An Efficient Text Input Method for Pen-based Computers

Toshiyuki Masui
Sony Computer Science Laboratory Inc.
3-14-13 Higashi-Gotanda
Shinagawa, Tokyo 141-0022, Japan
+81-3-5448-4380
[email protected]

ABSTRACT was not the one that the user intended to use, the user types
Pen-based computing has not yet taken off, partly because of a “next candidate key” until the correct word appears as the
the lack of fast and easy text input methods. The situation candidate.
is even worse for people using East Asian languages, where
thousands of characters are used and handwriting recogni- On almost all the pen-based computers available in Japan,
tion is extremely difficult. In this paper, we propose a new either RKC or handwriting recognition is supported. Text
fast text input method for pen-based computers, where text input is slow and tiring using either of the techniques, for
is not composed by entering characters one by one, but by the following reasons. Specifying the pronunciation of every
selecting words from a menu of candidates created by filter- input word using a soft keyboard takes a lot of time, and
ing the dictionary and predicting from context. Using our the user must convert the pronunciation to the desired Kanji
approach, users can enter Japanese text more than twice as strings with extra keystrokes. Handwriting recognition has
fast as recognition-based and other existing text input meth- more problems. First, the recognizer has to distinguish be-
ods. User studies and detailed analysis of the method are also tween thousands of characters, often making errors. Many of
given. the characters in the character sets have similar shapes, so it
is inherently difficult to make recognition reliable. Second,
KEYWORDS: Input devices, Pen-based input, Predictive in many cases, users do not remember the shape or the stroke
interface, Hand-held devices, International interfaces, POBox order of Kanji characters, even when they have no problem
reading them. Finally, writing many characters with many
INTRODUCTION
strokes on a tablet is very tiring. With these difficulties, it is
Although a variety of pen-based computers are available these
believed to be difficult to enter Japanese text faster than 30
days, they are not as widely used as keyboard-based comput-
characters a minute on pen-based computers, which is several
ers, partly because entering text is much harder on pen-based
times slower than using keyboards.
machines. Traditionally, handwriting recognition techniques
and the soft keyboard (virtual keyboard displayed on the We have developed a new pen-based text input method called
tablet of a pen computer) used to be the main techniques for POBox (Pen-Operation Based On eXample), where users can
entering characters on pen-based computers, although other efficiently enter text in any language, using menus, word
techniques have also been proposed[4][6]. However, using prediction and approximate pattern matching. The remainder
any of these techniques takes much longer to enter text than of this paper demonstrates the details of POBox.
with a standard keyboard.
The situation is worse for East Asian languages such as Chi- STRATEGIES FOR RAPID TEXT ENTRY
nese, Japanese, etc. These, unlike European languages, have There is a big difference between the speed of typing on
thousands of character faces. Even with a keyboard, it is keyboards and pointing to characters on soft keyboards of
not easy to enter a character. A variety of techniques for en- pen-based computers. Computer users can easily type more
tering text into computer have been investigated. The most than five characters per second, while it is very difficult to
widely-used Japanese input technique is “Roman-Kanji con- touch three character keys per second, accurately on the soft
version” (RKC), in which a user specifies the pronunciation keyboard of a pen-based computer. In contrast, the speed of
of a word with an ASCII keyboard, and the system shows the selecting an item from a list is faster with a pointing device,
user a word with the specified pronunciation1 . If the word and many keyboard-oriented text editors (e.g. Emacs) now
1 Japanese
have mouse interfaces. For this reason, forcing the user
characters consist of two character sets. Kanji characters,
to enter many characters should be avoided on pen-based
computers, while a better approach should allow the user to
Published in: select a word from a list of candidates, in a minimum number
Proceedings of the ACM Conference on Human of penstrokes. We took the following approach.
Factors in Computing Systems (CHI’98) (April
1998), ACM press, pp. 328–335. imported from China, contain both meaning and pronunciation, while Kana
characters only represent pronunciation.
 Figure 1: Initial display. Figure 4: Selecting “first” after releasing the pen from the
tablet.

Figure 2: Selecting the “F” key.

Figure 5: Selecting “we” after selecting “first”.

Figure 3: Selecting “first” by dragging.

Figure 6: After selecting “we”.

Using dynamic menus to show candidates and select

by “e,” “r,” “s,” “f,” “x” and “c” keys.) This pattern
words: The desired word can be picked up directly from a
matches words like “synergy” and “dynasty,” but since
pulldown or popup menu.
“dynamic” has higher term frequency than these words, it is
Dynamic query for dictionary search: As soon as the user shown in the candidate word list for the selection. The other
specifies a portion of the pronunciation or the spelling of a is pattern matching allowing errors. This strategy is effective
word using the soft keyboard on the tablet, POBox shows a when the user does not remember the correct spelling or the
menu of candidate words that match the input. pronunciation of a word. In this case, POBox automatically
looks for words whose spelling or pronunciation is closest to
Using term frequency and example phrases: The words the pattern and shows them as candidates. Users can even
which are most likely to appear at the insertion point in the text specify only a portion of a word to get the desired word in
are shown at the top of the menu. The likelihood is calculated the candidate list.
from the term frequency and context. For example, since the
word “interface” tends to come after “user,” it appears at the Simple dictionary adaptation: Newly selected words are
top of the menu after the user has selected “i” as the first put at the top of the dictionary, and are likely to be shown
character following “user.” at the top of the menu so that the dictionary reflects the
characteristics of the current text.
Dynamic approximate string matching for selecting can-
didate words: If the pattern specified by the user does not EXAMPLES
exactly match any of the words in the dictionary, POBox Entering English Text
automatically performs approximate string search based on First, for explanatory purpose, we show how to use POBox
the following two strategies. One is spatial approximation, for entering English text, although POBox is more effective
where adjacent characters on the soft keyboard are treated for entering Japanese and other East Asian languages. We
equally in the search. This strategy is effective especially used the ACM CHI’95 Electronic Proceedings CD-ROM to
when the soft keyboard is small and precise selection is diffi- create an English dictionary with term and phrase frequen-
cult. For example, if the user failed to tap the right position of cies. We extracted plain text files from all the HTML files
a soft keyboard and selected “dtns” to enter “dynamic,” in the CD-ROM, counted the occurrences of words and word
no word in the dictionary matches “dtns” and POBox auto- combinations, and created the dictionaries by sorting the en-
matically searches the dictionary using the less strict pattern tries by frequency order. The remainder of this section uses
“[ersdfxc][rtyfg][hjbnm][weasdzx],” based on the sentence (“First, we show our technique for entering En-
the arrangement of ASCII keyboard. (“d” key is surrounded glish text.”) as the sample input text for our example.
 Figure 7: Selecting “show” from the menu. Figure 10: Specifying “mdtrn” to get “Mediterranean”.

Figure 8: Selecting the “E” key. Figure 11: After specifying “comple”.

Figure 9: Moving to the “N” key and selecting “entering”. Figure 12: After specifying “cplm”.

Figure 1 shows the startup display of POBox. When the user times to enter the phrase “First, we show our technique for
touches the “F” key, the display changes to Figure 2, showing entering.” Notice that the user made no spelling errors with
the frequently used words that start with “F” in a pulldown this method, since all the input words were taken from the
menu. Since the word “first” is a frequently used word and is dictionary.
found in the menu, the user can drag the pen and highlight the
word “first” as shown in Figure 3, and then take the pen off Using Approximate String Matching
the tablet to complete the selection. Alternatively, if the user
does not make a selection from the pulldown menu of Figure With the approximate string matching feature, even when the
3, he can choose the desired word from the popup menu as user does not specify the correct spelling of a word, there is
shown in Figure 4. a good chance of finding the desired word among the can-
didates. Also, the user can specify only part of the spelling
After selecting “first”, the display changes to Figure 5. In the to find the desired word. For example, if the user does not
menu at the bottom, the words that often come after “first” remember the spelling of “Mediterranean,” he can specify
are listed in order of frequency. The word combination “first “mdtrn” to see the list of words which are close to the pattern
the” appears 27 times in the CHI’95 CD-ROM, “first and” and then can find the right word in the list (Figure 10.)
and “first time” appear 20 times, etc. Since the next word,
“we,” happens to be in the list because “first we” appears 13 The same technique can be used to enter a word that has a
times in the CD-ROM, the user can directly select “we” by common prefix. If the user tries to enter “complementary”
touching it in the menu. After selecting “we”, the display and specifies “comple,” he still cannot find the word in the
changes to Figure 6. In this case, “show” is not found in candidates in Figure 11, since there are many commonly
the menu, but it can be selected from the pulldown menu by used words that begin with “comple.” Instead, the user can
touching the “S” key as shown in Figure 7. specify the characters that better represent the word. As
shown in Figure 12, the user can obtain “complementary” by
After this, “our”, “technique” and “for” can be selected in specifying “cplm,” although other patterns such as “cpmt”
a similar manner. Touching the “E” key does not make the will also work.
system display the next intended word (“entering”) as shown
in Figure 8, but touching the “N” key next narrows the search
Entering Japanese Text
space of the dictionary and “entering” then appears in the
menu for the selection (Figure 9). With POBox, users can enter Japanese text much more easily
than RKC and handwriting recognition systems. We show
From start to finish, the user only had to tap the tablet 15 the example by using “
 Word Spelling/Pronunciation
the THE
of OF
to TO
and AND
... ...

Figure 16: Word dictionary.

Figure 13: Initial display in Japanese input mode. Context Word Spelling/Pronunciation
of the THE
in the THE
to the THE
... ... ...
as well as AS
into the THE
... ... ...

Figure 17: Phrase dictionary.

Figure 14: Selecting “ ”.

frequency. Here, “context” means the word(s) that precede
the input word. The top portion of the initial phrase dictionary
is shown in Figure 17. Of all the phrases (lists of more than
one words), “of the” occurs most often and hence appears at
the top of the phrase dictionary.
Whenever possible, POBox checks the context and the char-
acters specified by the user, and generates the list of candidate
words for the next user input. First, it checks the phrase dic-
Figure 15: Before selecting “ ”. tionary and looks for the dictionary entries whose context
match the current context and whose spelling match the user
input. If such entries are found, POBox puts them into the
”2 as a sample Japanese input text. Figure 13 shows the candidate list. Then it checks the word dictionary and looks
initial display of POBox in Japanese input mode. A Hira- for entries whose spelling match the user input. If no entry
gana character table is displayed for entering pronunciations, is found in both of the dictionaries, POBox tries to find more
instead of the Roman alphabet in English mode. candidate words by performing approximate string matching
described in the next section. After the user selects a word
The pronunciation of the first word “ ” is “ ”(i- from the menu, the newly selected word and phrase are put
ka-ni), and the user can select the word by choosing “ ”(i) at the top of the dictionaries.
and “ ”(ka) from the Hiragana keyboard, just like in the
English example. Figure 14 shows how the user can select A middle-sized natural language dictionary usually has 20,000
the word “ ” with the pulldown menu. The user can to 50,000 word entries, which occupies less than 500KB of
select the next word “ ” (pronounced “hon-shuhou”) memory without compression. With appropriate compres-
after selecting its pronunciation “ ”(ho) and “ ”(n). sion and indexing techniques, a word dictionary plus a phrase
dictionary can easily be packed into 1MB of memory.
In this way, the user can enter Japanese text by specifying
the pronunciation of the first portion of the word and then Approximate String Matching
selecting the desired word from the menu, just like specifying Our approximate string matching algorithm is based on Baeza-
the spelling for English words. The user can input the phrase Yates’ “shifter algorithm”[1], with our extensions for allow-
“ ” in 7 penstrokes, whereas the ordinary RKC ing errors and handling simple wildcard characters. The
method requires at least 20 penstrokes. shifter algorithm is also used in an approximate string match-
ing program agrep[7] (an extension to grep on UNIX),
DETAILS OF THE ALGORITHM
where wildcard characters are treated differently from ours.
Dictionaries and Word Prediction
The word dictionary is a set of 2-tuples fword, spelling/
In our algorithm, we limit the wildcard to the basic “.*”
pronunciationg sorted by the term frequency of the word.
pattern in order to achieve simple and fast processing.
The top portion of the English word dictionary is shown in
Figure 16. Since “the” appears more often than any other S
a b c a
word in the corpus, it resides at the top of the dictionary, with
*
its spelling “THE.” The phrase dictionary is a set of 3-tuples
fcontext, word, spelling/pronunciationg sorted by the phrase Figure 18: A state transition machine which accepts
“ab.*ca”.
2 “Here, we show an example of entering text using this method”
 rors. After reading “abra”, state A1 also becomes active,
q w e r t y
a s d f g h
showing that “abra” matches “ab.*ca” with one error.
z x c v b n
q
g d
q This state transition can be calculated with simple logic and
w
h f
w shift operations. For a short pattern with small ambigu-
a a
S
b c ity, POBox first creates a deterministic state transition table
s s
z
v x
z from the nondeterministic state transition diagram like the
n v
x * x one shown in Figure 20, and uses the transition table instead,
for faster processing. For example, the state machine in Fig-
Figure 19: A state transition machine with spatial ap- ure 20 can be converted to a deterministic state transition
proximation table with 32 states.
EVALUATION
POBox currently runs on UNIX(X11), Windows95, Newton,
a b c a
A2
Java VM, and Pilot. POBox for Pilot is the latest version, dis-
tributed to the public on the Web3 since July 1997, and down-
ε ε * ε ε
* * * * * loaded by more than 10,000 people in two months. Since it
* * * *
a b c a
is the most widely-used version of POBox, we used it for
A1
the evaluation, although it lacks the pulldown menu feature
ε ε * ε ε because of its limited processing power.
* * * * *
* * * *
S a b c a A set of inquiries asking the user’s background and impres-
A0
sions of POBox was also presented on the Web page for
*
downloading POBox, and 1,057 people answered the ques-
Figure 20: State transition machine which allows er- tions. Among the 967 people with experience in both POBox
rors. and Japanese handwriting recognition systems, 126 people
(13.0%) said they feel that POBox is as efficient as handwrit-
ing recognition systems, and 796 people (82.4%) said POBox
is more efficient. Among the 899 people with experience in
Figure 18 shows a nondeterministic state transition machine both POBox and RKC systems, 118 people (13.2%) said they
which accepts a regular expression “ab.*ca”. In the shifter feel that POBox is as efficient as conversion-based systems,
algorithm, a bit string is used to represent the status of this and 718 people (80.1%) said POBox is more efficient. Sev-
state machine. For example, the initial state is represented as eral people sent back comments saying that they feel POBox
“10000”, and it becomes “11000” after accepting an “a”. is the most effective pen-based Japanese input method they
The state machine can be extended to perform spatial ap- have ever used.
proximate search by adding transitions by adjacent characters To obtain more reliable data, we asked POBox users who
(Figure 19.) The state machine can also be extended to allow answered the inquiry to compare the text input time using
errors by adding extra rows of states as shown in Figure 20. POBox and other handwriting recognition systems4 . Of these
A0 is the accept state with no errors, and A1 and A2 are the users, we selected approximately 300 people who seemed to
accept states with one and two errors, respectively. Like most have reasonable experience with both POBox and handwrit-
spelling correctors, POBox treats character insertion, deletion ing recognition systems, independent of their performance
and substitution as errors. Figure 21 shows the state transition on the two systems, and 31 people agreed to perform the ex-
by “abracadabra”. After reading “ab”, state A2 becomes periment and sent back the test results. All of them are adult
active, showing that “ab” matches “ab.*ca” with two er- male, and most of them are engineers in various Japanese
companies. About half are in their thirties, three are in their
forties, all of them having enough experience on both POBox
a b c a a b c a and handwriting recognition systems.
We asked the participants to measure the entry time of a
sample Japanese text consisting of 53 Kanji/Kana characters
and 2 punctuation characters, under the following conditions:

a b r 1. writing the text on paper.

a b c a a b c a
2. entering the same text using POBox.
3. entering the text using conventional RKC.
4. entering the text using the participants’ favorite Kanji hand-
writing recognition systems on any architecture.
3 http://www.csl.sony.co.jp/person/masui/POBox/pilot.html
Figure 21: State transition by “abracadabra”. 4 We offered calling cards (a value of approximately $5) to the participants

as a token incentive to perform the test seriously.

 Number of trials difficulty of writing on a tablet. No correlation was observed
20 (a) Handwriting on paper between the speed of writing on paper and the speed of en-
15
tering text using handwriting recognition systems.
10 The average text input speed using POBox was about 40
chars/min, which is approximately twice as fast as con-
5
average = 67.1 (49.2 chars/min) ventional RKC or Zaurus’ handwriting recognition system.
0
100 200
While the fastest handwriting recognition times observed
Text input time (sec) were shorter than the slowest POBox users, every individual
20 tested performed better with POBox than with the handwrit-
15 (b) POBox on Pilot ing recognition system.
10 Approximate String Matching
We have not advertised the approximate string matching fea-
5
average = 81.5 (40.5 chars/min) ture very much on the Web page, but 448 people (43.4%) of
0
100 200
the users noticed this feature. Of these 448 users, only 30 of
them (6.7%) answered that approximate string matching was
20 (c) RKC on Pilot not useful for them.
15
DISCUSSIONS
10 Stochastic Analysis of the Dictionary
5
The total number of words in the CHI’95 CD-ROM is about
average = 148.9 (22.2 chars/min) 650,000, and the distribution of the frequency conforms well
0
100 200 to Zipf’s rank-frequency law. From the data, the probabil-
ity of finding the desired word in the candidate menu after
20 (d) Handwriting Kanji recognition entering the top portion of the spelling can be calculated
on Zaurus
15 by summing up all the frequencies of words that appear in
the menu after each penstroke. This is the case when us-
10
ing POBox without the prediction from context feature. The
5
average = 164.4 (20.1 chars/min)
result is shown in Figure 23.
0
100 200 When the system shows 10 candidates after each penstroke,
about 53% of the input words can be found in the menu after
Figure 22: Distribution histograms of text input time specifying one character, and about 92% of words can be
using different methods. found after three penstrokes. This means that 92% of the
words can be entered with four penstrokes, while about 50%
of the words in the CHI’95 CD-ROM consist of more than
The second and third tests were performed on the Pilot, which
four letters. This result shows that the menu-based text input
does not have a Kanji handwriting recognition system. There-
method of POBox is effective even without the prediction
fore, we asked the participants to use their favorite recognition
mechanism.
systems, instead.
The same analysis for the Japanese dictionary is shown in
Among the 31 participants, 10 people used the same handwrit- Figure 24. Since about 50 Hiragana characters are used for
ing recognition system available on a Zaurus PDA5 (made by
Sharp). Other people used various handwriting recognition
systems on PCs and other PDAs, but the recognition time was i : number of penstrokes
P(i,n) :Hit ratio before showing the menu
longer than on the Zaurus. The summary of the test result 1.0
is shown in Figure 22. Since not all participants completed 0.9
i =5
all experiments for the same number of times, the area of the i =4
i =3
0.8
histogram differ among the tests. i =2
0.7
i =1
Input Speed Comparison 0.6
Most of the participants could write the sample text on pa- 0.5
i =0
per faster than with any of the electronic text input methods.
0.4
(The average was about 50 chars/min.) Writing speed does
0.3
not vary significantly between people. On the other hand,
the text input speed using Zaurus’ Kanji handwriting recog- 0.2

nition system does vary considerably from person to person, 0.1

the average being about 20 chars/min. This is much slower 0
1 2 5 10 20 50 100
than writing on paper, because of the recognition error and n : Number of candidates in the menu

5 Zaurus was the most popular PDA in Japan at the time this experiment Figure 23: Probability of finding the desired word in
was performed. the menu (English text).
 1.0 1.0 1.0
i=5 i=5 i=5
i =4 i =3 i =4 i =3
0.9 0.9 0.9
i =4 i =3 i =2 i =2
0.8 0.8 0.8
i =2 i =1
i =1
0.7 0.7 0.7
i =1
0.6 0.6 0.6
i =0 i =0
0.5 0.5 0.5
i =0
0.4 0.4 0.4

0.3 0.3 0.3

0.2 0.2 0.2

0.1 0.1 0.1

0 0 0
1 2 5 10 20 50 100 1 2 5 10 20 50 100 1 2 5 10 20 50 100
(a) Without prediction (b) With prediction (c) With prediction and dictionary adaptation

Figure 25: Probability of finding the desired word in the menu.

i : number of penstrokes
P(i,n) :Hit ratio before showing the menu
i characters is known. If it takes Tk for a user to input
1.0
i=5
one character and it takes Ts (n) to select an item from the
i =4 i =3
0.9 menu with n items, the average total time for entering a word
i =2
0.8
(T (i; n)) can be calculated by the following formula:
0.7 T (i; n) = Ts (n)
0.6
i =1
+ (Tk + Ts (n))(1 0 P ( ; n))
0
+ (Tk + Ts (n))(1 0 P ( ; n))
1
P1
0.5
+ ...
=0 (Tk + Ts(n))(1 0 P (j; n))
0.4
= Ts (n) + j
0.3
i =0
0.2 If the user starts using the menu after entering at least i
0.1 characters, the average total time T (i) is calculated by the
0
following formula:
1 2 5 10 20 50 100
n : Number of candidates in the menu

Figure 24: Probability of finding the desired word in T (i; n) = i 1 Tk + Ts (n) +

X
1
(Tk + Ts (n))( 1 0 P (j; n))
the menu (Japanese text).
j =i

Japanese text input, most of the desired words can be found We assume that Ts (n) is proportional to n and Tk is a constant
in the menu after two or three penstrokes, while more than value, since POBox shows a menu of candidates according
four penstrokes are required using ordinary Kanji-conversion to the probability of the words, and the user cannot tell the
methods. ordering of the words in the menu beforehand. We calcu-
lated T (i; n) using P (i; n) for the two cases of slow and fast
Dynamic Analysis character input.
More accurate hit ratio of POBox menus can be calculated
by simulating the prediction and adaptation mechanisms of Slow Character Input: Figure 26 shows the calculated av-
POBox with real English text. Figure 25(a) shows the hit ratio erage time for entering a word where character input speed
calculated by using all the texts in the CHI’95 CD-ROM. The is slow and Ts (n) can be estimated to be n=10 and Tk is the
hit ratio with the prediction from context feature is shown in constant 1. In this case, without prediction, the minimum text
Figure 25(b), and the hit ratio with prediction and dictionary input time is obtained when i = 1 and n = 3, which means
adaptation is shown in Figure 25(c). Prediction from context using a three-entry menu after one penstroke without a menu.
is effective for increasing the hit ratio, especially when no With prediction, the input time is minimized when i = 0 and
input is specified for selecting words (i = 0). In this case, n = 3, which means using a three-entry menu from the start.

POBox displays the correct word among its 10 candidates This is because frequently-used words are displayed at the top
38% of the time, whereas this number drops to 26% when of the menu even before the user specifies characters for fil-
prediction is not used. tering the dictionary. The estimated average time for entering
words is smaller with prediction than without prediction.
Input Speed Estimation
Text input speed can also be estimated by dynamic analysis Faster Character Input: Figure 27 shows the average time
if the character input speed using the soft keyboard and the for entering a word, where character input speed is faster than
speed of menu selection is known. the previous example and Ts (n) is estimated to be n=3. In
this case, minimum input time is obtained when i = 0 and
From the dynamic analysis shown above, the hit ratio P (i; n) n = 1, which means predicting one word every time after
of finding a word in the menus with n items after selecting entering a character.
T(i,n) i=5 existing common GUI tools with the prediction mechanism,
6.0 i=4 POBox can greatly reduce the time for text input on pen-based
i=3 computers, especially for Japanese and other languages where
i=2 direct text input is not possible.
i=1
5.0 i=0 Greenberg[5] argued that it is convenient to put frequently
used tools close at hand, and showed that this technique
is useful for issuing text commands in his WORKBENCH
4.0 system. POBox resembles the WORKBENCH system in that
both frequently used words and recently used words always
appear close at hand at the top of the candidate list for quick
selection.
3.0
Fukushima et al.[3] showed that input word prediction can
5 10 5 10 n reduce the search space and the number of penstrokes for
Without prediction With prediction handwriting recognition of Japanese texts. Although they
reported that their prediction system could reduce input pen-
Figure 26: Text input speed estimation with slow char-
strokes from 10 to 40 percent, problems with handwriting
acter input. (Tk = 1, Ts (n) = n=10)
recognition still remain and the text input speed does not
increase dramatically.
T(i,n) i=5
i=4 CONCLUSIONS
i=3 We developed a new fast text input method for pen-based
7.0 i=2 computers based on dynamic query of the dictionary and
i=1 word prediction from context. With our method, the speed of
i=0
text input on pen-based computers greatly increases and for
6.0 the first time, pen computing becomes a viable alternative to
keyboard-based input methods.
ACKNOWLEDGEMENTS
5.0 We would like to thank Jun Rekimoto and Jeremy Cooper-
stock for giving us many valuable suggestions. We also thank
many POBox users who actually used it, sent comments to
us, and performed the evaluation tests.
4.0
REFERENCES
3 6 3 6 n 1. Baeza-Yates, R. A., and Gonnet, G. H. A new approach to
Without prediction With prediction text searching. Communications of the ACM 35, 10 (October
1992), 74–82.
Figure 27: Text input speed estimation with faster char-
acter input. (Tk = 1, Ts (n) = n=3) 2. Darragh, J. J., Witten, I. H., and James, M. L. The Reactive
Keyboard: A predictive typing aid. IEEE Computer 23, 11
(November 1990), 41–49.
In this manner, the fastest method for entering text depends 3. Fukushima, T., and Yamada, H. A predictive pen-based
on the relation between Tk =Ts (n) and P (i; n). Roughly Japanese text input method and its evaluation. Transactions of
speaking, when Tk =Ts (n) is very small (character input is Information Processing Society of Japan 37, 1 (January 1996),
very fast) as with a keyboard, the fastest way of entering text 23–30. in Japanese.
is entering characters without the use of menus. On the other
4. Goldberg, D., and Richardson, C. Touch-typing with a
hand, if Tk =Ts (n) is very large (character input is very slow),
using menus with many entries is faster. The two cases shown stylus. In Proceedings of ACM INTERCHI’93 Conference on
Human Factors in Computing Systems (CHI’93) (April 1993),
in Figure 26 and Figure 27 are between these extremes, and
Addison-Wesley, pp. 80–87.
POBox supports the entire spectrum.
5. Greenberg, S. The Computer User as Toolsmith. Cambridge
Related Work Series on Human-Computer Interaction. Cambridge University
Darragh’s Reactive Keyboard[2] predicts the user’s next key- Press, March 1993.
strokes from the statistical information gathered by the user’s 6. Venolia, D., and Neiberg, F. T-Cube: A fast, self-disclosing
previous actions and shows the predicted data for the selec- pen-based alphabet. In Proceedings of the ACM Conference on
tion. Unfortunately, the Reactive Keyboard is not usually Human Factors in Computing Systems (CHI’94) (April 1994),
useful for experienced computer users, since they can type Addison-Wesley, pp. 265–270.
much faster than selecting candidates from the menu. On
pen-based computers, however, people cannot enter charac- 7. Wu, S., and Manber, U. Agrep - a fast approximate pattern-
ters as fast as with keyboards, thus predictive methods like matching tool. In Proceedings of USENIX Technical Conference
POBox and the Reactive Keyboard are useful. By integrating (San Francisco, CA, January 1992), pp. 153–162.