International Journal of Science and Research (IJSR), India Online ISSN: 2319-7064

Myanmar Spell Checker

Aye Myat Mon1, Thandar Thein2
1, 2
University of Computer Studies, Mandalay, Myanmar
1
[email protected], [email protected]

Abstract: Natural Language Processing (NLP) is one of the most important research area carried out in the world of Human
Language. For every language, spell checker is an essential component of many of Office Automation Systems and Machine
Translation Systems. In this paper, we develop a Myanmar Spell Checker System which can handle Typographic errors, Sequence
errors, Phonetic errors, and Context errors. A Myanmar Text Corpus is created for developing Myanmar Spell checker. To check
Typographic Errors, corpus look up approach is applied. Myanmar3 Unicode is applied in this system so that it can automatically
reorder the character sequence. A compound misused word detection algorithm is proposed for Phonetic Errors checking and Bayesian
Classifier is applied for Context Errors checking. In this system, Levenshtein Distance Algorithm is applied to improve users’ efficiency
by providing a suggestion list for misspelled Myanmar Words. We provide evaluation results of the system and our approach can handle
various types of Myanmar spell errors.
Keywords: Levenshtein Distance Algorithm, Myanmar Spell Checker, Myanmar Text Corpus, Natural Language Processing, Naïve
Bayesian Classifier

1. Introduction In this work, first we study the details on Myanmar

Language to identify the problem area of Myanmar spell
Spell checking is one of the most vital and widely studied errors and then we develop Myanmar spell checker. It
NLP tasks, which is used in orde r to increase the success consists of two phases: spell errors detection and suggestion
rate of NLP applications. Many NLP applications like list generation.
Machine Translation Systems, Text to Speech Systems and The rest of this paper is organized as follows: Section 2
Information Retrieval Systems require automated spell describes the related work. Section 3 describes nature and
checking of text. Many different techniques for detection and collation of Myanmar Language. Section 4 depicts spell
correction of spelling errors are based on English. Since error patterns and Section 5 describes the proposed spell
every language has its own writing system, the techniques checker framework. Implementation of Myanmar Spell
that perform well for one language, may not perform that Checker System is presented in Section 6. Experimental
well for some other languages and they may even totally fail. results are depicted in Section 7. Finally, conclusion on this
English spell checker will fail on the first step of recognizing work is given in Section 8.
Myanmar word boundaries because in Myanmar, unlike
English, word boundaries are not marked with spaces.
Human language translation is a difficult task for natural
2. Related Work
language because there has language ambiguity and varies Many researchers have been worked for spell checker of
according to their features and nature. Myanmar word Asian Languages. Even though other Asian spell checker
transformations are similar to other Asian Language researches have been done for two decades, Myanmar spell
including Indian, Japanese, Thai and Chinese Language. In checker research is still in its infancy. There is a very little
our country, Myanmar Language is used as an official amount of work done in this field. In this section, we discuss
language so spell checker is an essential role for the briefly some of the related work and history in the area of
development of Machine Translation system. spell checking and suggestion generation.
Myanmar is also among the languages whose writing Adbullah et.al [13] proposed an alternative approach to
system is different from that of English and therefore check the spelling of Bangla text that used Finite State
existing techniques cannot be applied for Myanmar spell Automata (FSA) to probabilistically crate the suggestion list
checking. Myanmar word does not have white space for a misspelled word. They used backtracking to add each
between words so it is difficult to tokenize. Although each possible solution to the suggestion list. Their system was
Myanmar word can be identified by word boundary only handled non-word errors in Bangla text. UzZaman et.al
correctly, all words may not have meanings because they are [11] proposed for generating suggestion for typographical
not in the dictionary. The most common reasons for errors with the edit distance of 2 from the misspelled word,
misspelled and misused words are caused by phonetic which obviates the need for computing the edit distances of
similarity and typing error of Myanmar characters. The the entire lexicon from the misspelled word. In [8], their
categories of error words are: (i) Typographic Error which is phonetic encoding was based on the Soundex algorithm,
mistyped the key in the wrong order and accidentally type modified to match Bangla phonetics. Their approaches used
characters (ii) Phonetic Error which is pronounced the same by PHONIX and Metaphone variants do provide some
as the intended word but the spelling is wrong (iii) Sequence contexts. Their encoding is equally applicable in a wide
Error which often caused the wrong format of character range of text processing applications, from searching for
sequence and (iv) Context Error which is pronounced the patient records in medical database to matching names in
same as the intended word but the word is ambiguous for the census record. Dhanabalan et.al [6] presented Tamil Spell
input sentence. Checker by providing possible suggestions for erroneous

Volume 2 Issue 1, January 2013

333
www.ijsr.net
 International Journal of Science and Research (IJSR), India Online ISSN: 2319-7064

words. User has the provision to select the suggestion among zwfawmhtoH) reflects the differences between spoken and
the list, ignore the suggestion or add the particular word to written Myanmar, as spelling is often not an accurate
the dictionary. Htay et.al [15] presented Myanmar word reflection of pronunciation. Some writers are writing with
segmentation using syllable level longest matching the pronunciation and careless of spell error. In Myanmar
approach. They used a combination of stored lists, suffix Language, every isolated word has meaning. And also there
removal, morphological analysis and syllable level n-grams have compound words. But some words are cannot combine
to hypothesize valid words with about 99% accuracy. The as a compound word. If we combine the two words, the
author [2] presented an approach consists of an approximate compound word’s meaning will be changed. For example,
word matching method, an N-best word segmentation (pdrf;Ægreen ) (vef;Æfresh ). If we combine these two
algorithm and used a statistical language model. Word-based words their meaning will be changed as (pdrf;vef; Æ green
correction method is proposed for Optical Character and lush). Typist may misuse the word (vef;) with
Recognition errors. It outperforms the conventional character
(vrf;Æroad). The two words (vef; and vrf; ) have same
based correction method. Fassati et.al [14] addressed the
pronunciation but different meanings. There is no
problem of real word spell checking and proposed a
combination of (pdrf; and vrf;). Myanmar words
methodology based on a mixed trigrams language model.
Their model had been trained and tested with data from the collocation depends on the previous meaning of words. One
Penn Treebank. Their approach has been evaluated in terms word has different meanings and different usages. So spell
of hit rate, false positive rate and coverage. checker is major issue and challenge for all computerized
Golding [3] proposed a hybrid method for context applications of Myanmar Language. Myanmar syllables and
defined symbols are shown in Table 1.
sensitive spelling correction by combining Bayesian
classifier and decision list. They extracted semantic and
Table 1: Type of Myanmar Syllables and Defined Symbols
grammatical features from the context of members of
confusion set using corpora. Chaudhuri [5] described a new
Defined
novel technique of location and correction of non-word Syllables Type of Syllables
Symbols
error. They pinpointed the error position in a big majority of
cases and thus reduce the number of correct alternatives to a u-t Consonants C
large extent. Their approach was based on matching the
j- ? -s Medials M1
string in the normal as well as a reserved dictionary. They
combined with a phonetic similarity key based approach -S ? -G Medials M2
where phonetically similar characters were mapped into a
single symbol and a nearly phonetic dictionary was formed. a Vowels V1
In this paper, we propose a Myanmar spell checker system
for handing Myanmar spell errors by applying Myanmar -m?-d ?-D?-k ?-l?–J ?-H Vowels V2
Text Corpus, Levenshtein Distance Algorithm and Naïve
Bayesian Classifier. -f Final F

-h ?-; Tone T
3. Nature and Collation of Myanmar
Language
Myanmar language is a very rich language and use as an Common misused characters and sample words are
official language of the Union of Myanmar. A Myanmar shown in Table 2.
syllable has a base character, and may also have (or not) a
pre-base character, a post-base character, an above-base Table 2: Common Misused Characters and Sample
character and a below-base character. Syllables have to be Words
constructed. Each syllable boundary should begin with a
base consonant. Myanmar languages have 33 consonants and Common Misused Characters Sample Words
the consonant combines with vowel and sometime it includes
medial to form the complete syllables in Myanmar language. u?c?* url? *rl ? cHk;wHwm;? *Hk;wHwm;
Besides, it has not delimiter between syllables and words. p ? q ? Z ? ps Z&uf ? quf&u?f pm;? qm;
Myanmar words are collated being based on syllables. A
P?e tEkjrL? tPkjrL
Myanmar syllable encoded in Unicode can be broken into 5
parts for collation [9]: <consonants> <vowels> <medial> y?z?b?A zl;? bl;? Al; ? ykef;? bkef;
<final> <tone>. In particular sentence, Typographic Errors o?w oHk; ?wHk;
(Non word errors) and Cognitive Errors (Phonetic Errors)
are collocated with two or more syllables. But Context ‘' ? " "g; ? 'g;
Errors (Real word errors) are only one syllable, which are ,?& ,uf ? &uf
ambiguous for poor reader.
The Myanmar saying “the pronunciation is merely the M ? –s Mum; ?usm;
sound, whilst the orthography is correct” (a&;awmhtrSef? wf ? uf ? yf wwf ?wuf? wyf

Volume 2 Issue 1, January 2013

334
www.ijsr.net
 International Journal of Science and Research (IJSR), India Online ISSN: 2319-7064

ef ? rf vrf; ?vef; word, which in fact leads to a misspelling where the

misspelling is pronounced the same as the intended word but
Sample of compound misused words and correct words
the spelling is wrong which accidentally produce a real word
are shown in Table 3.
(e.g., misspelling ‘weather’ as ‘whether’).
Examples: (i) opfyifrsm; pd rf ; vrf ; pdkjynf onf/
Table 3: Sample of Correct Words and Misused Words
(ii) ig;rsm; udk aevSr f; onf/
Correct words Misused words (iii) uGefysLwm uG e f ,uf qufoG,fa&; pepf onf vGefpGm
toHk;0if onf/
uwfaMu; uyfaMu;
In example sentence (iii), there has no combination of
ajcvSrf; ajcvSef; (uGef,uf). ”. The two words ( ,uf Æ rake) and (&uf Æ
weave) have the same pronunciation but different meaning.
aevSef; aevSrf; The correct combination for those error words is
xGef,uf xGef&uf (uGef&ufÆnetwork). The correct sentence is uGefysLwm uG ef &uf
qufoG,fa&; pepf onf vGefpGm toHk;0if onf/ “Computer network
uGef&uf uGef,uf system is very useful”.
quf&uf Z&uf
Context Errors: They can be seen to be a subset of phonetic
csDwuf csDwwf errors which produce a real word error (e.g., misspelling
‘piece’ as ‘peace’), where the word is pronounced the same
pdrf;vef; pdrf;vrf; as the intended word (e.g., ‘dessert’ as ‘desert’) but the word
is ambiguous for the input sentence.
pGJvrf; pGJvef; Examples: (i) &moDOwkonf awm udk rD onf/
(ii) rS sm; twGif; um; r&yf&/
4. Types of Spell Errors (iii) iSufrsm; avxJ wGif jyef aeMuonf/
In example sentence (iii), there has ambiguous word (jyef).
Spelling error patterns can result generally from the mistakes
made by human. Generally, human-generated misspellings Some writer misused the context word like that the word (jyef
can be distinguished into four groups: (i) Typographic Errors Æreturn) is used. The correct word for that sentence is (ysH
(Non-word errors) (ii) Sequence Errors (iii) Phonetic Errors Æfly): iSufrsm; avxJ wGif ysH aeMuonf/ “Birds are flying in the
(Cognitive errors) and (iv) Context Errors (Real word errors)
air”. In Myanmar words, (jyef and ysH) have the same
Typographic Errors: These errors have been made by the pronunciation but difference meanings and difference
typist accidentally presses the wrong key, presses the keys in usages.
the wrong order (e.g., misspelling ‘edit’ as ‘adit’). These
errors are made assuming that the writer or typist knows how 5. Myanmar Spell Checker Framework
to spell the word but may have typed the word hastily Myanmar Spell Checker Framework consists of four
resulting in an error. For example, “ola usmifodkYoGm;onf” . In
components as shown in Figure 1. They are: (1) Myanmar
this sentence, mistyped word is (ausmif). The typist need to Text Corpus, (2) Tokenizer, (3) Spell Checker and (4)
type (;, tone). The word (ausmif) has no meaning and for the Suggestion Generator.
above sentence the correct word is (ausmif;Æschool). rdk;acgif
aomaMumifh ukefaps;ESkH; rsm; jrifhwuf vmonf/ In this sentence,
the word (eSk H ;) has meaningless. The correct word is
(ES k e f ; Æprice ).

Sequence Errors: These errors can be caused in writing

Myanmar words with wrong format sequence that may be
two or three combinations of consonants, medial or vowels.
For example, (pigeon, cdk Æ“c- d-k” as “c- -k d”) (two
combinations) and (dove, csdK;Æ “c- -s - d-K- ;” as “c -d s - K -
;”) or (“c- -s - d -K -;” as “c -s -K -d - ;”) (three combinations).
Figure 1. Myanmar Spell Checker Framework
Phonetic Errors (Cognitive Errors): They have been made
by a lack of knowledge of the writer (e.g., misspelling
‘separate’ as ‘saparate’). These errors are made when the
writer substituted letters they believe sound correct into a
Volume 2 Issue 1, January 2013
335
www.ijsr.net
 International Journal of Science and Research (IJSR), India Online ISSN: 2319-7064

5.1 Corpus Creation tokenization of a word boundary with Finite State Automata.
Corpus is a large and structured set of texts. It is used to An automaton can be said to recognize a string [1]. In
spell checker, checking occurrences or validating linguistic Myanmar3, start state is always started with Consonant (C)
rules on a specific universe. Besides it is a fundamental basis and “end sate” is represented with double circle. Each
of many researches in NLP. Building of the text corpus is character in the input string passes through the
very helpful for the development of spell checking. In this corresponding edges to the next state. In this way, it reaches
work, Myanmar text corpus is created manually to apply in the final state, and then automatons accept the input string
Myanmar Spell Checker system. It contains various sense and return a word with boundary. According to the
meanings of ambiguous Myanmar words, compound words Myanmar word collation rule (e.g., ေက်ာင္း = က -် ေ- -ာ င -္ -း <
and training sentences. All words are collected from example C M1 V1 V2 C F T > ), we define the Finite State Automata in
sentences of “Myanmar Grammar” [10], “Myanmar Words figure 3.
Commonly misspelled and misused books [7]”, “Ornagai
Dictionary” [16] and “Wxpy Dictionary” [17].

Figure 3. Finite State Automata for Tokenization of

Figure 2. Myanmar Corpus Structure Myanmar Syllable

Myanmar Syllable file is used for checking Typographic Examples of Myanmar Syllable collation
errors which consists of 1908 syllables. Myanmar ေက်ာင္း =က -် ေ- -ာ င -္ -း <C M1 V1 V2 C F T>
Compound Words files is used for checking compound လ်ွင္ = လ -် -ွ -င -္ <C M1 M2 C F>
misused errors which misused as phonetic errors, it also used ျမန္ = မ ျ - န -္ <C M1 C F>
for segmented words for the input string. In Myanmar
ေကာက္ = က ေ- -ာ က -္ < C V1 V2 C F >
Compound words file, which consists of 62582 compound
စိုက္ = စ -ိ -ု က -္ < C V2 V2 C F>
words. Myanmar Training sentences consists of 3600
sentences and average words in sentences is 12. Training
sentences are used for calculating the probabilities of
6. Implementation of Myanmar Spell Checker
Context words errors.
6.1 Detection of Typographic Errors
Non-word errors correction is an important task. Non-word
5.2 Tokenization error spelling correction is focused on the task of generating
Tokenization is a preprocessing step for this system. It is the and ranking a list of possible spelling corrections for each
process of breaking a stream of text up into words, phrases, words not existing in the corpus. It is also isolated words
symbols, or other meaningful elements called tokens. The errors checking and generating suggestion. The main steps of
list of tokens becomes input for further processing such as Typographic Errors checking process are:
parsing or text mining. Tokenization is useful both in 1. Look up the word in the corpus
linguistics and in computer science, where it forms part of 2. In case, the word exit, pass on to next word.
lexical analysis. Typically, tokenization occurs at the word 3. If the word is not found in the corpus, calculate the
level. However, it is sometimes difficult to define all similarity of the error words and word from corpus to
contiguous strings of alphabetic characters and to define generate suggestion list.
what is mean by a "word”. Tokens are separated by
6.2 Detection Phonetic Errors
whitespace characters, such as a space or line break, or by
punctuation characters. In languages such as English where Phonetic error is a special class of real words errors in which
words are delimited by whitespace, this approach is the writer substitutes a phonetically correct but
straightforward [19]. However, tokenization is more difficult orthographically incorrect sequence of letter for the intended
for languages such as Myanmar, Thai, Japanese, and Chinese words. Moreover, there exists a class of real word errors in
which have no word boundaries. which the misspellings result in a valid word. It occurs due
Myanmar text is a string of characters without explicit to the presence of words in the language having similar
word boundary, so it is hard to define word boundary. In this pronunciation but different meaning. In this paper, we
paper, we describe regular expression and pattern for
Volume 2 Issue 1, January 2013
336
www.ijsr.net
 International Journal of Science and Research (IJSR), India Online ISSN: 2319-7064

proposed Myanmar compound misused words detection {

algorithm which detect phonetic errors. To detect this type of 18. cmbWÅw1+w2;
19. If cmbW not contain in the bigram word list
errors, the system used proposed algorithm and bigram 20. {
model. Then generate suggestion list by applying 21. If w2 and w3 not contains in the CW and SW and
Levenshtein Distance Algorithm. w3 equals with "Oneword"
22. {
6.2.1 Longest Matching approach and Proposed 23. PerrorÅw2+w3;
24. If Perror not contain in the bigram word list
Algorithm for compound misused words
25. Calculate similarity of Perror with word from
In Myanmar language, the text is a string of character written corpus;
in sequence from left to right and word are not always 26. else Calculate similarity of cmbW with word
delimited by spaces although sentences are clearly delimited from corpus; 
by a sentence boundary marker “/”. The text is needed to be 27. }
segmented into words as the preprocessing step in order to 28. }
29. }
process phonetic error checking. Many methods for word
30. }
segmentation have been proposed. We apply with longest 31. }
matching approach to check phonetic error of Myanmar 32. else if w1 equal with "Oneword" and w2 equals with
compound words. Syllable level longest matching algorithm "Twoword"
is described by [15]. 33. {
After word segmentation, we check the segmented 34. If w1 not contain CW and SW
35. {
words with the proposed algorithm to check phonetic errors
36. cmbWÅw1+w2;
by applying corpus lookup approach. The proposed 37. If cmbW not contain in the bigram word list
algorithm is shown in figure 4. To detect compound 38. {
misused words, we calculate the probability of next word by 39. Calculate similarity of cmbW with word from
using bigram model depict in Equation 1. corpus;
40. }
41. }
P (W n − 1 , W n ) (1)
P (W n |W n −1 ) = 42. } end if
P (W n − 1 ) 43. SegÅSeg-j
44. }end while
If there have probability, we define as correct compound 45. Print correct words
word. If there have no probability we define as misused
compound words and calculate similarity of the two Figure 4. Proposed Algorithm for Detection of Compound
continuous words. Misused Words

1. Input : Segmented words 6.3 Suggestion Generator

SegÅlength of the segmented words After checking the whole sentence, the system detects error
SWÅpredefined stop word list
CW Åpredefined context words
words. If error words present in the input sentence, it
int jÅ1 calculate the similarity of the error word and word from
2. while(Seg>0) corpus. And then we generate the possible suggestion list
3. { w1ÅSeg(j-1); which rank according to the most likely candidate by using
4. w2ÅSeg(j); Collection sort. The errors words and correct words will
5. w3ÅSeg(j+1) have a Levenshtein distance less than or equal to 3 which are
6.
if w1and w2 are equal with “Oneword” and not contain
considered to get more similar Myanmar words.
in SW and CW
7. { 6.3.1 Levenshtein Distance Algorithm
8. PerrorÅw1+w2; There are many kinds of String Similarity Algorithms for
spelling checking such as Hamming Distance, N-grams,
9. if Perror not contain in the bigram word list Longest Common Subsequence (LCS), Cosine Similarity
10. {
and Levenshtein Distance (LD). Among these algorithms,
If w3 not contain in the CW and SW and equal with
"Oneword" Levenshtein Distance Algorithm is the best algorithm for
11. { W4ÅPerror+w3; two fuzzy strings. In information theory and computer
12. Calculate Similarity of w4 with word from science, the LD is a metric for measuring the amount of
corpus; difference between two sequences (i.e., the so called edit
13. } distance). A generalization of the LD allows the
14. else
transposition of two characters as an operation and produces
Calculate similarity of Perror with word from
corpus the number of operations need to be transformed from one
15. } word to another. LD is a measure of the similarity between
16. else if w1 equal with "Twoword" and w2 equals with two strings, which we will refer to as the source string (s)
"Oneword" and the target string (t). It is used in some spell checkers to
17. { guess at which word (from a corpus) is meant when a missed
If w2 not contain in the CW and SW
Volume 2 Issue 1, January 2013
337
www.ijsr.net
 International Journal of Science and Research (IJSR), India Online ISSN: 2319-7064

spelt word is encountered and operate Insert, Delete and present for word senses is that it looks at the words around
Substitute transformations. At the end, the bottom-right confusion set in a large context window. Each content word
element of the array contains the answer. The resulted contributes potentially useful information about which sense
distance is the number of deletions, insertions, or of the ambiguous word is likely to be used with it. The
substitutions required to transform s into t [18]. supervised training of the classifier assumes that we have a
corpus where each use of ambiguous words is labeled with
6.4 Detection and Suggestion Generation of Context
its correct sense. For context error detection and correction
Errors
tasks, giving a word w, candidate classification variables
In Myanmar Language, most of context words are Myanmar S=(S1,S2,….,Sk) that represent the sense of the ambiguous
verb. For example, the confusion set {rSD? rD} have the same word and the feature F=(f2,f2,….,fn) by that describe the
pronunciation but difference meaning. In the context word context in which an ambiguous word occurs, the Naïve
“rSD” which would translate to English word “base on some Bayesian finds the proper sense s for the ambiguous word w
fact or evidence”, “rD” which use to combine with other by selecting the sense that maximizes the conditional
Myanmar Noun and verb. For example: “be with (reach, probability P(w=si|F). Suppose C is the context of the target
time, limit) (as in vufvSrf;rD? tcsdefrD) ”and then it can use as word w, and F=(f1,f2,…,fn) is the set of features extracted
from context C, to find the right sense s! of w given context
part :before; prior to (as in roGm;rD). Confusion set consists of C, we have:
words that are likely to be misused in place of one another.
⎡ ⎤
We can see in the following sentence that misused “ rSD ” s ' = arg max ⎢ ∑ log P ( f j | w = si ) + log P (w = si )⎥ (2)
instead of “ rD ”. olonfblwm&HkodkhtcsdefrSDvmonf/” The correct ⎢⎣ f j ∈C ⎥⎦
S = (S1,S2,….,Sk) sense of the context words
word for that sentence is “ He come railway station on time”.
F = (f1,f2,….,fn) , the set of features extracted from
In Myanmar Language, all context words can correct by
sentence which an confusion word occurs,
statistical techniques exception for {zl;?bl;}{bJ ? yJ} confusion
P(si) = The probability of sense (ambiguous word ) si
set. In Myanmar words, (bl; and bJ) are always use as P(fj|si)= the conditional probability of feature fj with
negative statement. The two words always combine with (r), observation of sense si
for example: ra&;bl;/ (not write), rpm;bJaeonf/ (live The probability of sense si, P (si), and the conditional
probability of feature fj with observation of sense si, P(fj|si),
without eating) . Myanmar verb always use between r and
are computed via Maximum-Likelihood Estimation:
bl; / bJ for describe negative statement. Myanmar context
P (s i )= C (s i ) / C ( w ) (3)
(f | w = si )= C ( f j , si ) / C (si )
words (confusion set) are shown in table 4.
P j (4)
Table 4: Myanmar Context Words
Where C(fj,si) is the number of occurrences of fj in a context
Confusion set of sense si in the training corpus, and C(si) is the number of
yJ bJ occurrences of si in the training corpus, and C(w) is the total
number of occurrences of the ambiguous word w [4]. To
zl; bl; avoid the effects of zero counts when estimating the
conditional probabilities of the model , when meeting a new
zuf buf
feature fj in a context of the test dataset, for each sense si, we
rSD rD set P(fj|w=si) equal 1/C(w).

rSsm; jrm; i=position;

Step 1: Processing
rif rSif a. Segment input sentence by longest matching
b. Remove stop words from input
ysH jyef Step 2: Confusion set Lookup
a. Lookup possible sense from the corpus Confusion word
equal with
In the literature, the problem of context sensitive spelling { zl;?bl;?bJ ?yJ} go to step(3)
correction is commonly formulated as a disambiguation task b. if not equal go to step(4)
where ambiguity among words is model by confusion sets. A
confusion set C= {W1, W2,……,Wn} means that each words Step3: Confusion word equal with { zl;?bl;? bJ?yJ }
Wi in the set is ambiguous with each other words in the set. If sense equal with(bJ) and position of sense equal 1
Change (yJ);
Thus, if C={piece, peace} and either piece or peace
encountered in a text, the task is to decide which one was else if sense equal with (bl;) and position of sense equal 1
intended. This way of identifying the actual form from an Change (zl;);
observed or surface form is called Bayesian classification.
Figure 5 shows the process of context errors detection and else if sense equal with (zl; or yJ) and sense position 0 or 1
suggestion generation by using Naïve Bayesian classifier. equal with r
The idea of the Naïve Bayesian classifier which we will Change (bl; or bJ);

Volume 2 Issue 1, January 2013

338
www.ijsr.net
 International Journal of Science and Research (IJSR), India Online ISSN: 2319-7064

-else if sense equal (bl; or bJ) and sense position 0 or 1 is not

equal with r
Change (zl; or yJ);
-else
print correct;

Step4: a) Calculate Probability

-for all sense si of W do
-for all words fi in the vocabulary do
P(fi| si)= C(fi,si)/ C(si)
-end Figure 6. Experimental result of Phonetic Errors Detection
-end
-for all senses si of W do
P(si)= C(si)/ C(w)
-end

b) Disambiguation
-for all sense si of W do
-score (si)=log P(si)
-for(all words fi in the context window c do
-score (si)=score(si)+log P(fi|si)
-end
-end
Choose s’=arg max score(si)
Figure 7. Similarity Score of Suggestion Generation for
Figure 5. Process of Detection and Suggestion Generation
Typgraphic Errors and Phonetic Errors base on Levenshtein
of Myanmar Context Errors
Distance Algorithm

7. Experimental Results
Accuracy
The performance of this system is evaluated in terms of Sentence Types
(%)
precision, recall and F-measure. Precision (P) means the
Test sentence in the corpus 98%
percentage of the correct word suggested by the system
which is divided by total number of error detected by the Test sentence that are partial words
91%
system. Recall (R) means the percentage of correct words include in the corpus
suggested by the system which is divided by the total number Test sentence that are not include in
82%
of sentence. F-score is the mean of recall and precision, that the corpus
is F= 2PR / (P+R). Testing sentences are used for evaluation
Table 5. Context Errors Detection and Suggestion
which consists of words include in corpus, test sentence that
Generation Results
are not exactly same sentences in corpus, and new words.
Corpus size is larger and larger because the tested sentences
are manually added to the corpus to get accuracy for new
words which are not included in corpus.
In this system, we tested with 500 sentences to get the
accuracy of the system. The average numbers of words
includes in one sentence is 12 words. Figure 6 shows the
accuracy of correctly detected on the testing sentences with
the compound words errors detected algorithm. Figure 7
shows similarity score suggestion generation for
Typographic errors and Phonetic errors by using Levenshtein
Distance Algorithm. In that figure, suggestion generation of
Typographic errors get 100% accuracy. But, at the Phonetic
errors, 91% similarity score of suggestion list are generated. Figure 8. OverAll System Evalutaion Results on Accuray of
Table 5 shows the accuracy of context errors detection and Correct words Vs. No. of Sentences
suggestion generation results. Average accuracy of overall
system gets 95% precision, 92.33% recall and 93% f-score. 8. Conclusion
We implemented a spelling checker system for Myanmar
language which can handle Typographic errors, Sequence
errors, Phonetic errors and Context errors. A Myanmar Text
Corpus is created and Mynamar3 Unicode is applied for
implementing the Myanmar Spell Checker system. We
applied Levenshtein Distance Algorithm, for generating
Volume 2 Issue 1, January 2013
339
www.ijsr.net
 International Journal of Science and Research (IJSR), India Online ISSN: 2319-7064

suggestion list. The proposed algorithm is very useful in [15] H.H. Htay, K. N. Murthy, Myanmar Word Segmentation
checking compound misused word errors of Myanmar using Syllable Level Longest Matching, Proceeding of
language. This system emphasized on Myanmar sentences the 6th Workshop on Asian Language Resources, 2008.
[16] http://www.ornagai dictionary.com
which follow Myanmar grammar rules and it cannot handle
[17] http://www.Wxpy dictionary.com
Parli words. This system can be applied in Myanmar NLP [18] http://www.Enclyclopedia.com/ Natural language
applications. Evaluation results show that this system can understanding / Levenshtein_distance.html
provide promising accuracy. [19] http://www.wikipedia.com/Tokenization.hml

References
[1] Lewis II, P.M, Rosenkrantz, D.J, Stearns, R.E, Compiler
Author Profile
Design Theory, Addison_Wesley Publishing Company,
Third Printing November, 1978.
[2] M. Nagata, Context-Based Spelling Correction for
Japanese OCR, Proceeding of the 16th International
Conference on Computational Linguistics, 05-09, 1996. Aye Myat Mon is currently pursuing Ph.D degree
[3] Golding, A, A Bayesian hybrid method for context program in University of Computer Studies, Mandalay, Myanmar. I
sensitive spelling correction, In Proceeding of 3rd got M.C.Sc from University of Computer Studies, Yangon in 2008.
workshop on Vary Large Corpora, 3, Jun 1996. I am also an assistant lecturer. My current research interest is
Natural Language Processing.
[4] Christopher D. Manning, Foundations of Statistical E-mail:[email protected]
Natural Language Processing, The MIT Press,
Cambridge, Massachusetts, London, England, 1999
[5] B. Baran Chaudhuri, Reversed word dictionary and
phonetically similar word grouping based spell-checker
to Bangla text, LESAL workshop, 2001.
[6] T Dhanabalan, Ranjani Parthasarathi, T V Geetha,
Tamil Spell Checker, Resource Center for India Thandar Thein received her M.Sc. (Computer
Science) and Ph.D. (Information Technology) degrees in 1996 and
Language Technology Solutions, TDIL newsletter, 2004, respectively from University of Computer Studies, Yangon
2003,Tamilnadu, India. (UCSY), Myanmar. She did her post doctorate research fellowship
[7] Myanmar Words Commonly Misspelled and Misused in Computer Engineering Department of Korea Aerospace
Book, Department of Myanmar Language commission, University, the scholarship awarded by Korea Research Foundation
Ministry of education, Union of Myanmar July, 2003. Grant funded by the Korean Government. She is a faculty member
[8] N. UzZaman, M. Khan, A Bangla Phonetic Encoding of UCSY since 1996. Currently she is a professor and guiding the
Ph.D students. Her current research interests are in the areas of
for Better Spelling Suggestion, Center for Research on Software Aging, Virtualization, Green Computing, Wireless Sensor
Bangla Language Processing, Proceeding of 7th Networks and Natural Language Processing.
International Conference on Computer and Information E-mail:[email protected]
Technology, Dhaka, Bangladesh, Dec, 2004.
[9] Stribley, K, Collation of Myanmar in Unicode, 22,
August 2005
[10] Myanmar Grammar, Department of Myanmar Language
commission, Ministry of education, Union of Myanmar
June 2005.
[11] N. UzZaman, M. Khan, A Comprehensive Bangla
Spelling Checker, Center for Research on Bangla
Language Processing, Proceeding of International
Conference on Computer Processing on Bangla, 2006.
[12] ျမန္မာစာ ျမန္မာ စကား , Department of Myanmar
Language commission, Ministry of education, Union of
Myanmar June 2007.
[13] Md. Munshi Abdullah, Md. Zahurul Islam, Mumit
Khan, Error tolerant Finite State Recognizer and String
Pattern Similarity Based Spelling Checker for Bangla,
Proceeding of 5th International Conference on Natural
Language Processing (ICON), 2007.
[14] D. Fossati, B. D. Eugenio, A Mixed Trigrams Approach
for Context Sensitive Spell Checking, Proceeding of the
8th International Conference on Computational
Linguistics and Intelligence Text, 2007.

Volume 2 Issue 1, January 2013

340
www.ijsr.net