uk [17] GKOS, http://www.gkos.com/ [18] Lyons, K., Starner, T, Plaisted, D., Fusia, J., Lyons, A., Drew, A. and Looney, E. W. (2004) Twiddler Typing: One-Handed Chording Text Entry for Mobile Phones, Proc. of CHI 2004, ACM Press, 671-678. [19] Chelghoum, K., Margenstern, M., Martin, B. and Pecci, I. (2004), Cellular automata in the hyperbolic plane: proposal for a new environment. Proc. of ACRI 2004, Springer, 678-687. [20] Brooke, J. SUS - A quick and dirty usability scale. http://www.cee.hw.ac.uk/~ph/sus.html