Shape Writer
Transcript of Shape Writer
Seminar Report On
SHAPE WRITING TECHNOLOGY
Submitted by
PREETHA V K
In the partial fulfillment of requirements in degree of
Master of Technology (MTech) in Computer and Information Science
DEPARTMENT OF COMPUTER SCIENCE
COCHIN UNIVERSITY OF SCIENCE AND TECHNOLOGY
KOCHI682022
2007
ACKNOWLEDGEMENT
I thank GOD almighty for guiding me throughout the seminar. I would like to
thank all those who have contributed to the completion of the seminar and helped me
with valuable suggestions for improvement.
I am extremely grateful to Prof. Dr. K Poulose Jacob, Director, Dept.of computer
Science, for providing me with best facilities and atmosphere for the creative work
guidance and encouragement. I would like to thank my coordinator, G. Santhosh Kumar,
Lecturer, Dept.of computer Science, CUSAT, for all help and support extend to me. I
thank all staff members of my college and friends for extending their cooperation during
my seminar.
Above all I would like to thank my parents without whose blessings; I would not
have been able to accomplish my goal.
2
ABSTRACT
This paper gives an overview of Shape Writer—a system for fast highaccuracy
text entry and command input on penbased computers. Shape Writer views words as
patterns mapped on a stylus keyboard. Using pattern recognition the system recognizes
users’ pengestures partially independently of scale and translation. The system has
shown potential in being a faster and more accurate replacement of handwriting
recognition, and in contrast to other specialized writing systems, it has a builtin smooth
learning curve. This paper describes the User Interface information and the recognition
architecture that is capable of pattern matching the user’s input against 10,000–60,000
patterns with very low latency (_ 35 ms on a 1GHz Tablet PC). It is concluded by
focusing some of the main features of shape writer.
3
CONTENTS
1. INTRODUCTION…………………………………….. ……... 4
2. BASIC CONCEPTS OF SHAPE WRITING…………. ……... 4
3. INFORMATION AND CONSTRAINTS…………………….. 5
4. USER INTERFACE………………………………………….. 6
4.1. Keyboard Design
4.2. Error Correction
4.3. Feedback
5. MULTI CHANNEL ARCHITECTURE……………………… 12
5.1 Template Pruning
5.2 Shape Channel Recognition
5.3. Location Recognition Channel
5.4. Channel Integration
6. INSTALLATION……………………………………………... 19
7. CONFIGURATION………………………………………….. . 19
8. FEATURES OF SHAPE WRITER……………………………. 20
9. CONCLUSION……………………………………………….. 22
10. REFERENCE ………………………………………………. 23
4
1. INTRODUCTIONShape Writing is a novel form of writing that uses pen strokes on graphical
keyboards to write text which is invented by Shumin Zhai and Per Ola Kristensson, IBM
researchers. Shape writing, previously known as Shorthand Aided Rapid Keyboarding
(SHARK), is a writing method designed to enable users to enter text efficiently at a faster
rate than previously possible on mobile phones, handheld computers and other mobile
devices. It can also be used with external tablets connected to desktop computers for
those who need an effective alternative to two handed touch typing on a physical
keyboard. Unlike the hand writing systems which rely on predefined symbols based on
either widely practised scripts (ie.natural handwriting) or novel symbols (such as
Graffiti), shape writing is defined on and used with a graphical keyboard.
2. BASIC CONCEPTS OF SHAPE WRITING
A shape writing system (current system is called Shape Writer) displays a
graphical keyboard to the user. Instead of tapping each individual letter key explicitly and
precisely, the user slides the pen over all the letter keys in a word sequentially on the
graphical keyboard.
Shape writing is easy. Rather than tapping individual keys, one simply draws a
continuous line from letter to letter on a graphical keyboard. The resulting pattern is
recognized by Shape Writer as a “sokgraph” (Shorthand On Keyboard as a GRAPH).
Figure 1 shows the ideal trace (left) and one actual and acceptable input stroke (right) for
the word “the”. The user's pen trace can deviate and sometimes not even cross some of
the intended keys as long as it is closer to the intended word's trace than any other word
traces in a lexicon.
5
The ideal trace is called a “sokgraph” (shorthand on keyboard as graph). In other
words, a sokgraph is the continuous trace that is formed by serially connecting the center
points of the keys on a graphical keyboard for a string of letters (normally a word).To
shape write a word, the user draws a pen stroke on the keyboard that approximates the
sokgraph of the word. Once the pen stroke terminates(by lifting the pen for example),the
shape writing system in principle compares the pen stroke on the keyboard with all
sokgraphs generated from a lexicon and returns the closest match. Shape writing is
inherently error tolerant, allowing noises such as hand tremors or faster and more
“sloppy” writing.
3. INFORMATION AND CONSTRAINTS
Text entry can be viewed as a communication system through which information
is transmitted from the user to the computer. With shape writing the writer encodes his
message geometric shapes defined on a graphical keyboard. Such a process is noisy both
because one often has to cross irrelevant letters in order to reach the intended letter and
because the writer can be sloppy and miss some relevant letters. In other words the stroke
drawn is often imperfect. The shape writing recogniser decodes the message from the
code + noise and outputs the message.
The most basic way of capturing the lexical level of language regularities is the
notion of a lexicon that consists of all permissible letter combinations including words,
parts of words, names and acronyms. The use of lexicon is a critical aspect of shape
writing. Using lexical constraints a shape writing system allows many irrelevant letters to
6
be crossed and still have the intended word returned. Taking fig 2 as an eg. , the letters
intended to be crossed were writing, but the actual pen stroke crosses wrtosise
tesing. Shape writer can still return the word “writing” because the lexicon
constraints eliminate all illegitimate letter strings. An intended word can still be
recognised although irrelevant letters between intended letters are crossed or even if some
of the letters in a word are missed by pen stroke.
Figure 2. Shape writer on a Tablet PC
4. USER INTERFACE
Shape writing is built on a foundation of pattern recognition of users’ input. As
such, there is an added complexity layer between the user and the system. Since pattern
recognition can result in mis recognitions the added complexity layer is another “failure
point” of the interactive system that does not exist in a traditional software keyboard. To
alleviate the issue effective user interfaces are required that simultaneously prevent errors
and minimize user effort to correct them when they do happen.
7
4.1. Keyboard Design
A stylus keyboard is an alternative text entry interface to handwriting recognition.
A stylus keyboard (sometimes called virtual or “graphical” keyboard) is an image of a
keyboard displayed on a touchsensitive screen. The user enters text by serially tapping
the keys with a pen. The first stylus keyboards used the traditional QWERTY keyboard
layout. However QWERTY is a poor choice, since QWERTY was originally designed to
prevent mechanical jamming in typewriters. As a result the frequent letter key
combinations are distributed to the left and right sides of the keyboard, which results in
frequent zigzag movements when typing text with a pen.
ATOMIK (Alphabetically Tuned and Optimized Mobile Interface Keyboard) was
designed to address this problem. It has the following three features.
First, ATOMIK has higher movement efficiency for stylus typing than any other existing
stylus keyboards. This was achieved by a Metropolis optimization algorithm in which the
keyboard was treated as a "molecule" and each key as an "atom". The "atomic"
interactions among all of the keys drove the movement efficiency – defined by the
summation of all Fitts' law movement times between every pair of keys, weighted by the
statistical frequency of the corresponding pair of letters in English towards the
minimum.
Second, the layout was alphabetically tuned. There is a general tendency that letters from
A to Z run from the upper left corner to the lower right corner of the keyboard. This gives
novice users a cue to look for letters that are no yet memorized.
Third, the letter connectivity of the most common words. Many common words or
comment fragments of words, such as "the" and "ing" are totally connected.
8
Figure 4.1 Shape Writer on ATOMIC Shape Writer on QWERTY
4.2. Error Correction
All text entry methods inevitably lead to errors. Therefore it is important to enable
fast and flexible error correction mechanisms. For instance, with a standard desktop
computer setup, the user can use the BACKSPACE, DELETE, HOME, END and arrow
keys to move the caret around in the document, and delete text in both the forward and
backward direction using the DELETE and BACKSPACE keys). The defacto direction
manipulation text editing interface used by the graphical user interfaces (GUI) allows the
user to move the caret at any desired position in the text. In fact, the search for
comparatively more effective tools for text editing led to the publication of a now famous
computer input device. Efficient error correction can be improved when words rather than
characters should be corrected. For this purpose the shape writing system uses a user
interface component called an edit buffer. Words appear in the edit buffer to the right and
pushes existing words to the left. When words cross the left edge they are synthesized into
keyboard key strokes and injected into the operating system’s key stroke dispatch queue.
4.2.1 Correcting a Confusion Error
With continuous shape writing, words are written on a wordbyword basis. This
means an error results in an entire word being incorrect. The system alleviates this
problem by providing several pengesture functions for intuitive editing. Figure shows
how a user deletes the two words and editing by simply crossing them. Any individual
word or sequence of words can be deleted by a crossing action
9
To enable efficiency while simultaneously reducing novice users’ frustration, the
selection process of an alternate word supports two different selection methods. With the
first method the user first taps on the word with the pen to reveal the list of alternate
words. Then the user selects the desired word from the list with another tap with the pen.
In the second method, the user taps and holds the pen down on the word. Then the user
slides the pen to the desired word and lifts up the pen. The second method is more fluid
but informal user testing revealed that some users are used to pulldown menu widgets
appearing after a tap only. These users were confused when the menu immediately
disappeared when they lifted the pen.
10
4.2.2. Correcting an Out-of-Vocabulary Error
Another possible error is the outofvocabulary (OOV) error. If a user wants to
write a word that does not exist in the lexicon the user has to add that word explicitly.
Since shape writing recognition does not rely on training data words can be immediately
added to the system’s lexicon. If a user taps a word using the software keyboard the
system automatically performs a check to see if the word exists in the system lexicon. If it
does not, the user’s tapped word is drawn with a surrounding dashed rectangle to create
an affordance for the user to click on it (see Figure a). When the user clicks on the word,
the user can select the ADD TO LEXICON function from the pulldown menu (see
Figure b). The new word can be immediately shape written by the user afterwards.
4.3. Feedback
4.3.1. Display of Recognized Word and Reinforcement of the Ideal Shape
As soon as the user lifts up the pen or finger, the recognized word is displayed.
The recognized word is displayed at the point of penup since that is the most likely
position where the user’s focus of visual attention will be. In addition the ideal shape of
the recognized word is displayed for a brief period of time (600 ms). The display of the
11
ideal shape reinforces the shape of the word to the user. Figure below shows the ideal
shape for the word system displayed over the software keyboard. The dot indicates the
starting position of the ideal shape.
4.3.2. Minimizing PenTrace Clutter
Displaying the ink of the pen as the user articulates a pengesture is advantageous.
First, it gives the user information that the pen motion is still recorded. Second, it
provides a sense of orientation to the user on where the pen has traveled on the keyboard.
However, some words are longer than others and when the pen has moved back and forth
on the keyboard the visual clutter from the pen trace becomes distracting. A solution is to
progressively fade the tail part of the pengesture when the trace increases in length
12
4.3.3. Morphing Visualization
A morphing algorithm is implemented to help novice users understand what part
of their pen traces contributes to the match of the recognized word. After the user has
lifted up the pen the complete pen trace of the user is drawn with blue ink and the ideal
shape of the word is drawn in red ink (see Figure 3.3.3). Thereafter both traces are re
sampled into an equal large number of equidistant sample points. Next, the sample points
that are indexed at the same position in the user’s pen trace and the ideal word shape are
connected by imaginary lines. A pair of two imaginary lines is then formed into an area
by connecting the four vertices in the linepair in sequence. This area is painted by a low
translucent blue color. The visual area explicitly communicates the spatial distance
between the pen trace and the ideal word shape to the user. To create a stronger visual
effect the user’s pen trace is gradually transformed into the ideal word shape over time
(Figure 3.3.3). The intermediate forms of the user’s pen trace are found by linear
interpolation.
13
5. MULTI CHANNEL ARCHITECTURE OF GESTURE RECOGNITION ENGINE
Three characteristics set the SHARK gesture system apart from other gesture
recognition systems. First, the number of sokgraph gestures each individual user may
need is much greater than the gesture repertoire of an alphanumeric recognizer such as
Graffiti or Jot, or a typical command gesture recognizer based on a linear machine, often
used in HCI research projects. Second, unlike natural longhand cursive handwriting
recognition, a fraction of the sokgraph short hand gestures, particularly those for short
words, can be identical or similar in shape; therefore shape alone may not provide
sufficient information to recognize the user’s intent. Third, a set of sokgraphs defined on
a keyboard layout constitutes a symbolic system novel to the user. This is both an
advantage and a disadvantage. It is an advantage because each sokgraph has a unique
ideal prototype, in contrast to natural hand writing in which even the same letter can be
written in perfectly legitimate but very different styles. It is also a disadvantage because
there is not a natural corpus of sokgraphs that can be collected, precluding many of the
standard datadriven machine learning approaches to recognition.
Fig. 4.3.3. The user has written the word computer. The user’s pentrace is gradually morphing towards the ideal word shape
14
Figure 5. A multichannel architecture for sokgraph recognition.
In consideration of these requirements we developed a multichannel recognition
system (see Figure 4). Each channel does not necessarily have enough discriminative
power, but the collective information from the multiple channels can separate the
sokgraphs sufficiently. The two core channels are a shape recognizer and a location
recognizer. The former classifies a pen gesture according to the normalized (scale and
translation invariant) shape of the pen gesture. The latter classifies a pen gesture
according to the absolute location of the gesture on the keyboard.
Both the shape channel and the location channel draw their recognition templates
from a lexicon. The SHARK paradigm requires the lexicon to include all (but just
enough) words a particular user needs in regular writing. This lexicon can be constructed
with various methods. It can be a preloaded standard dictionary, or a list of words
extracted from the user’s previously written documents, including emails and articles, or
words added by the user. In practice it is a combination of all.
15
5.1 Template Pruning
Due to the massive vocabulary required for the SHARK system, an initial pruning
component first filters out a large number of the sokgraph templates from entering later
stage recognition channels. An effective filtering mechanism is based on the start and end
positions of the sokgraph templates, normalized in scale and translation. Compute the
starttostart and endtoend distances between a sokgraph template and the normalized
unknown input gesture. If either of the two distances is greater than a set threshold, the
template will be discarded. This pruning process was implemented efficiently by only
storing the coordinates of the start and end points of a template pattern in a linked list.
Traversing the list and collecting the templates that pass this filter is thus an inexpensive
operation, even for large datasets.
5.2 Shape Channel Recognition
The most basic means of sokgraph classification is based on the shape
information contained in the user’s input gesture. There are many approaches to online
shape similarity measurements. Most relevant to the current work are those methods used
in pengesture recognition or handwriting recognition. Pengesture recognition commonly
uses trained classifiers based on the classic linear machine. Handwriting recognition
systems use a multitude of techniques and approaches, including neural networks, hidden
Markov models, and model matching. An early model based approach to cursive script
recognition system pioneered the use of socalled elastic matching in handwriting
recognition. While being outperformed in cursive script recognition by statistical
classifiers nowadays, elastic matching retains the valuable property of not requiring any
training at all. Elastic matching in cursive script recognition measures the spatial
similarity of two patterns by comparing the pointtopoint correspondences in sequence,
allowing certain elasticity if a nearby point has a shorter spatial distance than the
corresponding point.
Shumin Zhai and Per Ola Kristinsson initially experimented with using elastic matching
for the SHARK2 system, but in early testing they found that elasticity in the matcher is an
16
undesirable classification property when the template space is crowded with nearby
competing patterns. They found that they could significantly reduce the error rate by
introducing a linear matcher, a proportional matcher.
Proportional Shape Matching: The proportional shape matching distance between an
unknown pattern u and a template pattern t is defined as
where N is the total number of sampling points of the patterns. It is easy to see that the
elastic matching algorithm presented by Tappert with zero lookahead reduces to
Equation. Before applying Equation the patterns are resampled into N equidistant points
and normalized in scale and location. Normalization is achieved by scaling the largest
side of the bounding box of a pattern to a predetermined length L:
where W and H are the original width and height of the bounding box. Finally translate
the pattern’s geometric centroid to the origin in the coordinate system. The final result of
the shape channel is an approximate scale and translation invariant distance measure of
the similarity between the patterns, based on the average sum of the equidistant sample
points’ spatial distance.
5.3. Location Recognition Channel
The second core channel of sokgraph recognition in the current architecture examines the
absolute location of the user’s gesture trace on the keyboard. The rationale for having
such a channel is twofold: 1. Location information provides an increased recognition
resolution of sokgraphs; 2. Location is part of the user’s memory of a sokgraph and
therefore will be reproduced during gesture production.
17
Shape Confusion and Location Information By sokgraph shape alone some words
may be near or in complete confusion with each other. To gain a quantitative baseline
understanding of sokgraph confusion, present a brief analysis of the confusion pairs of
sokgraphs in a 20K lexicon on an ATOMIK layout. As shown in Table 1, there are a small
fraction of words that have identical sokgraphs. Considering normalized sokgraph shape
only (independent of scale and location), there are 1117 pairs of words that have identical
sokgraphs (confused pairs), for example root vs. heel, mend vs. shea, abe vs. ids, can vs.
cam, and ben vs buy. Many of these conflicts are not natural or complete English words.
For example, in a 20K lexicon, the word at conflict with du; as conflicts with lo, oz, by,
ny and ft; rjr conflicts with sas
If we consider shape plus the starting key position, the number of confusion pairs reduces
to 519. If we consider shape plus ending key position, the number is 522. If we consider
shape plus both the start and ending key positions, the number of confusion pairs reduces
to 493. Examples of confusion pairs with identical start and ending positions include
refuge vs. refugee, webb vs. web, and traveled vs. travelled. 284 pairs (58%) of these
remaining confusions pairs are Roman numerals, such as “lxvi” vs. “lxxxvi”, “xci” vs.
“xcii”, and “mmxvii” vs “mmxviii”. Table 1 also shows the numbers of confusion pairs
for the sokgraphs defined on QWERTY layout. In conclusion, location cue helps to
reduce sokgraph ambiguity.
18
Location Memory Since in the SHARK paradigm, users initially use and learn a
sokgraph by tracing the letters, the sokgraph location on the SK plays an important role in
the memory of the each individual sokgraph. Even for a completely memorized sokgraph,
the user is likely to remember its approximate location on the SK together with its shape,
particularly the beginning and ending positions.
Location Channel Algorithms location algorithm computes the distance of the
unknown input trace u to the template (ideal) trace t of word w on the SK (now both u and
t are absolute). t is defined by the lines connecting the centers of the letters that constitute
w. Both t and u are resampled to a fixed number N of equidistant points. The location
channel distance is defined as:
where N denotes the number of sampling points in the patterns. r is the radius of an
alphabetical key. This means that we form an invisible “tunnel” of one key width that
contains all letter keys in w. A perfect distance score of zero is given when the entire
gesture input trace u is within this tunnel of t . Otherwise, the sum of the spatial pointto
point distances is used. In other words, Equations (36) give special weight to traces that
are contained within the tunnel of radius r whose path is formed by serially connecting all
19
the individual keys used in a word. This was based on the observation from actual use
that when all letters in a word is traced (“tunneling”), one would expect the word to be
recognized no matter what the shape of the trace is. are weights for
different pointtopoint distances. The shape of α (i) can be set in various ways. For
example it could be dynamically trained through a large amount of data when available. It
can also be prescriptively set. Currently use a function that gives the lowest weight to the
middle point, and the rest of the points’ weights increase linearly towards the two ends.
This is because when producing a gesture it is easier for the user to pay visual attention to
the beginning and ending points than the rest of the locations.
5.4. Channel Integration
The shape and location channels output distance scores between an unknown
gesture and templates drawn from the lexicon. These distances in the two channels are not
on a common scale and cannot be directly compared. The issue of multiple classifier
integration of distances or scores is not new in pattern classification. Several methods are
proposed using such methods as voting or distance to rank conversion. Bouchaffra et al.
present a method of deriving probabilities from handwriting recognizer scores based on
training. As is common in engineering, a reasonable assumption is that the distance from
an input gesture to the template of the intended word (in either channel) follows a
Gaussian distribution. In other words, if an input gesture has distance x to a template y,
the probability of y being the targeted word can be calculated using the Gaussian
probability density function:
where μ = 0 and σ can usually be obtained through training from large amount of data.
σ reflects how sensitive a channel is. For example if σ equals to one key radius, those
templates whose distance to the input gesture are greater than one key width ( 2σ ) have
20
practically zero probability of being the intended sokgraph. In the current system σ used
as a parameter to adjust the weight of the contribution of each channel. The greater σ is,
the more flat the p(x) distribution will be, and hence the less discriminatory the channel is
when it is integrated with the other channel (hence less weight). As a pruning
measure all candidates (templates) with x > 2σ are discarded without further processing.
Among the remaining candidates w∈W, the marginalized probability of a word w with
distance x being the user intended target word is:
Finally integrate the probabilities from the two channels using Bayes’ rule and obtain a
confidence score for the word:
where e p (w) S′ and p (w) L′ are the probability scores from the shape and location
channel respectively; S W and WL are the sets of the remaining candidates that passed the
2σ threshold pruning stage in the shape and location channel respectively. The result of
the integration process is a ranked list of the templates ordered by confidence.
6. INSTALLATION
Shape Writer requires Sun Java runtime version 5 or above to be present and the
install will fail and tell you so if it is not detected on your system. Once the current Java
runtime is installed the program will setup properly. This is not a standard Windows XP
installer so it might be important where you install it for it to work properly. Once Shape
Writer is installed it is invoked by executing the shark.exe file created by the install.
21
7. CONFIGURATION
SHARK has a settings area that provides good control over the user's experience
with the program. The first tab in the SETTINGS is the GENERAL tab that lets you
select if you are left or right handed. This is a big boon for lefthanded users as SHARK
will adapt to the different writing styles of the two groups. You can also turn off the
animated morphs I mentioned earlier in this section. The last setting you can make here is
to enable/ disable the phantom keyboard. The phantom keyboard is an interesting feature
that some users will really like. Since text is entered into SHARK by moving over the
keypad some users will have problems since the hand may obscure the keyboard on the
screen. The phantom keyboard produces a second identical (but nonfunctional) keyboard
alongside the real one. This lets you input words and still see the whole keyboard during
the process. The ink strokes can be toggled to show on the phantom keyboard too.
The next tab under SETTINGS is the LEXICON tab where you can specify a
special lexicon file to use instead of the integrated one. It is also here that you can add or
remove words from the lexicon. Following the LEXICON tab is the KEYBOARD tab
where you can toggle between the ATOMIK and QWERTY key layouts with a preview of
the particular layout selected. There is also a place here to use special layouts but I
wouldn't mess with these without an intimate knowledge of the program. It is an
interesting inclusion by the SHARK programmers as it would allow any layout to be used
by the program.
The last SETTINGS tab is the RECOGNITION tab where virtually every
technical setting of the actual recognition engine can be tweaked.
8. FEATURES OF SHAPE WRITER
Shape writing is easy. Rather than tapping individual keys, one simply draws a
continuous line from letter to letter on a graphical keyboard. The resulting pattern is
recognized by ShapeWriter as a “sokgraph” (Shorthand On Keyboard as a GRAPH).
Shape writing works on QWERTY and other graphical keyboard layouts. Customized
22
layouts can be made for special purposes or different languages. For high performance,
IBM has designed a layout based on ATOMIK, an optimized graphical on screen
keyboard. Sokgraphs defined on this layout are efficient to produce, tolerant to error, and
easy to remember. The following factors make shape writing particularly powerful.
Efficiency: Rather than articulating one letter at a time (longhand), shape writing allows
the user to write word level sokgraphs a form of shorthand.
Human sensitivity to geometric patterns: A person’s ability to recognize, memorize and
draw patterns is remarkable. Shape writing capitalizes on this remarkable human
capability. Drawing patterns with a stylus is fluid, dexterous and fun.
Intelligent pattern recognition: ShapeWriter is "intelligent". The number of legitimate
words (ranging from thousands to tens of thousands in a lexicon) is only a fraction of the
number of all letter permutations (tens of millions). ShapeWriter takes advantage of the
regularities of words formation and recognizes user's ink trace on keyboard with
maximum flexibility and error tolerance. An intended word can still be recognized
although irrelevant letters between intended letters are crossed or even if some of the
letters in a word are missed in the stylus trace.
Ease of learning: Shape writing bridges initial ease of use with eventual high
performance by embedding learning in use. In psychology terms, for initial ease of use,
the user interface needs to be recognitionbased – action by visual guidance. To reach
high performance, however, the user interface should support recallbased skills. In shape
writing, these two modes are gradually connected. One shifts from recognition to recall
over time. The graphical keyboard serves as a visual map and a training wheel from
careful visual tracing towards a fluid form of shorthand writing.
Ease of error correction: Underneath each word in ShapeWriter's text stream editor is a
list of probable alternative words that can be selected with one additional pen stroke. One
can also delete or insert words anywhere in the text stream editor.
ShapeWriter also supports Command Strokes a fluid and efficient form of penbased
command.
23
9. CONCLUSION
Shape Writer recognizes continuous shorthand gestures and features a smooth
learning curve for users. Beginning users of shape writer traces the letter keys comprising
a word and over time the memory of the sokgraphs build up in the users’ memory. After
some usage users can quickly flick the sokgraphs without looking much at the keys. In
addition to text entry, shape writer can be used as a command recognizer for various
applications. To understand all issues involved and the full potential Shape Writer,
possibly even beyond mobile computing as a form of speed writing, requires a great deal
more research in the future.
.
24
10. REFERENCE
1. ZHAI, S. AND KRISTENSSON, P.O. 2007. Introduction to shape writing. In
MacKenzie, I.S. and TanakaIshii, K. (Eds.), Text Entry Systems, 139158. San Francisco:
Morgan Kauffman.
2. KRISTENSSON, P.O. 2004. Large Vocabulary Shorthand Writing on Stylus
Keyboard. Licentiatavhandling, Linköping University, Sweden.
3. KRISTENSSON, P.O. AND ZHAI, S. 2004. SHARK2: a large vocabulary
shorthand writing system for penbased computers. In Proceedings of the 17th Annual
ACM Symposium on User Interface Software and Technology (UIST ‘04). ACM Press:
43 52.
4. Per Ola Kristensson: Discrete and Continuous Shape Writing, for Text entry
and control, Dissertain No. 1106, 2007
5. ZHAI, S. AND KRISTENSSON, P.O. 2003. Shorthand writing on stylus
keyboard. In Proceedings of the ACM Conference on Human Factors in Computing
Systems (CHI ’03). ACM Press: 97104.
6. www.shapewriter.com
7. www.almaden.ibm.com
7. www.wikipedia.org
25