AUTOMATED WORD PREDICTION IN BANGLA LANGUAGE USING STOCHASTIC LANGUAGE MODELS

International Journal in Foundations of Computer Science & Technology (IJFCST) Vol.5, No.6, November 2015
DOI:10.5121/ijfcst.2015.5607 67
AUTOMATED WORD PREDICTION IN BANGLA
LANGUAGE USING STOCHASTIC LANGUAGE
MODELS
Md. Masudul Haque 1
, Md. Tarek Habib2
and Md. Mokhlesur Rahman3
1
Dept. of Electrical and Computer Engineering, North South University, Bangladesh
2
Dept. of Computer Science and Engineering, Daffodil International University,
Bangladesh
3
Dept. of Computer Science and Engineering, Prime University, Bangladesh
ABSTRACT
Word completion and word prediction are two important phenomena in typing that benefit users who type
using keyboard or other similar devices. They can have profound impact on the typing of disable people.
Our work is based on word prediction on Bangla sentence by using stochastic, i.e. N-gram language model
such as unigram, bigram, trigram, deleted Interpolation and backoff models for auto completing a sentence
by predicting a correct word in a sentence which saves time and keystrokes of typing and also reduces
misspelling. We use large data corpus of Bangla language of different word types to predict correct word
with the accuracy as much as possible. We have found promising results. We hope that our work will
impact on the baseline for automated Bangla typing.
KEYWORDS
Word prediction, stochastic model, natural language processing, corpus, N-gram, deleted interpolation,
backoff method.
1. INTRODUCTION
Auto complete or word completion works so that the user types the first letter or letters of a word
and the program provides one or more higher probable words. If the word he intends to type is
included in the list he can select it, for example by using the number of keys. If the word that the
user wants is not predicted, the user must type the next letter of the predicted word. At this time,
the word choice(s) is altered so that the words provided begin with the same letters as those that
have been selected or the word that the user wants appears it is selected. Word prediction
technique predicts word by analyzing previous word flow for auto completing a sentence with
more accuracy by saving maximum keystroke of any user or student and also reduces misspelling.
N-gram language model is important technique for word prediction. We use large data corpus for
training in N-gram language model for predicting correct Bangla word to complete a Bangla
sentence with more accuracy.
Word prediction means guessing the next word in a sentence . Word prediction helps disabled
people for typing,speed up typing speed by decreasing keystrokes,helps in spelling and error
detection and it also helps in speech recognition and hand writing recognition. Auto completion
decreases misspelling of word. Word completion and word prediction also helps student to spell
any word correctly and to type anything with fewer errors [1].

68
Figure 1.Word completion vs. word prediction. Suggested words are highlighted in yellow color.
We survey many techniques to predict upcoming words of a sentence in different languages
especially for English Language But there is no satisfactory analysis on Bangla language to
predict words in a sentence. So we apply some N-gram language model, backoff and deleted
interpolation techniques to predict Bangla words in a sentence. Word prediction is very important
and complex task in natural language processing (NLP) to predict the correct word to complete a
sentence in a very meaningful way.
We use statistical prediction technique such as N-gram technique as for example unigram,
bigram, trigram, bakeoff propagation, deleted interpolation. We also use large data set of text
word in Bangla which is collected from different news paper.
The paper is constructed as follows: related work in section 2, introduction of N-gram based word
prediction in section 3, methodology in section 4, implementation in section 5, result analysis in
section 6 and conclusion in section 7.
2. RELATED WORK
In an analysis of predicting sentences [2] researcher developed a sentence completion method
based on N-gram language models and they derived a k best Viterbi beam search decoder for
strongly completing a sentence. We also observed use of Artificial Intelligence [3] for word
prediction. Here syntactic and semantic analysis is done using the chart bottom-up technique for
word prediction. Another researcher suggests an approach [4] of word prediction via a Clustered
Optimal Binary Search Tree. They suggest using a cluster of computer to build optimal binary
search tree which also contain extra link so that bigram and the trigram of the language also
presented to achieve optimal performance of word prediction. Here the researcher does a lexical
analysis of most probable appeared word in users’ text.
In a paper of a learning classification based approach word prediction [5] they present an
effective method of word prediction using machine learning and new feature extraction and
selection techniques adapted from Mutual Information (MI) and Chi square (Χ2
). Some
researchers use N-gram language model for word completion in Urdu language [6] and in Hindi
language [7] for detecting disambiguation in Hindi word. There are some related works also on
Bangla language using N-gram language model such as grammar checker of Bangla language [8],
checking the correctness of Bangla word [9] and verification of Bangla sentence structure [10].
There are different word prediction tools such as AutoComplete by Microsoft, AutoFill by
Google Chrome, TypingAid [11], LetMeType [12] etc. There are some tools for word completion
in Bangla language such as Avro [13] . Some other Bangla software provide word completion
features only but theses software do not give word prediction or sentence completion features.
That is why we show word prediction process by using N-gram language model to complete a
Bangla sentence in our paper.
Tes| |
Test
Testing
Testify
Testimony
(a) Word completion.
|
(b) Word Prediction.

69
3. N-GRAM BASED WORD PREDICTION
N-gram language model is a type of probabilistic language model where the approximate
matching of next item is very high. Probability is based on counting things or word in most cases.
The probability of a word depends on the previous word which is called Markov assumption.
Unigram looks single item from a given sequence. Bigram is called first-order Markov model
which looks one word into the past and trigram is second-order Markov model which looks two
words into the past and quadrigram is third-order Markov model which looks three words into the
past and similarly an N-gram language model is N-1 Markov model which looks N-1 words into
the past[14]. Thus the general equation for this N-gram approximation to the conditional
probability of the next word in a sequence, w1,w2, ...,wn, is:
(1)
If N = 1, 2, 3 in (1), the model becomes unigram, bigram and trigram language model,
respectively, and so on.
For example, using unigram probability of the sentence " I want to eat " is,
P(I want to eat) = P(I) × P(want) × P(to) × P(eat)
Now, using bigram probability of the sentence "I want to eat " is ,
P(I want to eat) = P( I | <start>) × P(want| I) × P(to | want) ×P( eat | to)
Now, using trigram probability of the sentence "I want to eat " is ,
P(coming back to the) = P(I | <s> <s>) × P(want | <s> I) × P(to | I want) × P(eat | want to)
Now if any of these four words are not in the training corpus then the probability of the sentence
will be zero for the cause of multiplication. So in this statistical method if we want to consider
these words then we need a huge data corpus that must contain all the words of the language. So
there may arise the problems like:
• Many entries in corpus are with low frequency
• Probability of a word sequence will be very low or zero
If an entry does not exist in corpus then the probability of the sentence will become zero because
of multiplication.
To solve the problem, we have applied back-off and deleted interpolation model. In the backoff
method, for N = 3 in (1), i.e. for a trigram model, the word sequences will follow trigram
probabilities at first; if it could not match then word sequences will follow bigram model; if it
also could not match then word sequence will follow unigram model and predict at least a word.
As the system is for word prediction and it is playing with word sequence, so though any
probability of word sequence is zero for multiplication, it will predict at least a word. The
Backoff N-gram modelling is a nonlinear method. The difference is that in backoff , if there are
non-zero trigram counts, it rely on trigram counts and don’t interpolate bigram and unigram
counts at all [14]. The N-gram version (N = c) of backoff model can be represented as follows:
)
n
Nn
|w
n
P(w)
n
|w
n
P(w
1
1
1
1
−
+−
≈
−

70
( ) ( )............ ............, 02 1 2 1( 1) ( 1)|P w w w w if C w w w wi ii i i ii c i c >− − − −− − − −
( ) ( )
( )
1 ............ ............
............
2 1 2 1( 2) ( 1)
2 1( 2)
, 0
and 0
w w w w w w wii i i ii c i c
w w w wii ii c
P if C
C
α − − − −− − − −
− −− − >
=
( ).......
1 2( 1)|P w w w wi i ii c =
− −− − (2)
( )1
.
.
.
,c iP wα −
otherwise
The deleted interpolation algorithm combining different N-gram orders by linearly interpolating
all three models when they are computing any trigram. The detailed interpolation formula is as
follows:
(3)
3.1. WHY N-GRAM BASED WORD PREDICTION?
We choose N-gram based word prediction system because these are more statistical approach to
predict upcoming word in a sentence with more accuracy and N-gram language models provide a
natural approach to the construction of sentence completion systems. To measure the probabilities
by statistical model like N-gram data is divided into training set and test set. We compute the
probabilities of a test sentence from trained corpus which is designed very carefully. Suppose a
word “ ” occurs 400 times in a corpus of one million Bangla words. Then the estimated
probability for the word “ ” is 4 0 0
0 .0 0 0 4.
1 0 0 0 0 0 0
=
4. METHODOLOGY
Our approach starts with a sentence fragment and come up with a word using a stochastic
language model as shown in Fig. 2. We use five stochastic language models, namely unigram,
bigram, trigram, backoff and deleted interpolation.
Conditional frequency distributions come into play when language model learning takes place for
word prediction. A conditional frequency distribution refers to collection of frequency
distribution for the same experiment, run under different conditions. We intend to predict a word's
text (outcome), based on the text of the word that it follows (context). To predict the outcomes of
an experiment, a training corpus is examined first, where the context and outcome for each run of
the experiment are known. When presented a new run of the experiment, the outcome that
( ) ( ) ( ) ( ) ( ) ( ) ( )1 1 1
2 1 1 2 2 1 2 2 1 3 2
^
| |P n n nw w w w P w w P w w w P wn n nn n n n n n n nwλ λ λ− − −= + +− − − − − − − −
Fig. 2. Word prediction as the task of mapping a sentence fragment into a predicted word.
Language
Model
Input Output
Sentence
Fragment
Predicted
Word

71
occurred most frequently for the experiment's context is simply chosen. The total process is
shown in Fig. 3.
5. IMPLEMENTATION
Our work starts with a training corpus of size 0.25 million words. The corpus has been
constructed from the popular Bangla newspaper the daily “Prothom Alo”. The corpus contains
14,872 word forms. We have taken some test data in a file from the training corpus. After that
comprehensive unigram, bi-gram and tri-gram statistics have been automatically generated and
stored the sentences with predicted words in a file. Thus we have constructed N-gram models of
word prediction by counting frequencies of words in a very large corpus, i.e. database and
determine probabilities using N-gram. We employ a Java program in order to perform our
experiments using unigram, bigram, trigram, backoff and deleted interpolation language models.
First, we start working on unigram model. We split the entire corpus into two parts, namely
training part and testing part. We use holdout method [14] for selecting the proportion of data
reserved for training and for testing. We split the corpus at the proportion of two-thirds for
training and one-third for testing. In order to avoid model overfitting problem, i.e. to have low
training error as well as low generalization error, we use a validation set. In accordance with this
approach, we divide the original training data into two smaller subsets. One of the subsets is used
for training, while the other one, i.e. the validation set, is used for calculating the generalization
error. We fix two-thirds of the training set for model building while the remaining one-third is
used for error estimation. We repeat holdout method five times in order to find the best model.
After finding out our best model, we compute the accuracy of the model from the test set. We
evaluate the performance of the classifier by varying the length of test sentence. Likewise we
work on bigram, trigram, backoff and deleted interpolation models. All results are shown in Table
I and Fig. 4. For deleted interpolation model, we empirically found the linear weights λ1 = 0.5, λ2
= 0.33 and λ3 = 0.17 such that they sum to 1:
∑=
=
3
1
.1
i
iλ (4)
So, the probability equation for our deleted interpolation becomes as follows:
1 2 1 2 1
ˆ( ) 0.5 ( ) 0.33 ( ) 0.17 ( ).n n n n n nn n n
P w w w P w w w P w w P w− − − − −= + + (5)
We get the following probability for the following sentence or sample test data.
Test sample no. 1
Test sentence fragment:
Word to be predicted:
For unigram model, we get the following results.
Word to predicted:
P( ) = P( ) × P( ) × P( ) × P( ) ×
P( ) × P( ) × P( ) = 5.09E-5 × 1.19E-4 × 8.48E-5 × 5.94E-4 × 1.02E-4 × 2.38E-4*
0.01
Fig. 3. The approach for building our language model for predicting word.
Training
Corpus
Model
Learning
Model
Trained
Model
Applying
Test Sentence
Fragments
Predicted
Words

72
For bigram model, we get the following results.
Word predicted:
P( ) = P( | <s>) × P( | ) × P( |
) × P( | ) × P( | ) × P( | ) × P( | ) = 0.0 × 0.33 ×
0.14 × 0.2 × 0.03 × 0.17 × 0.14
For trigram model, we get the following results.
Word predicted:
P( ) = P( | <s> <s>) × P( | <s> )
× P( | ) × P( | ) × P( | ) × P( |
) × P( | ) = 0.0 × 0.0 × 1.0 × 1.0 × 1.0 × 1.0 × 1.0
For backoff model, we get the following results.
Word predicted:
P( ) = P( | <s> <s>) × P( | <s>
) × P( | ) × P( | ) × P( | ) × P( |
) × P( | ) = 0.0 × 0.0 × 1.0 × 1.0 × 1.0 × 1.0 × 1.0
For deleted interpolation model, we get the following results.
Word predicted:
P( | ) = 0.5 × P( | ) + 0.33 × P( | ) + 0.17 ×
P( ) = 0.5 × 1.0 + 0.33 × 0.07 + 0.17 × 0.1E-3
Test sample no. 2
Test sentence fragment:
Word to be predicted:
For unigram model, we get the following results.
Word predicted:
P( ) = P( ) × P( ) × P( ) = 0.0 × 1.02 E-4 × 0.01
For bigram model, we get the following results.
Word predicted:
P( ) = P( | <s>) × P( | ) × P( | ) = 0.0 × 0.0 × 0.33
For trigram model, we get the following results.
Word predicted: Empty string, because there is no word following in the training
corpus
For backoff model, we get the following results.
Word predicted:
P( ) = P( | <s>) × P( | ) × P( | ) = 0.0 × 0.0 × 0.33

73
For deleted interpolation model, we get the following results.
Word predicted:
P( | ) = 0.5 × P( | ) + 0.33 × P( | ) + 0.17 × P( )
= 0.5 × 0.0 + 0.33 × 0.33 + 0.17 × 5.1E-5
6. RESULT ANALYSIS
In order to understand the merits of our work in predicting words in Bangla, we need to delve into
all results found. We see from Table I and Fig. 4 that trigram, backoff and deleted interpolation
Table I. Results of all models used
No of Words in Test Sentence
Accuracy
Unigram Bigram Trigram Backoff
Deleted
Interpolation
5 21.89% 46.89% 52.16% 52.16% 52.16%
6 26.53% 49.53% 62.68% 63.18% 62.68%
7 23.89% 46.89% 65.32% 66.32% 67.95%
8 17.26% 44.26% 60.05% 61.05% 62.68%
9 28.53% 49.53% 67.95% 68.45% 67.95%
10 26.53% 49.53% 69.58% 71.58% 70.58%
11 20.26% 44.26% 65.32% 65.32% 62.68%
12 28.16% 52.16% 67.45% 68.95% 67.95%
13 21.26% 44.26% 62.68% 63.18% 62.68%
14 23.26% 44.26% 67.95% 68.95% 67.95%
15 18.63% 41.63% 67.95% 67.95% 67.95%
16 18.63% 41.63% 60.05% 60.55% 60.05%
17 17.26% 44.26% 65.32% 65.32% 62.68%
19 21.89% 46.89% 67.95% 68.45% 65.32%
20 18.63% 41.63% 41.63% 41.63% 41.63%
Fig. 4. Graphs of the results of all models used.

74
language model have performed almost in the same trend-line. Bigram model perform modestly,
whereas unigram models performance is obviously very poor. The average accuracies of all
models deployed are shown in Table II and Fig. 5. We see, from Fig. 5 and Table II, that the
average accuracies of trigram, backoff and deleted interpolation model are close, where the
accuracy (63.5%) of the backoff model is the maximum. Some results of predictions by all
models deployed are given below.
Performance of word predictor depends on some important component such as which language
model is used, the average length of the sentence in the language, amount of trained data set etc .
In our work we use 5 to 20 words in a sentence of 100 sentence for each execution from .25
million trained data set. We see from Fig. 5 that initially performance of word prediction
increases with the increase of N in N-gram language model.
4. CONCLUSIONS
N-gram based word prediction works well for English but we use for Bangla language which is
more challenging to get 100% performance because it depends on training corpus of large data.
We use more smoothed corpus data and increase data corpus size in future to get higher
performance. Here in our work we use single word prediction but we can develop our process to
predict a set of word to complete a sentence in a very meaningful way.
REFERENCES
[1] Auto Complete [Online Document], Available at: https://en.wikipedia.org/wiki/Autocomplete
[2] Steffen Bickel, Peter Haider and Tobais Scheffer, (2005), "Predicting Sentences using N-Gram
Language Models," In Proceedings of Conference on Empirical Methods in Natural language
Processing.
Table II. Results of all models used
Model Average Accuracy
Unigram 21.24
Bigram 45.84
Trigram 63.04
Backoff 63.50
Deleted Interpolation 62.86
Fig. 5. Graphs of the results of all models used

75
[3] Nestor Garay-Vitoria and Julio Gonzalez-Abascal, (2005), "Application of Artificial Intelligence
Methods in a Word-Prediction Aid," Laboratory of Human-Computer Interaction for Special Needs.
[4] Eyas El-Qawasmeh, (2004), "Word Prediction via a Clustered Optimal Binary Search Tree,"
International Arab Journal of Information Technology, Vol. 1, No. 1.
[5] Hisham Al-Mubaid , (2007), "A Learning-Classification Based Approach for Word Prediction", The
International Arab Journal of Information Technology, Vol. 4, No. 3.
[6] Qaiser Abbas, (2014), "A Stochastic Prediction Interface for Urdu", Intelligent Systems and
Applications , Vol.7, No.1, pp 94-100 .
[7] Umrinder Pal Singh, Vishal Goyal and Anisha Rani, (2014), "Disambiguating Hindi Words Using N-
Gram Smoothing Models", International Journal of Engineering Sciences, Vol.10, Issue June, pp 26-
29.
[8] Jahangir Alam, Naushad Uzzaman and Mumit khan, (2006), "N-gram based Statistical Grammar
Checker for Bangla and English", In Proceedings of International Conference on Computer and
Information Technology.
[9] Nur Hossain Khan, Gonesh Chandra Saha, Bappa Sarker and Md. Habibur Rahman, (2014),
"Checking the Correctness of Bangla Words using N-Gram", International Journal of Computer
Application, Vol. 89, No. 11.
[10] Nur Hossain Khan, Md. Farukuzzaman Khan, Md. Mojahidul Islam, Md. Habibur Rahman and
Bappa Sarker, (2014), "Verification of Bangla Sentence Structure using N-Gram," Global Journal of
Computer Science and Technology , vol.14 ,issue-1 .
[11] TypingAid auto completion tool, Available at
http://www.autohotkey.com/community/viewtopic.php?f=2&t=53630.
[12] LetMeType auto completion tool, Available at http://letmetype.en.softonic.com/download?ex=SWH-
1776.0#downloadingor .
[13] Avro Keyboard for auto completion tool in Bangla typing, Available at
https://www.omicronlab.com/avro-keyboard.html.
[14] Daniel Jurafsky and James H. Martin, (2000), Speech and Language processing, USA: Prentice-Hall,
Inc.
AUTHORS
Md. Masudul Haque has obtained graduate degree in Computer Science and
Engineering from North South University and expert level professional degree in ICT
from BUET. Now he is employing as a Programmer in the Local Government
Engineering Department under the ministry of Local Government of Bangladesh. He
has several publications in international and national journals and conference
proceedings. His research interest is in Database Security, Web Security, Artificial
Intelligence, Pattern Recognition and Natural Language Processing.
Md. Tarek Habib obtained his M.S. degree in Computer Science and Engineering
(Major in Intelligent Systems Engineering) and B.Sc. degree in Computer Science from
North South University in 2009 and BRAC University in 2006, respectively. Now he is
an Assistant Professor at the Department of Computer Science and Engineering in
Daffodil International University. He is much fond of research. He has had a number of
publications in international and national journals and conference proceedings. His
research interest is in Artificial Intelligence, especially Artificial Neural Networks,
Pattern Recognition, Computer Vision and Natural Language Processing.
Md. Mokhlesur Rahman was born at Satkhira in Bangladesh. He received his
B.Sc degree in Computer Science & Engineering from Mawlana Bhashani Science and
Technology University, Bangladesh in 2010. His research interest is in Artificial
Intelligence.He worked as a SENIOR WEB DEVELOPER at Squadro Solutions, Dhaka,
Bangladesh in 2010. In the following year, he joined to International Acumen Limited as
an ASSISTANT ENGINEER, Dhaka, Bangladesh. At the end of 2012, he joined as a
LECTURER to Prime University, Mirpur-1, Dhaka, Bangladesh.

AUTOMATED WORD PREDICTION IN BANGLA LANGUAGE USING STOCHASTIC LANGUAGE MODELS

More Related Content

What's hot (18)

Viewers also liked (20)

Similar to AUTOMATED WORD PREDICTION IN BANGLA LANGUAGE USING STOCHASTIC LANGUAGE MODELS (20)

More from ijfcstjournal (20)

Recently uploaded (20)

AUTOMATED WORD PREDICTION IN BANGLA LANGUAGE USING STOCHASTIC LANGUAGE MODELS