Entering Text with A Four-Button Device
Kumiko Tanaka-Ishii and Yusuke Inutsuka and Masato Takeichi
The University of Tokyo
7-3-1 Bunkyoku Hongo, Japan
fkumiko, inu, takeichig@ipl.t.u-tokyo.ac.jp
Abstract
This paper presents the design of a text-entry device
that requires only four buttons. Such a device is ap-
plicable as the text interface of portable machines and
as an interface for disabled people. The text-entry
system is predictive;; the basis for this is an adaptive
language model. Our evaluation showed that the sys-
tem is at least as ecient for the entry of free text as
the text-entry systems of current-generation mobile
phones. The system requires fewer keystrokes than a
full keyboard. After adaptation, one user reached a
maximum speed of 23 wpm.
1 Introduction
Electronic machinery is becoming smaller;; recent de-
velopmentsin palmtop andmobile-phone technologies
oer dramatic examples of this process. Since smaller
machines are more portable, their users have freer ac-
cess to information. Here, however, the user interface
is a major issue.
If a machine is being used as a medium for person-
to-person communications, a natural interface might
be speech-based. Forother tasks, however,like brows-
ing through Internet pages or looking up databases,
the most natural tool for control and data entry is
still the keyboard.
Mobile machines oer little surface space, so only
a few buttons are available for the entry of text. The
most representativeformistheuseof10keys for text
entry on mobile phones. However, even smaller ma-
chines continue to appear, suchaswatch-sized com-
puters. It might not be possible to equip such ma-
chines with more than four or ve buttons. Ques-
tions then arise. Is it possible to enter text with a
small number of buttons? What about four buttons?
How ecient can wemake this?
Other potential applications for text entry with
four buttons include the control panels of oce ma-
chines and household machines. Although these ma-
chines increasingly contain functions that allow ac-
cess to the Internet, sucient surface space for a full
keyboard is often not available. Another potential
application is in text-entry interfaces for elderly and
disabled people. A report (of Advanced Design of
Integrated Information Society, 2000) indicates that
keyboard operation is the highest hurdletotheuseof
computers by the aged. The situation is even worse
for people with hand-related disabilities. A text-entry
device with four largebuttons might facilitate human-
machine communications bysuch people.
The idea of decreasing the number of keys on
the keyboard in itself is not new. The oldest re-
alization of this idea is the stenotype keyboards.
With the recent popularity of mobile machines, re-
searchers have become increasingly interested in one-
handed keyboards(Mathias et al., 1996). Most of the
work to date in this eld has been related to mo-
bile phones. Text entry on current-generation devices
remains cumbersome, so innovative companies(Tegic
9, 2000)(ZI-Corp., 2000)(Slangsoft, 2000) have pro-
posed predictive methods for the more eciententry
of text by implementing a method that had rst been
proposed some years earlier(Rau and Skiena, 1994).
The results of several studies have veried its e-
ciency(James and Reischel, 2001)(Tanaka-Ishii et al.,
2000), so the technology looks promising in the con-
text of mobile phones.
Our study goes further in decreasing the number of
buttons than the above-cited studies. In our study,
we tried various text-entry methods and found the
predictive method to be the best. As far as we know,
no other study that includes the application of a lan-
guage model has yet been carried out in this context;;
neither has the eciency of this approach been ex-
amined. Additionally, the major contribution is our
study of the potential power of a language model by
examining its actual eciency on a device with few
buttons.
In the next section, we rst show how text is en-
tered via our TouchMeKey4 keypad.
2 An Example
Figure 1-1 shows the GUI for the TouchMeKey4 key-
pad. Nine buttons are visible, with four on either side
of the central boxes plus a `quit' button on the right-
hand side. In this paper, we only count those buttons
that are only used for the entry of characters that
is, the four on the right-hand side. We also impose
the constraint that the buttons may only be pressed
one at a time, because the inclusion of key-chords in-
creases the actual number of buttons by including the
combinations of keys.
Six or seven letters of the alphabet are assigned to
each of the buttons. The no. 1key has `abcdef', the
no. 2 key has `ghijkl', the no. 3key has `mnopqrs',
and the no. 4 key has `tuvwxyz'. The small letters
Figure 1: Entering the word \technology" with the
TouchMeKey4 keypad
are assigned to the same keys as the corresponding
capital letters. All other ASCII characters other than
the alphanumeric characters are assigned to the no.
4key.
Suppose that wehavejustentered the string `hu-
man language'. The text appears in the upper box
in the middle of the window (the upper text-box in
Figure 1-1). We now wish to enter the word `tech-
nology'. Words are entered through a single-tap-
per-character form of predictiveentry;; a key is only
pressed once to enter a character. For example, the
no. 4 button is pressed once to enter the `t' of `tech-
nology'. Toenter the subsequent `e', the no. 1 button
is pressed once.
After the no. 1 button has been pressed, the Touch-
MeKey4 windowisasshown in Figure 1-2
1
. Here, we
see two dierences from Figure 1-1. The rst is that
`41' appears in the box in the middle of the window.
This indicates the string that the user has just en-
tered. The second change is that some words have ap-
peared in the lower box in the middle of the window
(a list-box that we call the `candidate-box'). These
words are the candidate words that correspond to the
user's input, `41'.
Eachpressof abutton bythe usermakesthe Touch-
MeKey4 system automatically search the dictionary
for candidates. The candidates include longer words
as well as, if such words exist, words of the same
length as the entered sequence of digits. The can-
didates are thus all words that begin with one letter
from `tuvwxyz' followed by one letter from `abcdef'.
For example, `text', `was', and `vendors' are candi-
dates, as is the two-character candidate `we'.
The numerous candidates are sorted into an order
before they areplaced in the candidate box and shown
to the user. The order is according to word probabil-
ity as determined on the basis of PPM (prediction
by partial match), which has been proposed in the
information-theory domain. A detailed description is
given in x4, but we summarize the method's essence
here as part of our explanation of Figure 1. The rele-
vance of each candidate is measured by statistics from
two sources.
Base dictionary the unigram statistics collected
fromahuge corpus of newspaper data, and
User corpus the ngram statistics obtained from a
small personal document supplied by the user.
In this example, the Base dictionary is constructed
from one year of issues of the Wall Street Journal
(WSJ) that contains 93 000 dierent words and the
User Corpus is a computer magazine that contains 10
000 words. The particular User corpus is the reason
for the appearance of the relatively uncommon word
`vendors' among the top ve candidates (Figure 1-2).
Our target `technology' appears as the second-
ranked candidate. In selecting this word, the user
highlights it by using the down button on the left-
1
Note that the most recently pressed button is framed bya
thick line.
hand side of the window(Figure 1-3) and then presses
the enter button (Figure 1-4). We see that the se-
lected candidate now appears in the upper text-box
2
.
In describing our realization of the TouchMeKey4
system outlined above, the following four questions
are discussed in the remainder of this paper:
Interface Is some method other than that described
above suitable for text entry with a four-button
device?
Candidate Estimation How can the system esti-
mate the relevance of each candidate?
Key Assignment How should characters be as-
signed to the individual buttons?
Number Of Keys What is the minimum numberof
keys required? Is the entry of free text with only
two buttons reasonably ecient?
3 Interface
Various methods for the entry of text via a four-
button device are conceivable. The biggest choice is
whether or not to adopt a predictive method.
3.1 Non-PredictiveEntry Methods
Let's start by considering the case where we don't
adopt prediction. This means that we need to enable
the exact entry of the individual characters via the
four buttons. One method of this type involves as-
signing an order to the characters on eachkey;;akey
is then pressed i times to obtain the i-th character
(we call this the multi-tap method). This method is
commonly applied on mobile phones.
However, there are two problems with this method.
Firstly, the user often needs to press a key numerous
times to obtain a single target character. Secondly,
there is an ambiguity in the user action when two
characters assigned to the same button are to be en-
tered one after another ('aa' requires the entry of '11'
that can also be 'b'). This situation requires the use
of an escape.
A second possible method is to press a rst button
to select it, and then enter the number i to select the
ith character which is assigned to the rst button.
For example, on many mobile phones, `o' is obtained
by pressing the no. 6 key and then the no. 3 key,
since `o' is the third letter on the no. 6key. However,
if the number of letters on each key is greater than
the number of keys, entry of the higher i values is
implausiblydicult. With theTouchMeKey4system,
for example, a system for the easy entry of fth and
sixth characters, etc., is not possible.
In short, the free entry of text turns out to be too
dicult with a four-button device unless we adopt
2
As with any system where a predictive method is applied,
the weak pointofTouchMeKey4 is the processing of unknown
words which do not appear in the dictionary. Therefore, it
is important that the Base dictionary contains a richvocabu-
lary. When, however, an unknown word occurs, it may still be
entered character bycharacter by using the methods described
inx3.1, or the system may be connected with a larger dictionary
via a network.
Table 1: Data used in this work
name WSJ ZIFF JA
usage base user user
dictionary corpus corpus
domain newspaper computer scientic
magazine paper
Total no. wrds 6.6 6.2 9.0
(million wrds)
No. di. wrds 93 99 173
(thousand wrds)
Wrds in common 100 40 18
with Base
Dictionary (%)
Wrds: Avr. len. 4.45 4.71 4.41
(L
avr
)
Test document - 1900 1826
(no. wrds)
No. di. wrds - 868 761
in test doc.
prediction. This is so even for the case of English, the
written form of which has relatively few characters,
andiseven more so for languages with large numbers
of characters such as Chinese, Japanese, or Thai (78
characters). We are thus obliged to use prediction.
3.2 Predictive Text entry
Generally, there are twoways to predict candidates.
The rst is the single-tap method. The earliest
appearance of this idea was at the beginning of the
80's in Japan, in discussions of processing systems
for Japanese text(Co.Ltd., 1982);; more recent work
has been concerned with mobile phones (James and
Reischel, 2001)(Tanaka-Ishii et al., 2000).
The second wayisprediction by prefix. Given a
user input, the system searches for words with the
corresponding prefix.
This method of collecting candidates to be oered
to the user has been particularly successful in the en-
try of Chinesetext. The method has alsobeen applied
to certain text-entry systems in the man-machine in-
terface domain, too (Masui, 1999).
As the description of x2 indicates, the combination
of the two methods is adopted in our TouchMeKey4
system. It thus needs to process many candidates
for a single user entry. The mechanism of estimating
levels of relevance for the words is explained in the
next section.
4 Applying an Adaptive Language
Model in Candidate Estimation
As was summarized in x2, the PPM (prediction by
partial match) framework is used by TouchMeKey4
to estimate the relevance of candidates. Its charac-
teristic is that the word distribution is adapted to the
style of the user's corpus.
PPM was originally proposed as an adaptive lan-
guage model for use in improving the compression
rates of arithmetic coding. The estimation of prob-
abilities by PPM thus guarantees a lowering of the
entropy of the language model. PPM has successfully
been adapted to the user-interface domain in certain
previous works(Tanaka-Ishii et al., 2001)(Ward et al.,
2000).
Broadly, PPM interpolates the n-gram counts in
the user corpus and the statistics in the base dictio-
nary. The following formula is used to estimate a
probability for the ith word w
i
, P(w
i
):
P(w
i
)=
kmax
X
k=;1
u
k
P
k
(w
i
) (1)
Here, k, the order, indicates the number of words be-
fore w
i
that are used in the calculation of P
k
(w
i
).
For example, P
2
(w
i
) is estimated on the basis of the
occurrence of w
i;1
and w
i;2
. P
k
(w
i
) is calculated as:
P
k
(w
i
)=
c
k
(w
i
)
C
k
(2)
where C
k
is the frequency of the order k as a context,
and c
k
(w
i
) is the frequency with which w
i
occurs in
that context. P
k
(w
i
) when k = -1 describes a base
probability that is obtained from the base dictionary.
For other k, P
k
(w
i
) is calculated from statistics ob-
tained from User corpus. Finally, u
k
is a weighting
probabilitythatismultiplied to P
k
(w
i
)toobtainthe
nal P(w
i
). Of the manystudies of u
k
(Teahan, 2000),
we have chosen PPM-A(Bell et al., 1990), the sim-
plest, because our preliminary experiments showed no
signicant dierence in performance among the meth-
ods we tried.
We decided to utilize this PPM framework because
the context is the most suitable item of information
for the elimination of irrelevant candidates. Small
machines are in a personal context, and oce and
household machines are used in particular contexts.
With this method, the language model is adaptable
on the y. This is achieved by simply accumulating
the user's newly entered text at the end of the user
corpus.
In this paper, the Base dictionary contains the uni-
gram probabilities obtained from Wall Street Journal
as was explained in x2. We prepared various User
corpora,: three in English, three in Japanese and two
in Thai. Of these, the characteristics of two of the
English User corpora that are used in x6aregiven in
Table 1.
5 Key Assignment
The assignment of characters to the respective but-
tons is one determinant of the eciency of text en-
try. For example, if all characters from `a' to `w' are
assigned to the rst key and `x', `y', and `z' are re-
spectively assigned to the second, third and fourth
keys, the performance in word prediction will clearly
be bad. The problem of key assignment remains even
when we have eliminated such extreme possibilities,
Table 2: Key assignments and entropy
N
K
Lab. Groups of characters Entropy
10 - (S
0
)(S
1
)(abc)(def)(ghi)(jkl) 0.73
(mno)(pqrs)(tuv)(wxyz)
5 A (abcdef) (ghijkl) 1.09
(mnoS
0
)(pqrsS
1
) (tuvwxyz) 1.09
5 B (S
0
S
1
) (abcdef) (ghijkl) 1.40
(mnopqrs) (tuvwxyz)
5 C (S
1
mno) (abcpqrs) 1.14
(deftuv) (ghiwxyz) (jklS
0
)
4 A (abcdef) (ghijkl) 1.40
(mnopqrs) (tuvwxyzS
0
S
1
)
4 B (S
1
abc) (defghi) 1.56
(jklmnopqrs) (tuvwxyzS
1
)
4 C (S
0
jkl) (abcmno) 1.76
(defpqrstuv) (ghiqxyzS
0
)
3 A (S
1
abcdef) (ghijklmno) 1.95
(pqrstuvwxyS
0
)
3 B (S
1
ghipqrs) (abcjkltuv) 1.91
(defmonqxyzS
0
)
3 C (S
1
jklpqrstuv) (abcdefmno) 2.13
(ghiwxyzS
0
)
2 (S
0
abcdefghijkl) 3.28
(mnopqrstuvwxyzS
0
)
because there are many plausible assignments. We
thus need to be able to measure the performance of a
key assignment.
One way to measure this is to experimentally de-
cide it by automatically entering some documents (as
will be described in the x6 later in this paper). How-
ever, the result of such a test is dependent on the test
document which is used. Lower-level settings, such
as key assignments, should, as much as is possible, be
for general-purpose use.
Having key sequences as C and the target word
as W, the task of the system is to estimate a better
W from C. Information theory provides us with a
tool for estimating the uncertaintyofthis task: the
average conditional entropy. The denition of this
quantity, H(WjC), is given by:
H(WjC) (3)
= 
w;;c
P(C = c)H(W = wjC = c)
= ;
w;;c
P(C = c;;W = w)logP(W = wjC = c)
where P(C = c) is the probability of the input se-
quence c and P(W = wjC = c) is the conditional
probability of words for the given c. When the es-
timation of W is less certain, H(WjC) has a larger
value. The lower the entropy, the less uncertain the
estimation of the word. Therefore, the conditional
entropy is suitable as a method for the evaluation of
key assignments.
One other factor that we need to consider at this
point is the order of the alphabet. English has an
alphabet order that even children know. If this order
is neglected and the letters `ajxgukh' are assigned to
a given key, the interface will become dicult for the
beginners, although it might be the most ecientfor
a well-trained user. Therefore, the key assignments
had better reect such linguistic tradition.
We took this into consideration in generating some
possible key assignments. Table 2 is a list of the as-
signments and their values of conditional entropyas
calculated on the basis of one year of issues of WSJ.
The rst column shows the total number of keys (be-
low denoted by N
K
). We here consider the situa-
tions where there are ve, three, and two, as well as
four, buttons. The second column gives a label for
each of the key assignments. In the third column,
the characters to be assigned to the respective but-
tons are grouped in parentheses. For example, 4-A
indicates an assignment to four keys with 'abcdef' as-
signed to the rst button, 'ghijkl' to the second but-
ton, 'mnopqrs' to the third, and 'tuvwxyz' and other
ASCII symbols to the fourth. The capital letters are
assigned to the same keys as the corresponding small
letters. S
0
and S
1
indicates the non-alphabetic ASCII
symbols
3
. Note that 4-A corresponds to the Touch-
MeKey4 assignment whichwesaw in Figure 1-1. The
groupings with the label C are more random than
those with other two.
In general, entropyvalues fall as the number ofkeys
increases. This is a readily comprehensible result;; a
larger number of keys eases the task of estimation,
thus making it less uncertain. When we compare the
values for assignments to the same numbers of keys,
we see that the entropyvalues dier considerably. For
example, the entropy of 5-B indicates more uncer-
tainty than the other 5-x assignments. The entropy
value is the same as for 4-A, although the number of
keys in use is dierent (this is comprehensible when
we look at the similar character groupings of 4-A and
5-B).
In this paper, we evaluate the use of key assign-
ments with the label A on TouchMeKey4, since they
havelower entropyvalues than the other settings.
6 Evaluation
6.1 Number of Keystrokes
We attached an automatic text-entry routine
to TouchMeKey4 and measured the numbers of
keystrokes that are needed per word. The number
of keystrokes is the sum of the keystrokes required
for the input and selection operations. Keystrokes
for selection are counted to be n when choosing the
nth-candidate.
In the prediction of words, there are multiple points
wherethe targetwordmaybechosen. For example, in
Figure 1, the word`technology'appears asthe second-
best choice after the user has typed in `41'. The user
may select the target at this pointortype in another
`1' to indicate the `c' and increase the target's rank.
The automatic routine only chooses the target after
it has appeared as the best candidate;; otherwise, the
3
These symbols are categorized into two groups according
to the categorization used on mobile phones.
4
6
8
10
12
14
16
18
20
0 5000 10000 15000 2000
Avr. No. of keystrokes / word
learning data size(words)
"baseline"
"2keys"
"3keys"
"4keys"
"5keys"
"10keys"
Figure 2: No. keystrokes with learning of ZIFF
4
6
8
10
12
14
16
0 5000 10000 15000 2000
Avr. No. of keystrokes / word
learning data size(words)
"baseline"
"2keys"
"3keys"
"4keys"
"5keys"
"10keys"
Figure 3: No. keystrokes with learning of JA
routine continues to enter the word. When the full
length of the word has been entered, the target word
of the current ranking is chosen.
Figure 2 and 3 show the relation between the
amount of learning data (horizontal axis) and the
number of keystrokes per word (vertical). The test
document is indicated in Table 1 (7th row). Data of
the same kind but from dierent positions in the test
data are used for the learning data and the test data.
The respective lines indicate learning when N
K
is
10, 5, 4, 3, and 2. The larger the N
K
, the lower
the line. The horizontal solid line (around 5.5 taps
per word) indicates the baseline, the average num-
ber of keystrokes needed to process a single word on
a full keyboard. This is calculated as L
avr
(in Ta-
ble 1)+1(for space). Note that TouchMeKey4 auto-
matically enters the space.
When there is no learning data, TouchMeKey4
needs far more keystrokes than the baseline. How-
ever, after the learning of ten thousand words, the
number of keystrokes goes below the baseline when
N
K
> 4.
In order to see the results at macroscopic scale, Ta-
ble 3 shows the results after the learning of 50 thou-
sand words. The values indicate per-word keystrokes,
and the percentages in the parentheses show the ra-
Table 3: No. keystrokes per word with learning of 50
000 words of the user corpus
total number ZIFF JA
of keys (N
K
)
10 4.36(16.7%) 4.15(16.1%)
5 4.83(18.9%) 4.84(13.2%)
4 5.19(22.4%) 5.06(17.8%)
3 5.97(29.5%) 6.08(19.1%)
2 11.87(35.4%) 10.63(28.0%)
tios bywhichthenumbers of keystrokes decrease as
compared with the case of no learning of a user doc-
ument. We see that the numbers of keystrokes are
reduced by about 30% for both ZIFF and JA. When
N
K
=4,thevalue falls to around 5.1. Since L
avr
+1
is around 5.5, TouchMeKey4 provides a reduction in
numbers of keystrokes of almost 9 % as compared
with a full keyboard. Supercially, this looks like a
small gain. However, it is surprising that, even with 4
buttons, text maybeentered with fewer keystrokes
than with a full keyboard.
When N
K
= 3, on the other hand, the number of
keystrokes remains at around 6.0 per word. There-
fore, when N
K
=3, the system requires a larger num-
ber of keystrokes than the baseline.
TouchMeKey4 also runs in Japanese and Thai, so
we executed analogous experiments with those lan-
guages. We obtained very similar graphs in these
cases. To resume, here are our observations across
three languages:
 Learning is indispensable for systems with small
N
K
values to perform better than the baseline.
 However, a large amount of learning data is
not necessary (text with ten thousand words is
enough).
 When N
K
= 3, the number of keystrokes does
not fall below the baseline.
6.2 Speed
Eightsubjectswerehired totest TouchMeKey4: three
in English, three in Japanese, and twoinThai.Two
of the subjects for the English are nativespeakers of
Japanese. The other subjects were the nativespeak-
ers of the languages in the respective tests.
The subjects were told to do 10 sessions of test-
ing. Each session is 30 minutes long;; the subject was
told to continue to enter the given text as quickly as
was possible and without pausing during eachofthe
sessions. The vocabulary of the given text is solely
from the learned user corpus. TouchMeKey4 learned
a 10-thousand-word user corpus before it was handed
to the subjects. For the text entry, they were given
hardware controllers that work with TouchMeKey4.
Figure 4 gives the results on speed. The horizontal
axis describes the sessions and the vertical axis shows
the average numbers of words per minute (wpm) in
each session. The respective lines indicate the speed
4
6
8
10
12
14
16
18
20
22
24
1 2 3 4 5 6 7 8 9 10
speed (wpm)
session
"English-1"
"English-2"
"English-3"
"Japanese-1"
"Japanese-2"
"Japanese-3"
"Thai-1"
"Thai-2"
Figure 4: Speed
of the subjects over time. After 5 hours training, en-
try by each of the subjects was at some rate above
12 wpm. The speed of entry by the multi-tapping
method on a mobile phone is in the range from 5 to
10 wpm(James and Reischel, 2001), so TouchMeKey4
obviously allows higher rates of text entry. Further-
more, the speeds are comparable to those obtained
with the single-tapping method on mobile phones (7
to 25 wpm(James and Reischel, 2001)). One subject
set the record, reaching 23 wpm.
This speed is comparable to that of an expert with
the single-tapping method. Predictive text entry thus
prevented deterioration of performance, despite the
number of buttons being decreased from 10 to 4.
With regard to human learning, the more highly
the subject was trained, the faster he or she became.
The speeds of some subjects who had had diculties
at the beginning of the tests had doubled by the end.
Language-by-language comparison reveals that
Japanese text entry was fastest. Although the en-
tropy value for Japanese is by far greater than the
values for Thai and English (4.05 for Japanese and
1.13 for Thai), the Japanese subjects managed well
with TouchMeKey4 because they are accustomed to
the use of predictive text entry in kana-kanji conver-
sion.
Wemust admit that TouchMeKey4 places a heavier
cognitive load on users than does text entry via a
full keyboard (40 to 60 wpm) or a stylus and virtual
keyboard (32.5 wpm(Zhai et al., 2000)). However, we
regard the speed as satisfactory in comparison with
those achieved by using single-tap-by-character entry
systems on mobile phones.
7 Conclusion
Wehave presented TouchMeKey4, a text entry device
that requires only four buttons, and aspects of its de-
sign and testing. Several characters of the alphabet
are assigned to each of four buttons and the user en-
ters a document with the aid of a predictive text-entry
system. The device is realized by software that esti-
mates the word that most probably corresponds with
the user input. The estimate is based on an adaptive-
statistical language model. Firstly, a base model is
constructed from a large body of text from newspa-
pers;; the model is then adapted to the local context
by using a smaller user corpus of 10 thousand words.
Our evaluation shows that text entry with this sys-
tem is as ecient as text entry on a mobile phone.
The number of keystrokes is reducible to a number
below that required on a full keyboard. The average
speed of our test subjects was 14 words per minute,
and the fastest subject recorded 23 wpm. We also dis-
cussed the possibilityofentering text via even fewer
keys, e.g., three keys. However, the usability of a
three-key device turned out to be questionable, be-
cause the number of keys needed to enter text be-
comes large.
Our future direction will be to investigate the ac-
tual application of the system on smaller machines
and to develop a text-entry system for use by elderly
and disabled people.

References

T.C. Bell, T.G. Cleary, and T.H. Witten. 1990. Text
Compression. Prentice Hall.

Toshiba Co.Ltd. 1982. Japanese entry system. In
Japanese Patent 57-185528 (in Japanese).

C. James and K. Reischel. 2001. Text input for mo-
bile devices: Comparing model prediction to actual
performance. In Proceedings of the Annual SIGCHI
Conference.

T. Masui. 1999. Po box an ecient text input
method for handheld and ubiquitous computers. In
the ACM Symposium on User Interace Software
and Technology, pages 113{119.

E. Mathias, I.S. MacKenzie, and W. Buxton. 1996.
One-handed touch typing on a qwerty keyboard.
In Human Computer Interaction,volume 11, pages
113{155.

Research Association of Advanced Design of
Integrated Information Society. 2000.
http://www.pref.toyama.jp/sections/
1012/matsplan/kadai/k10in.

H. Rau and S. Skiena. 1994. Dialing for docu-
ments:an experiment in information theory. In
UIST.

Slangsoft. 2000. Slangsoft home page.
http://www.slangsoft.com.

K. Tanaka-Ishii, Y. Inutsuka, and M. Takeichi. 2000.
Japanese input system with digits{Can Japanese
be input only with consonants?|. In Human Lan-
guage Technology Conference 2001.

K. Tanaka-Ishii, Y. Inutsuka, and M. Takeichi. 2001.
Personalization of text input systems for mobile
phones. In NLPRS.

W.J. Teahan. 2000. Probability esti-
mation for ppm. In NZCSRSC'95.
http://www.cs.waikato.ac.nz/ wjt/papers
/NZCSRSC.ps.gz.

Tegic 9. 2000. Tegic 9 home page.
http://www.t9.com.

D. Ward, A.F. Blackwell, and D.J.C. MacKay. 2000.
Dasher:a data entry interface using continuous ges-
tures and language models. In the ACM Sympo-
sium on User Interace Software and Technology,
pages 129{137.

S. Zhai, M. Hunter, and B.A. Smith. 2000. The
metropoliskeyboard-an explorationof quantitative
techniques for virtual keyboard design. In UIST.
ZI-Corp. 2000. ZI home page. Available from
http://207.229.18.241/.
