Predictive Text System for Bahasa with Frequency, n-gram,
Probability Table and Syntactic using Grammar
Derwin Suhartono, Garry Wong, Polim Kusuma and Silviana Saputra
Computer Science Department, Bina Nusantara University, K H. Syahdan 9, Jakarta, Indonesia
Keywords: Predictive Text, Word Prediction, n-gram, Prediction, KSPC, Keystrokes Saving.
Abstract: Predictive text system is an alternative way to improve human communication, especially in matter of
typing. Originally, predictive text system was intended for people who have flaws in verbal and motor. This
system is aimed to all people who demands speed and accuracy in typing a document. There were many
similar researches which develop this system that had their own strengths and weaknesses. This research
attempts to develop the algorithm for predictive text system by combining four methods from previous
researches and focus only in Bahasa (Indonesian language). The four methods consist of frequency, n-gram,
probability table, and syntactic using grammar. Frequency method is used to rank words based on how
many times the words were typed. Probability table is a table designed for storing data such as predefined
phrases and trained data. N-gram is used to train data so that it is able to predict the next word based on
previous word. And syntactic using grammar will predict the next word based on syntactic relationship
between previous word and next word. By using this combination, user can reduce the keystroke up to 59%
in which the average keystrokes saving is about 50%.
1 INTRODUCTION
Conventional process of typing documents using a
typewriter has become obsolete due to technological
advances. This is clear as computer can help people
considerably in many daily activities. This progress
can also be felt when a computer help to predict
which words are going to be typed by user. The
ability to predict the word that is going to be typed
by user is often referred to predictive text system. It
is also often called as the word prediction system.
Predictive text is a part of the research in the
field of artificial intelligence especially on natural
language processing (NLP). NLP is a field of
computer science that focused on getting computers
to perform useful tasks involving human language,
tasks like enabling human-machine communication,
improving human to human communication, or
simply doing useful processing of text or speech
(Jurafsky and Martin, 2008), and predictive text
itself is a technique that helps the input process that
was used by people with disabilities or not on a
desktop system, handheld devices, and augmentative
communication systems (MacKenzie and Ishii,
2007). Originally, predictive text system was
intended for people who have flaws (defects) in
verbal and motor (Vitoria and Abascal, 2006). Over
time, the usage of this system began to change and
now is aimed to all people who demands speed and
accuracy in typing a document.
In the previous research, there are several
prediction methods in predictive text, such as
prediction using frequencies, prediction using word
probability tables, syntactic prediction using
probability tables, syntactic prediction using
grammar, and semantic prediction (Vitoria and
Abascal, 2006). N-gram is also introduced by other
researchers as another prediction method in a
predictive text system (Verberne, et.al, 2012).
However, if only one of the methods is used, it
will not efficient and effective enough for a
predictive text system. It is because the weakness of
the chosen method cannot be supported by another
method so when the system is being evaluated by
considering the value of keystrokes per character
(KSPC) and keystrokes saving, the result is not
satisfactory. Therefore, there are many other
researches that try to combine the existing methods
to achieve maximum result.
This research focused on predictive text in
Bahasa (Indonesian language) by combining some
prediction methods that are expected to be a more
305
Suhartono D., Wong G., Kusuma P. and Saputra S..
Predictive Text System for Bahasa with Frequency, n-gram, Probability Table and Syntactic using Grammar.
DOI: 10.5220/0004756603050311
In Proceedings of the 6th International Conference on Agents and Artificial Intelligence (ICAART-2014), pages 305-311
ISBN: 978-989-758-015-4
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
optimal solution than the previous researches.
Bahasa was selected as the focus of this research
because many existing researches used foreign
languages such as English, Swedish, etc., as its main
focus. Lack of knowledge of the Indonesian people
about good and proper grammar is the other reason.
It caused slang or colloquial language which is often
unconsciously used in the official documents that
should have used the proper language.
This research is expected to help Indonesian
people in the process of typing a document more
quickly and precisely based on the proper
Indonesian language.
2 RELATED WORKS
Previous research that discuss about measuring
performance of predictive text system was a research
about keystrokes per character (KSPC) and
keystroke saving (MacKenzie, 2002). The result of
the research said that the smaller value of KSPC will
give a better performance for the system. A year
later, there was another research to evaluate the
accurate measurement of predictive system in case
of typing errors caused by users (Soukoreff and
Mackenzie, 2003). This evaluation was done by
minimum string distance (MSD) error rate and
KSPC. The research results were a new equation for
MSD error rate and development of KSPC formula.
Using this development, the used bandwidth which
represents useful information that was transferred
will be determined. Besides, it could determine the
wasted bandwidth and the total error rate.
The next research was a survey that revealed
several factors associated with the predictive text
system (Vitoria and Abascal, 2006). The research
stated that there are eight important factors that
affect a predictive text system. They were size of the
text block, dictionary structure, prediction method,
effect of the language used, the system interface,
system adaptability, system usability, and other
special features. The result stated that there are five
prediction methods that can be used in predictive
text. They are prediction using frequencies,
prediction using word probability tables, syntactic
prediction using probability tables, syntactic
prediction using grammar, and semantic prediction.
The survey also concluded that the result of a
predictive text system was expressed in terms of
keystroke saving and hit ratio or predictive accuracy
of a system can be considered as another measure
tools of predictive text system.
In 2008, there was a research that found a
standard of keystroke saving in evaluating a word
prediction system (Trnka and McCoy, 2008). The
result of this research stated that there are two limits
or boundaries that can become a standard evaluation
of a word prediction system. The two limits are
theoretical keystroke saving limit and vocabulary
limit.
Furthermore, there was a new development of
the predictive text system by incorporating some
prediction methods, such as using the rules of
English grammar to help text prediction and by
adapting to the amount of word usage frequency
(Nalavade, Mahule and Ketkar, 2008). The research
result declares the incorporation of these methods
can reduce KSPC by 26.91% compared to the T9
predictive text system.
The combination of semantic methods,
frequency, and part-of-speech model on keypads
was used in the next research (Gong, Tarasewich,
and MacKenzie, 2008). The result showed that it can
improve the text entry speed by 10% and reduce
errors as much as 20% depending on the keypads. A
year later, subsequent research did a combination of
syntactic and semantic method (Ganslandt, Jorwall,
and Nugues, 2009). The result declared that it can
reduce KSPC error in the Sweden corpus as much as
12.4%. In addition, when the combination of
syntactic and semantic coupled with the bigram
method, it can reduce the error up to 29.4%.
The next research was about a predictive text
system based on n-gram method (Verberne, et.al,
2012). N-gram was known as buffer and there are
two forms of buffer types (n-gram) which are
'current prefix of the word' and 'buffer15'. The
'buffer15' gave a better result than 'prefix of the
current word'. The summary of the combination of
predictive methods can be seen in Table 1.
3 PROPOSED ALGORITHM
The purpose of this research is to develop predictive
text system by combining some prediction methods
that hopefully can give smaller KSPC value than
previous researches. Methods that are used in this
research are:
3.1 Frequency
Frequency method is used to rank words in the word
table. It is based on how many times the word were
typed by the user. This method works by adding the
value of used word incrementally. By using this
method, predictive text system will offer words that
ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence
306
Table 1: Predictive Methods from Previous Research.
Researchers Year
Prediction
Methods
Result
Nalavade,
Mahule, and
Ketkar
2008
Frequency
and rules of
English
grammar
Decrease
KSPC by
26.91%
Gong,
Tarasewich,
and
MacKenzie
2008
Semantic
methods,
frequency
and part-of-
speech
model
Improve
text entry
speed by
10% and
decrease
error as
much as
20%
Ganslandt,
Jorwall, and
Nugues
2009
Syntactic
and
semantic
method
Decrease
KSPC
error as
much as
12.4%
Ganslandt,
Jorwall, and
Nugues
2009
Syntactic,
semantic
method and
bigram
Decrease
error up
to 29.4%
are frequently used by user.
3.2 Probability Tables
Probability table is a table that is designed to store
data. The data are predefined phrases from
Indonesian dictionary and corpus that has been
trained. Phrases are stored as static so user can select
faster on the prediction.
3.3 n-gram
N-gram is as a buffer that can be trained to predict
the next word based on previous word. In this
research, the used n-gram is bigram as the
differences between bigram and trigram do not
produce a significant difference and trigram makes
computing more complex. Therefore, bigram is the
most appropriate choice for this research. Training
result from bigram will be stored into probability
table.
3.4 Syntactic using Grammar
In this method, the system predicts the next word
based on syntactic relationship between previous
word and next word that has a greater frequency.
This relationship can be determined from data
training by n-gram. When data training is finished, it
will show the best probability of syntactic
relationship that can be used.
Database structure of this predictive text system
will contain three tables: word table, probability
table, and syntactic relationship table. Word table
contains all proper words that exist in Indonesian
dictionary: Kamus Besar Bahasa Indonesia (KBBI),
3
rd
edition.
Probability table is a table that contains
predefined phrase from KBBI and result from
training process using n-gram (bigram). Meanwhile,
syntactic relationship table is a table that contains
data about probability of syntactic relationship from
trained words. This table will be used as a reference
table for prediction to predict next word from the
greatest to the least probability of syntactic
relationship.
The sequential steps for predictive text system to
produce the desired word prediction are:
1. User types first character of desired word.
2. Predictive text system will trace words from
word table which its first character similar
with user typed.
3. System will offer collection of words sorted
by frequency value from bigger to smaller
and the highest syntactic relationship. If the
desired word is found, user can choose the
word by pressing predefined buttons on the
keyboard. In this research, predefined buttons
are listed from number one (1) to seven (7).
Afterwards, frequency of chosen word will be
incremented.
4. If the desired word is not found, user can type
the next character and return to second step
or word typed until complete.
5. Later on, when user presses space bar button,
system will show next prediction from
trained word by n-gram method and
predefined word as a phrase that are stored in
probability table. The word has the biggest
probability of syntactic relationship from
previous word.
6. When desired word is found, frequency from
the selected phrase or trained words by n-
gram will also be incremented either in the
word table or in the probability table.
7. After pressing space bar button, if the desired
word is not found, user can repeat the first
step. Then, when desired word is found, user
will press space bar button and n-gram
(bigram) will learn by catching two words in
front of space bar sign.
PredictiveTextSystemforBahasawithFrequency,n-gram,ProbabilityTableandSyntacticusingGrammar
307
8. System will look for the syntactic grammar
and store it into syntactic relationship table
from words that just stored and learned by n-
gram (bigram) in probability table. It will be
used for the next learning to decide the best
probability of syntactic relationship by
adding its frequency.
9. The process will be repeated from the
beginning to the last step until all words have
been completely typed.
In this research, the performance of predictive
text system was measured by using KSPC formula
without concerning errors or mistakes made by the
user and keystroke saving. By this limitation, the
used KSPC formula is adopted from MacKenzie and
Ishii (2007) as follows:
KSPC
∑
X
∈
∑
X
∈
Details of above formula: KSPC is the value of
keystrokes per character, w is a word in the word
model W, Kw is the number of keystrokes required
to enter w, Fw is frequency count for w, and Cw is
the number of characters in w. The reason of using
this KSPC formula is based on previous researches
that mostly use KSPC formula to evaluate the
performance of predictive text system. KSPC value
for QWERTY keyboard is one (1) because each
buttons represent a single character. KSPC value
must be lower than one (1) or the smallest KSPC
value for better performance on predictive text
system.
KSPC value represents how many keystrokes are
needed to type a document. Meanwhile, keystroke
saving represents how many keystrokes that are
saved. The used keystroke saving formula is adopted
from Trnka and McCoy (2008) which was stated in
below formula:
X100%
Where:
KS = Keystroke saving.
keys
normal
= The number of keystrokes for every
words.
keys
withprediction
= Number of keystrokes that required
to entry a word with predictive text
system.
Depicted from above formula, Keystroke Saving
(KS) is the amount of how many keystrokes have
been saved by the predictive text system.
4 RESULT AND DISCUSSION
To make sure this research goal is achieved, the
predictive text system is tested by using the
comparison of three prediction method groups.
Those method groups are shown in Table 2.
Table 2: Method Groups.
Where:
x = Used
- = Unused
The testing method uses the best case scenario
when user does not make any mistakes in typing
articles or documents. The test data or sample was
collected manually from www.liputan6.com.
Liputan6 is a news program from one of Indonesia’s
most popular television news channel called Surya
Citra Televisi (SCTV). It is known for delivering
actual, sharp, and trusted news in Indonesia. The 8
(eight) articles adopted from 4 (four) categories or
topics are collected from Liputan6.com’s online
article on 13
th
September 2013. The selected topics
are business, politic, health, and sport.
Steps of testing the predictive text system in this
research are:
1. Data will be trained for each prediction
method. Prediction method is divided into
three groups as shown in Table 2. For the
details, those method groups are: Dictionary
(prediction is only based on dictionary and
probability table without n-gram), Frequency
(prediction is based on dictionary with
frequency method and probability table
without n-gram and syntactic using
grammar), and Syntactic (prediction is based
on dictionary with frequency, probability, n-
gram, and syntactic using grammar).
2. Each method will train 2 (two) articles in one
topic sequentially.
3. When the first article has been typed, the
KSPC value will be recorded. The process
will be repeated to the second article.
4. Those articles will be tested again, and the
KSPC value will be recorded.
5. Then, the process will continue to the next
topic. Follow the second step until the fourth
step with the next topic.
ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence
308
6. Furthermore, KSPC value of each method
group will be shown and compared to find
the most effective prediction method.
For testing, the algorithm of this research was
implemented to a desktop application. It was built in
C# programming language and Microsoft Access as
the database that contains all of the proper words
and phrases. The database contains about 42,000
proper Indonesian words and about 17,000 most
used Indonesian phrases. All of the words and
phrases were obtained from KBBI, 3
rd
edition that
was published by Indonesia’s official language
organization.
The application was designed so user can type
with the help of predictive text system. User can
directly type the article in the application as shown
in Figure 1.
Figure 1: Main Display of Application.
In Figure 1, user can choose the prediction
method as explained in first step and there are two
buttons, which are “Reset” and “Statistics” buttons.
By pressing “Reset” button, it will clear all words
from text entry area and by pressing “Statistics”
button, it will show a new window as shown in
Figure 2.
Figure 2: Statistics Window.
In Figure 2, there are 4 (four) statistical values that
are displayed. Character Qty represents the length of
all words from the article. Number of Keystroke
represents how many words that user typed for the
article. Number of Backspace represents how many
times that backspace button pressed by user. KSPC
represents the result from calculation of KSPC
formula as stated before.
When user types one character in the application,
it will give prediction suggestion. User can choose
the desired word by pressing the number that
represents the word as shown in Figure 3.
Figure 3: Prediction Suggestion.
When tester does those steps above, it will be
shown in Figure 4, Figure 5, and Figure 6.
Figure 4: Prediction with Dictionary.
In Figure 4, predictive text system only makes a
prediction based on dictionary and probability table
without n-gram. And in Figure 5, the system is based
on dictionary with frequency method and probability
table without n-gram and syntactic using grammar.
PredictiveTextSystemforBahasawithFrequency,n-gram,ProbabilityTableandSyntacticusingGrammar
309
Figure 5: Prediction with Frequency.
In Figure 6, the system makes a prediction based on
dictionary with frequency method, probability table,
n-gram, and syntactic using grammar.
Figure 6: Prediction with Syntactic.
After testing in accordance to the steps above,
the result are shown in Table 3, Table 4, and Table
5.
In Table 3, all of the words are typed which
included the punctuation marks, foreign language,
and people or organization name. In this table, there
Table 3: Test Results for Training the Articles.
Where:
Topic = Selected topics from source.
Article = Article’s sequence number.
KSPC = Keystroke per character from each method
groups.
Average = KSPC average from each method group.
are no significant numbers as it is the first time that
system learns (for Frequency and Syntactic). The
result shows that the KSPC value is still high.
After training the data, the articles are tested
again included the punctuation marks, foreign
language, and people or organization name. And the
result is shown in Table 4.
Table 4: Test Results for Testing the Articles.
In Table 4, there are many differences that occur
from the result, especially in Frequency and
Syntactic method group. Both of them have smaller
KSPC value than Table 3. In Dictionary method
group, there is no difference from the previous
experiment because the prediction is only based on
dictionary (KBBI) and probability table without n-
gram. But the result of Syntactic method group is
not satisfactory because of the limitation of
dictionary and based on previous research which
stated keystrokes saving in practice can achieve 50
until 60% (Trnka and McCoy, 2008).
Based on the previous experiment, the articles
are tested again and focused solely on Bahasa
(without foreign language and people or
ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence
310
organization name). The result is shown in Table 5.
Table 5: Test Results for Testing the Filtered Articles.
In Table 5, the result is much better than before.
It shows that Syntactic method group is the most
effective combination for predictive text system and
can help people to save the keystroke about 50%.
5 CONCLUSIONS
Based on the test results, it can be concluded that the
most effective method is Syntactic method group for
Bahasa (prediction is based on dictionary with
frequency, probability table, n-gram, and syntactic
using grammar methods). It can save the keystrokes
until 50% (average) from each article with the best
saving is 59% and the lowest is 41%. In this
research, there are still many limitations for this
predictive text system, caused by vocabulary limit.
This research cannot find the newest edition of
dictionary (Kamus Besar Bahasa Indonesia,
4
th
edition) because it is not released as a digital data
yet, as so many articles contain special name or
acronym that is not supported by the system. With a
better and complete Bahasa database, the predictive
text system should be able to improve the keystroke
saving up to 60% focused solely on Bahasa.
REFERENCES
Ganslandt, S., Jorwall, J., Nugues. P., 2009. Predictive
Text Entry using Syntax and Semantics. Proceedings
of the 11th International Conference on Parsing
Technologies (IWPT). Association for Computational
Linguistics.
Gong, J., Tarasewich, P., MacKenzie, I. S., 2008.
Improved Word List Ordering for Text Entry on
Ambiguous Keypads. Proceedings of the Fifth Nordic
Conference on Human-Computer Interaction -
NordiCHI 2008. ACM.
Jurafsky, D., Martin, J. H., 2008. Speech and Language
Processing: An Introduction to Natural Language
Processing, Computational Linguistics, and Speech
Recognition. Pearson, Prentice Hall. New Jersey, 2
nd
edition.
MacKenzie, I. S., 2002. KSPC (keystrokes per character)
as a characteristic of text entry techniques.
Proceedings of the Fourth International Symposium
on Human–Computer Interaction with Mobile
Devices. Springer-Verlag.
MacKenzie, I. S., Ishii, K. T., 2007. Text entry systems:
mobility, accessibility, universality. Morgan
Kaufmann Publishers. San Francisco.
Nalavade, D., Mahule, T., Ketkar, H., 2008. PreText: A
Predictive Text Entry System for Mobile Phones.
Proceedings of the World Congress on Engineering
2008 Vol III. International Association of Engineers.
Soukoreff, R. W., MacKenzie, I. S., 2003. Metrics for text
entry research: An evaluation of MSD and KSPC, and
a new unified error metric. Proceedings of the ACM
Conference on Human Factors in Computing Systems
- CHI 2003. ACM.
Trnka, K., McCoy, K. F., 2008. Evaluating Word
Prediction: Framing Keystroke Savings. HLT-Short
'08 Proceedings of the 46th Annual Meeting of the
Association for Computational Linguistics on Human
Language Technologies: Short Papers Pages 261-264.
Association for Computational Linguistics.
Verberne, S., Bosch, A. V. D., Strik, H., Boves, L., 2012.
The effect of domain and text type on text prediction
quality. EACL '12 Proceedings of the 13th
Conference of the European Chapter of the
Association for Computational Linguistics.
Association for Computational Linguistics.
Vitoria, N. G., Abascal, J., 2006. Text prediction system: a
survey. Universal Access in the Information Society.
Springer-Verlag.
PredictiveTextSystemforBahasawithFrequency,n-gram,ProbabilityTableandSyntacticusingGrammar
311
