Building and testing optimized keyboards for specific text entry
Gregory Francis
1
and Claire Oxtoby Purdue University Department of Psychological Sciences 703 Third Street West Lafayette, IN 47907-2004 28 July 2004 Revised: December 14, 2004
Running head: Designing optimal keyboards
1
E-mail:
[email protected]; phone: 765-494-6934. This work was supported in part by the
Army Aeromedical Research Laboratory (Clarence E. Rash) under the auspices of the U.S. Army Research Office Scientific Services Program administered by Battelle (Delivery Order 0429, Contract No. DAAD1902-D-0001). The views, opinions, and/or findings contained in this report are those of the authors and should not be construed as an official Department of the Army position, policy, or decision, unless so designated by other documentation.
Abstract As computers are introduced into ever more devices with new methods of inputting information, there has been interest in how to optimally design the information input system. The study demonstrates that predicted differences in keyboard designs indicate substantial benefits for a keyboard that has been optimized for a specific set of text. The predicted differences are validated by an experiment. Key words: Fitts law, human-computer interaction, keyboard, optimization
Designing optimal keyboards
3
Introduction A key obstacle to the adoption of computers for a variety of new uses involves finding a method for entering text information. Despite improvements in voice and handwriting recognition, the alphabet keyboard remains one of the best technologies for entering a large variety of information accurately and quickly. However, the standard ten-finger keyboard that dominates information input on desktop and laptop computers is not practical for many new situations. Thus, a key issue is how to design alternative keyboards that can be used in these new situations. For example, many U. S. Army military helicopters now include computers that process and display information about flight paths, engine status, weapon targeting, and geographic locations. A keyboard is provided for crew-members to enter much of this information, but because of severe space constraints the standard QWERTY keyboard is not feasible. Figure 1a shows a schematic of the keyboard for the (now-canceled) RAH-66 reconnaissance helicopter. A standard keyboard would be of limited benefit anyhow, as possible positions of the keyboard in the cockpit and other crew member tasks prohibit the use of two hands for entering text. As in the keyboard of the RAH-66, many alternate keyboards have letters arranged alphabetically, which may offer some benefits in terms of foreknowledge of where letters will be located, but probably is not optimized with regard to entering information as quickly as possible. Mavor, Gal, Sawyer and Christ (1987) noted that the design of a keyboard makes a difference for pilots to enter information quickly with minimal interference for other flight tasks. –Figure 1– Another situation familiar to many people is the design of keyboards for entering text information into personal digital assistants and mobile phones. Early designs replicated the QWERTY keyboard commonly used for ten-finger typing, but required the user to press individual letters with a stylus pen. It was soon recognized that the QWERTY keyboard design was not well suited to one-finger typing, and alternative keyboard designs appeared that were optimized for one-finger, or stylus, data entry. Some examples of alternative
Designing optimal keyboards
4
designs include the FITALY (Textware Solutions, 1998); OPTI (MacKenzie & Zhang, 1999), and ATOMIK (Zhai, Hunter, & Smith, 2002) keyboards. The layout of the alphabet keys for the FITALY keyboard is shown in Figure 1b. Many of these new designs were based on optimization strategies. The FITALY keyboard was designed, among other things, to minimize the time required to enter text. It achieved this through consideration of the frequency of using individual letters and the frequency of letter-to-letter transitions. Letters that were commonly paired together in text were placed close to each other. Likewise, the ATOMIK keyboard was created by an optimization algorithm that minimized the time required to move between pairs of letters (using Fitts (1954) law as an estimate of movement time). Zhai et al. (2002) includes an excellent discussion of using optimization techniques for keyboard design. After reviewing a variety of keyboard designs, Zhai et al. suggested that researchers had identified designs that were probably as close to optimized as could be found. However, the benefits of an optimized design necessarily depend on the validity of the optimizing factors. A keyboard optimized with respect to the frequency of letter pairs is valid only if the underlying letter pair frequencies accurately represent the data being entered by users. Previous keyboard designs have been based on a large word corpus (e.g., Mayzner & Tresselt, 1965) that identifies the frequency of words and letter pairs in general use. This corpus may not be representative for the specific use of a keyboard, and a keyboard optimized to a specific text for a specialized situation may be superior to a keyboard optimized for a general corpus. For example, it seems very likely that helicopter pilots, medical doctors, and lawyers, would enter text data that have different frequencies of letter pairs than what are reported in standard word frequency tables. Thus, a keyboard optimized for these specific uses (and many others) might be superior for such users. Francis (2003) and Francis and Rash (2003) created a software program (described below) that can be used to build an optimized keyboard for a specific corpus of text. The next section describes the development of two keyboards that were optimized for two different sets of text. The subsequent section then describes an experimental test of how well human participants were able to enter text into the different keyboards.
Designing optimal keyboards
5
Designing optimized keyboards Francis and Rash (2003) described a software program called KeyboardTool that can create optimized keyboard designs for any specified text corpus. The program is derived from an earlier program called MFDTool that creates optimized multifunction displays (MFDs) (Francis, 1999, 2000, 2003; Francis & Rash, 2002). Data entry keyboards are MFDs with a hierarchy of information that is only one level deep. These programs make it easy to apply and modify the optimization approaches used in the creation of the FITALY and ATOMIK keyboards. The design of an optimized keyboard with KeyboardTool requires four types of information. First, the physical arrangement and size of buttons must be specified. This is done with a graphical user interface in the KeyboardTool program. The physical layout we used is shown in Figure 2. Second, the labels for the keys must be identified. For the keyboards discussed here, the labels include the lower-case letters of the alphabet and the space bar (upper-case letters, numbers, and non-alphanumeric symbols could be treated in a similar way). Third, the time required to move between every pair of buttons must be given. KeyboardTool provides calculations of a variety of movement times, and for the optimizations here the movement times were based on Fitts’ law (Fitts, 1954). Fourth, a corpus of text must be provided. We made two different keyboard designs, based on two different sets of text. One text was a list of Chinese sounding names and the other a list of Russian sounding names (both written in the English alphabet). The texts are shown in Table 1. –Figure 2– –Table 1– For the provided physical arrangement of the keys and labels, KeyboardTool finds the assignment of letters to keys that minimizes the time required to enter the given text. Other constraints can also be imposed on the optimization process. For example, in all of the designs discussed below, the space label was fixed to the large button in the bottom row. The optimization of letters to keys then worked around this constraint. For the calculation
Designing optimal keyboards
6
of movement time, each name was treated as a separate unit. Thus, the movement time within a word was optimized, but there was no effort to optimize from the last letter of one name and the first letter of a subsequent name. Figures 3a and b show the keyboard designs that KeyboardTool reports as optimal for the Chinese names and Russian names, respectively. One can observe the effects of the optimization process by noting how one would have to move a finger (or stylus) from key to key to enter the names. On the Chinese keyboard, the letter pairs for words such as wang, mei, zhao, qiao and qian are all very close to each other. Such placement tends to minimize the predicted time needed to enter these words. Likewise, on the Russian keyboard the letter pairs in words such as rosh, vlad, kozel, leonid, and czenko tend to be very close together. –Figure 3– Of course, it is not surprising that there is no keyboard that makes every letter pair for every word close together. The best that can be done is to find the arrangement across all combinations of words and letter pairs that minimizes movement time. As a result, some words are expected to be entered more slowly than others. For example on the Chinese keyboard the name jun hee is predicted to be entered the fastest (576 milliseconds) and when qian is predicted to be the slowest (1227 milliseconds) among the Chinese names. For the Russian keyboard, the predicted fastest to enter name is fefan onosh (1431 milliseconds), while the predicted slowest is sergey kozel (1911 milliseconds) among the Russian names. All of these predicted times are based on Fitts law, which provides a prediction of how long each movement should take. The sum of the times needed to move to each key in a name generates a prediction of the time needed to enter a given name. The predicted average time to enter a Chinese name on the keyboard optimized for Chinese names is 979 milliseconds. The predicted average time to enter a Russian name on the keyboard optimized for Russian names is 1663 milliseconds. The difference in average entry times reflects the generally longer Russian names in the texts and the specific combinations of letter pairs that determine the effectiveness of the optimization. In general, if many words share common letter pairs, the optimization will be more effective.
Designing optimal keyboards
7
Another way to characterize the effect of the optimization is to compare how long it takes to enter text that was not the basis for the optimization. The model predicts that entering the Russian names on the keyboard optimized for the Chinese names will take, on average, 2042 milliseconds. The model predicts that entering the Chinese names on the keyboard optimized for the Russian names will take, on average, 1255 milliseconds. Thus, there is a predicted average advantage of 276 milliseconds for entering the Chinese names on the Chinese keyboard versus the Russian keyboard. Likewise, there is a predicted average advantage of 379 milliseconds for entering the Russian names on the Russian keyboard versus the Chinese keyboard. This finding suggests that optimization of a keyboard for entering this specific set of text might be advantageous. Overall, the predicted differences between the keyboards are fairly substantial. Users should be substantially faster entering text on the keyboard optimized for that text than on the other keyboard. Of course the optimization is only valid if the underlying model of movement times is an accurate description of users’ behavior. Because the optimization is based on an idealized model of movement times, using Fitts law, the model movement times are more appropriately recognized as minimum movement times rather than predicted values. The minimum movement times would be achieved only by highly practiced users who were very familiar with entering text on these keyboards. Nevertheless, we hypothesized that even novice users might benefit from the close pairing of letter pairs on the optimized keyboards. The next section describes an experiment to test this hypothesis.
Testing the optimized keyboards One of the keyboards in Figure 3 was displayed on a touch screen monitor, along with an additional button labeled Submit and an additional window that displayed information to be entered in the keyboard. Figure 4 shows a screen shot of the windows as they appeared to a participant. A participant used a stylus (the erasure end of an unsharpened number 2 pencil) to tap on the touch screen and enter the requested information. While a touch screen is not the same as a physical keyboard, many keyboards are implemented this way.
Designing optimal keyboards
8
Moreover, to the extent that Fitts law identifies entry time, the movements and targets are the same for a touch screen and a physical keyboard. –Figure 4– To make the task more engaging, participants were told that they were to use the keyboard to enter names into a security checking system. A drawing of a face was shown with each name (faces were randomly assigned to names), again to keep the participant engaged in the task. To gain practice using the stylus, monitor, and keyboard, each participant was first asked to enter letter pairs in the English alphabet (ab, bc, ..., yz, za). These 27 letter pairs were presented in a random order. Since part of the motivation of this practice was for the participant to become familiar with the location of letters on the keyboard, he or she was asked to read the letter pair and identify which keys to strike before making any entries. After entering the letter pair, the participant clicked on the Submit button, which started the next trial. If the participant entered an incorrect letter, the trial was repeated later. After all of the practice trials were completed, the information window displayed a face and name, and the participant was to enter the name as quickly as possible, without making any mistakes. The participant entered both the first (surname) and last (family name) shown with a space in between. Feedback as to when a keypress was made by the participant was provided in text fields above the keyboard. The space bar was entered between the first and last name, and had the effect of shifting the display of feedback from the upper to the lower textfield. After clicking the Submit button, a label below the keyboard indicated whether the entry was valid (matched the given name) or invalid, and the next trial began. If an entry mistake was made, the trial was repeated later in the experiment at a random position among the remaining trials. Twenty participants were recruited from the Purdue University subject pool, and they received course credit for their participation. Ten participants used the keyboard optimized for the Chinese names and the other ten participants used the keyboard optimized for the Russian names. Each participant entered both the Chinese names and the Russian names,
Designing optimal keyboards
9
in a random order, for a total of at least 34 experimental trials (more if a mistake was made and a trial repeated). The entire experiment took approximately half an hour.
Results Entry times were recorded from the moment the first letter of a name was clicked to the moment the last letter of the same name was clicked. Entry times were recorded only on correct trials. Incorrect trials were repeated by the participant later in the experiment and were fairly uncommon. On average, a participant repeated 3.7 trials one time and 0.7 trials were repeated two times. No participant ever repeated a trial more than twice. Pilot work identified that participants tended to have very long pauses after entering the first name and before starting the last name. Presumably, the participant read the first name, entered it, and then read the second name before entering it. To factor out this pause, entry time for a name was computed as the sum of entry times for the first and last names separately. Thus, the time to move from the last letter of the first name to the space key and the time to move from the space key to the first letter of the last name were not included in the entry time. This time surely corresponds to cognitive processing, and while such processing is relevant to the task of entering text on a keyboard, it is not something that would be expected to differ with the keyboard design. Removing this time decreased a source of noise. There are several ways of judging whether the optimization of a keyboard was effective. The first is to compare the observed entry times against the entry times predicted by the model (based on Fitts’ Law). Figures 5a and b plot the average entry time (across participants) for the Chinese and Russian names, respectively. The filled circles correspond to entry on the keyboard optimized for Chinese names. The open symbols correspond to entry on the keyboard optimized for the Russian names. –Figure 5– Although the experimental movement times are nearly an order of magnitude larger than the movement times predicted by Fitts’ law (note the scales of the axes), the experimental
Designing optimal keyboards
10
movement times are well predicted by the model. The scale difference probably reflects the lack of practice the participants had with the stylus and monitor. The correlation between the predicted and measured movement times is r = 0.88, which is highly significant with a two-tailed test (df = 66, p < 0.001). Figure 5 also indicates the effect of the optimization. The entry times for the Chinese names tend to be faster for the Chinese keyboard than for the Russian keyboard (mean difference across all names was 1172 milliseconds). This difference is highly significant with a paired (by name) t-test (df = 16, t = 3.4, p < 0.004). Likewise, entry times for the Russian names tended to be faster for the Russian keyboard than for the Chinese keyboard (mean difference across all names was 4120 milliseconds). This difference was also highly significant with a paired t-test (df = 16, t = 7.23, p = 0.000). Thus, the experimental data support the model predictions, and demonstrate that keyboard optimization for a specific corpus can lead to a significant decrease in entry time, compared to a keyboard optimized for a different text corpus. Learning a new keyboard will surely require a substantial amount of training for maximum use (Anson, George, Galup, Shea & Vetter, 2001). It is noteworthy that even the novice users of the keyboards created for this study show the effects of optimization. We predict that the effects would be even stronger for users with extended practice. For future study of this important issue, it is worth noting that effects of practice can be found in the current study. Figure 6 shows the average time needed to move from one key to another (keypress entry time) for the characters of each trial for each keyboard. The curve on the left indicates the times for the practice trials and the curve on the right shows the times for the experimental trials. –Figure 6– There is tremendous variability in the learning curves. This reflects the many sources that contribute to keypress entry time. In addition to variability across participants, the characters to be entered were randomly assigned to the practice and experimental sessions. Nevertheless, there are clear trends in the learning curves, with keypress entry times grad-
Designing optimal keyboards
11
ually decreasing. Quantitative analysis using a one-tailed t-test on a within subject design verifies these trends. For the Chinese keyboard, participants were, on average, 239 milliseconds slower on the first ten experimental trials than on the last ten experimental trials (df = 9, t = 2.29, p < 0.025). For the Russian keyboard, participants were, on average, 214 milliseconds slower on the first ten experimental trials than on the last ten experimental trials (df = 9, t = 2.10, p < 0.05). Of course, there are other important issues involved in learning to use a new keyboard. For example, learning to use a new keyboard might interfere with a person’s ability to use a standard 10-finger keyboard. In principle, if one could quantify such interference, it might be possible to design an alternative keyboard that minimizes interference and minimizes entry times. Until such a quantitative model is available, though, it seems that such interference is inevitable whenever people have to use different types of keyboards. Under such circumstances, it seems best to have the non-standard keyboard be optimized to reduce entry times.
Conclusions It is quite straightforward to use KeyboardTool (Francis & Rash, 2003) to design an optimal keyboard for any provided corpus of text. Thus, it is now feasible to build specialized keyboards for a variety of different situations. Since a user must learn a new keyboard design whenever the standard 10-finger keyboard cannot be used, it only makes sense to then build a keyboard that best fits the needs of the specific user. We have demonstrated that such optimizations can make a significant and important impact on text entry time. While we think this design approach can immediately be put to use in a variety of situations, it is prudent to consider when and how an optimization will be most effective and when it might encounter some limitations. First, we are not necessarily advocating the use of optimized keyboards over the standard 10-finger keyboard. There are many benefits to standardization and if an environment can allow for standard 10-finger typing, this would likely be the best choice. Nevertheless, there
Designing optimal keyboards
12
are many environments where the standard keyboard is not feasible and it is here that we advocate creating specialized keyboards. If the text corpus is very large and/or non-specialized, the relative frequencies of letter pairs will likely mirror the frequencies in standard word corpuses. In such a situation, the optimization is unlikely to provide substantial benefits beyond other keyboards. It remains an empirical question as to how small or specialized the specific corpus must be to gain a benefit from optimization. It also depends on how big an optimization benefit must be for it to be worth the trouble to create the optimized keyboard. Likewise, one must address how much difficulty there will be for users to learn to use the optimized keyboard. If learning an optimized keyboard is much more difficult than learning a more standard keyboard, the optimized keyboard may not be practical. All of these issues can only be addressed with further empirical study. A similar issue involves identification of the text corpus. Some keyboards are used differently by different people. For example, the co-pilot in an aircraft might enter some types of text, while the pilot might enter other text. If they share the same keyboard, then it should be optimized for both sets of texts. Likewise, if a person has to use physically different keyboards in different situations, it might be prudent to give the keyboards the same character layout and thereby minimize interference effects across keyboards. The text corpus used to create the optimized keyboard deign would then include the text being entered in different circumstances. Another limitation involves the model of time to move between keys. The current study modeled movement time with Fitts law, which correlated very well with the measured movement times in the study. However, if users can enter text with more than one finger, or use a system of tabbing to select letters, Fitts law will not be at all appropriate, and an optimized keyboard based on Fitts law will not really be optimal at all. KeyboardTool can generate models that estimate movement times for a tabbing type of system (and several other possibilities as well), but it cannot yet deal with two- (or more) finger typing. Twofinger typing is more complicated to model because the time to strike a pair of keys depends on the positions of the fingers for the previous two (or more) keys. In principle though, if
Designing optimal keyboards
13
a model of two-finger typing can be quantified, the approach taken here can be modified to use such a model to build an optimized keyboard. The approach used by Shieh and Lin (1999) to model ten-finger typing on Dvorak and QWERTY keyboards may be worth further development. Unlike some other optimization approaches (Zhai et al., 2002), KeyboardTool does not try to optimize the physical location of the keys. Instead, the physical location and size of the keys must be provided to the program, and it then optimally assigns letters to the different keys. Thus, an optimized arrangement for one physical keyboard might be surpassed by a different physical keyboard. This limitation is not overly restrictive, however, since constraints on hardware designs are often imposed by manufacturing details. Moreover, with a valid model of text entry time one can produce optimized key layouts for many different physical designs and then choose the best. Finally, there may be concerns that a specialized keyboard will be difficult to use, and that any predicted benefits in entry time will not be worth the time needed to learn how to use the keyboard. This, of course, is an empirical issue that needs to be measured with further study. Nevertheless, two important comments can be raised. The first comment is that in many new keyboard applications, some type of new keyboard must be learned or the product will not exist. In such a case, an optimized keyboard is surely better than a nonoptimized keyboard. The second comment is that one can add additional constraints to the optimization process, so that the resulting keyboard will seem somewhat familiar to a new user. For example, one could add a constraint that had letter pairs on the new keyboard be close together if they were also close together on the standard 10-finger keyboard. Because of the different physical layouts, this constraint likely cannot be perfectly satisfied, and it must be balanced against the goal of minimizing entry time for the text corpus. KeyboardTool allows a designer to weight the important of these different kinds of constraints. Thus, despite the limitations of the approach, we feel that optimization of keyboards for a specific text corpus is likely to prove useful in many different situations. We envision, for example, the development of PDAs and similar devices that have keyboards specialized for various military uses, medical doctors, and inventory entry systems. Such systems may
Designing optimal keyboards dramatically speed the entry of text with a variety of modern computing systems.
14
Designing optimal keyboards
15
References Anson, D., George, S., Galup, R., Shea, B. & Vetter, R. (2001). Efficiency of the Chubon versus the QWERTY keyboard.. Assistive Technology, 13, 40-45. Fitts, P. M. (1954). The information capacity of the human motor system in controlling the amplitude of movement. Journal of Experimental Psychology, 47, 381–391. Francis, G. (1999). A software tool for the design of multifunction displays. (USAARL Report No. 99-20). U.S. Army Aeromedical Research Laboratory: Fort Rucker, AL. Francis, G. (2000). Designing multifunction displays: An optimization approach. International Journal of Cognitive Ergonomics, 4, 107–124. Francis, G. (2003). MFDTool: A software program for designing optimal multifunction displays. Behavior, Research Methods, Instruments & Computers, 35, 236–243. Francis, G., & Rash, C. E. (2002). MFDTool (Version 1.3): A software tool for optimizing hierarchical information on multifunction displays. (USAARL Report, No. 2002-22.) U.S. Army Aeromedical Research Laboratory: Fort Rucker, AL. Francis, G., & Rash, C. (2003). Optimization of keyboard design for specialized text entry. Proceedings of the 2003 Human Factors and Ergonomics Society, Vol 47, pp. 734–736. MacKenzie, I. S. & Zhang, S. X. (1999). The design and evaluation of a high-performance soft keyboard. Proceedings of the CHI 99 Conference on Human Factors in Computing Systems. Mavor, A. S., Gal, C. A., Sawyer, C. R., & Christ, K. A. (1987). A comparison of keyboard designs for cockpit applications. (Technical memorandum 24-878) U.S. Army Human Engineering Laboratory: Aberdeen Proving Ground, MD. Mayzner, M. S. & Tresselt, M. E. (1965). Tables of single-letter and digram frequency counts for various word-length and letter-position combinations. Psychonomic Monograph Supplements, 1, 13–32.
Designing optimal keyboards
16
Textware Solutions. (1998). The Fitaly one-finger keyboard. http://fitaly.com/fitaly/fitaly.htm. Reardon, M. J.& Francis, G. (1999). Reducing the risk of aviator-multifunction display interface problems with human factor models and optimization design methods. SAFE Journal, 29, 100–106. Shieh, K. & Lin, C. (1999). A quantitative model for designing keyboard layout. Perceptual and Motor Skills, 88, 113-125. Zhai, S., Hunter, M. & Smith, B. A. (2002). Performance optimization of virtual keyboards. Human Computer Interaction, 17, 89–129.
Designing optimal keyboards
17 Table Captions
Table 1. The two text corpuses used in the study.
Designing optimal keyboards
18
Chinese names
Russian names
wang zhou
bashka star
ya huang
roman tinov
yu munchi
tysha urman
chu mei
sergey kozel
jing chi
elena minsk
yung shek
vlad iosif
bing kui
olga krysk
tao fong
alexi lenko
qiao mu
malga pansof
zhi chen
dmitry isov
ping yu
fefan onosh
hong wang
ania kshtov
tsai miao
dorofi rozn
jun hee
unka zaval
when qian
czenko tumak
zhao ming
leonid rosh
rong ah
albast zary Table 1:
Designing optimal keyboards
19 Figure Captions
Figure 1. Two examples of keyboards that differ from the standard 10-finger design. (a) A schematic of the keyboard in the RAH-66 helicopter. (b) A schematic of the FITALY keyboard designed by Textware Solutions for Palm Pilot computers. Figure 2. The physical layout of the keyboard used in the current study. The space key was restricted to be on the large bar on the bottom row. All other letters could be assigned to any of the other keys. Figure 3. The optimal assignments of letters to keys for the two text corpuses. (a) The optimal keyboard for the Chines names. (b) The optimal keyboard for the Russian names. Figure 4. The two windows used in the experiment. The large window on the right contained the keyboard, a Submit button, and two text fields to provide feedback on the participant’s text entries. In this case the Chinese keyboard is being used to enter a Russian name. The small window on the left listed the information the participant was to key into the keyboard. The faces made the task somewhat more engaging. A label at the bottom of the keyboard window provided feedback and turned yellow if an entry mistake was made. Figure 5. Scatter plots comparing the actual entry time of each name and keyboard combination (averaged across participants) against the predicted entry time. (a) As predicted, entry of the Chinese names is faster on the keyboard optimized for the Chinese names. (b) As predicted, entry of the Russian names is faster on the keyboard optimized for the Russian names. Figure 6. Learning curves for the two keyboards. Each plot shows the average time needed to move from one key to another on each trial. Both keyboards show substantial learning effects across the experimental session. (a) Results for the keyboard optimized for the Chinese names. (b) Results for the keyboard optimized for the Russian names.
(a)
RCL INS
ENTER
DEL
(b)
A
B
C
D
E
F
G
H
I
J
K
L
1
2
3
M
N
O
P
4
5
6
Q
R
S
T
7
8
9
U
V
W
X
-
0
.
/
Y
Z
CLR ALL
Z
V
C
H
W
K
F
I
T
A
L
Y
N
E
G
D
O
R
S
B
Q
J
U
M
P
X
Figure 1:
Figure 2:
(a)
(b)
Figure 3:
Figure 4:
Average actual entry time (ms)
(a)
13000 Chinese names
12000 11000 10000 9000 8000 7000
Chinese keyboard Russian keyboard
6000 5000
600
800
1000 1200 1400 1600 Predicted entry time (ms)
1800
2000
(b) Average actual entry time (ms)
24000
Russian names
22000 20000 18000 16000 14000 12000 Chinese keyboard Russian keyboard
10000 8000 1400
1600
1800 2000 2200 2400 Predicted entry time (ms)
Figure 5:
2600
(a)
3000
Average keypress time (ms)
Chinese keyboard 2500
2000
1500
1000
500
(b)
0
10
20
30 Trial
40
50
60
3000
Average keypress time (ms)
Russian keyboard 2500
2000
1500
1000
500 0
10
20
30 Trial
Figure 6:
40
50
60
Biographies Gregory Francis (Purdue University) received his PhD in Cognitive & Neural Systems from Boston University in 1993. Claire Oxtoby (Purdue University) received her BS in Psychology from Purdue University in 2004.
