HMM Based Part-of-Speech Tagger for Bahasa Indonesia

Abstract

In this paper, several methods are combined to improve the accuracy of HMM based POS tagger for Bahasa Indonesia. The first method is to employ affix tree which covers word suffix and prefix. The second one is to use succeeding POS tag as one of the feature for HMM. The last method is to use the additional lexicon (from KBBI-Kateglo) in order to limit the candidate tags resulted by the affix tree. The HMM model was built on 15000-tokens data corpus. In the experiment, on a 15% OOV test corpus, the best accuracy was 96.50% with 99.4% for the in-vocabulary words and 80.4% for the OOV(out of vocabulary) words. The experiment showed that the affix tree and additional lexicon is effective in increasing the POS tagger accuracy, while the usage of succeeding POS tag does not give
much improvement on the OOV handling.

Full text

On Proceedings of 4
th
International MALINDO (Malay and Indonesian Language) Workshop, 2
nd
August 2010
HMM Based Part-of-Speech Tagger f or Bahasa Indonesia
Alfan Farizki Wicaksono
School of Electrical Engineering and Informatics
Institut Teknologi Bandung
Bandung, Indonesia
farizki@comlabs.itb.ac.id
Ayu Purwarianti
School of Electrical Engineering and Informatics
Institut Teknologi Bandung
Bandung, Indonesia
ayu@stei.itb.ac.id
Abstract—In this paper, several methods are combined to
improve the accuracy of HMM based POS tagger for Bahasa
Indonesia. The first method is to employ affix tree which covers
word suffix and prefix. The second one is to use succeeding POS
tag as one of the feature for HMM. The last method is to use the
additional lexicon (from KBBI-Kateglo) in order to limit the
candidate tags resulted by the affix tree. The HMM model was
built on 15000-tokens data corpus. In the experiment, on a 15%
OOV test corpus, the best accuracy was 96.50% with 99.4% for
the in-vocabulary words and 80.4% for the OOV(out of
vocabulary) words. The experiment showed that the affix tree
and additional lexicon is effective in increasing the POS tagger
accuracy, while the usage of succeeding POS tag does not give
much improvement on the OOV handling.
Keywords : POS tagger, HMM method, affix tree, succeeding
POS tag
I.
I
NTRODUCTION
Part-of-Speech (POS) tagging is the process of assigning
part-of-speech tags to words in a text [7, 20, 5]. A part-of-
speech tag is a grammatical category such as verbs, nouns,
adjectives, adverbs, and so on. Part-of-speech tagger is an
essential tool in many natural language processing applications
such as word sense disambiguation, parsing, question
answering, and machine translation [12, 2].
Manually assigning part-of-speech tags to words is an
expensive, laborious, and time consuming task, hence the
widespread interest in automating the process. The main
problems in designing an accurate automatic part-of-speech
tagging are word ambiguity and Out-of-Vocabulary (OOV)
word. Word ambiguity refers to different behaviour of words in
different context. OOV word is a word that is unavailable in
the annotated corpus.
There are several approaches on POS tagging research, i.e.
rule based, probabilistic, and transformational based approach.
Rule based POS tagger assigns a POS tag to a word based on
several manually created linguistic rules [8]. Probabilistic
approach determines the most probable tag of a token given its
surrounding context, based on probability values obtained from
a manually tagged corpus [7]. The transformational based
approach combines rule based and probabilistic approach to
automatically derive symbolic rules from a corpus [3].
Bahasa Indonesia is the national language of Indonesia
which is spoken by more than 222 millions people [11]. It is
widely used in Indonesia to communicate in school,
government offices, daily life, etc. Bahasa Indonesia became
the formal language of the country, uniting its citizens who
speak different languages. Bahasa Indonesia has become the
language that bridges the language barrier among Indonesians
who have different mother-tongues. Even though, the
availability of language tools and resource for research related
to Bahasa Indonesia is still limited. One language tool that is
not yet commonly available for Bahasa Indonesia is POS
tagger system.
There are relatively little works on POS tagging system for
Bahasa Indonesia. Pisceldo et al [7] tried to develop POS
tagger for Bahasa Indonesia using Maximum Entropy model
and Conditional Random Field (CRF). The best performance of
Indonesian POS tagger reached in [3] is 97.57%. Triastuti [3]
also developed POS tagger for Bahasa Indonesia using CRF,
Transformation Based approach, and combining between CRF
– Transformation based approach. The best performance in [3]
reached 90.08%. Sari et al. [24] also applied Brill’s
transformational rule based approach in developing a POS
tagger for Bahasa Indonesia on a limited tagset trained on a
small manually and contextual features, they showed that the
method obtained an accuracy of 88%.
However, there is no deep study about developing POS
tagger Bahasa Indonesia by using Hidden Markov Model
(HMM). Hidden Markov Model is the established probabilistic
method for automatic POS tagger. Several languages have
adapted the HMM method in building the automatic POS
tagger [16, 17, 6, 1]. A POS tagger with HMM method was
proved to have better running time than any other probabilistic
methods [14]. In this study, we report our attempt in
developing a HMM based part-of-speech tagger for Bahasa
Indonesia
1
. We tried to make improvements such as using affix
tree to predict emission probability vector for OOV words and
utilizing information from dictionary lexicon (KBBI-Kateglo)
and succeeding POS tag.
II. M
ETHODS
A. Underlying Model
Both first order (bigram) and second order (trigram) Hidden
Markov Model are used for developing the tagger system.
Hidden states of the model represent tags and observation
1
We provide API (Application Programming Interface) of our tagger. This API is available at
http://students.itb.ac.id/home/alfan_fw@students.itb.ac.id/IPOSTAgger or http://www.informatika.org/~ayu/IPOSTAgger

states represent words. Transition probabilities depend on the
states, thus pairs of tags. Emission probabilities only depend on
current tag. Equation (1) is used to find the best POS tag
sequence in first order HMM (bigram) case.
∏ ∏
= =
−
−
××
=
n
i
n
i
iiii
ttn
twPttP
tPt
n
2 1
1
1...1
)|()|(
)(maxarg
1
.................. (1)
Fort the trigram case, the calculation is in equation (2).
∏ ∏
= =
−−
−
××
×
=
n
i
n
i
iiiii
ttn
twPtttP
ttPtPt
n
3 1
21
121...1
)|(),|(
)|()(maxarg
1
..... (2)
n
t
−1
is the best POS tag sequence for the input
words,
n
ww ...
1
are sequence of input word, and
n
tt ...
1
are
elements of the tag set.
)(
1
tP
in equations (1) and (2) are not
an ordinary unigram probability.
)|(
12
ttP
in equation (2) is
also not an ordinary bigram probability.
)(
1
tP
and
)|(
12
ttP
in both equations are conditioned on being the first
token in a sentence. In the implementation, we add a dummy
POS tag <START> before the first word (two dummies for
trigram case). Therefore,
)(
1
tP
in equation (1) is calculated
using
)|(
1
starttP
. And also,
)(
1
tP
and
)|(
12
ttP
in
equation (2) are calculated respectively using
),|( 1startstarttP
and
),|( 12 startttP
.
Transition and emission probabilities are estimated from a
tagged corpus. We use maximum likelihood probabilities
(MLE) method to estimate this value.
Another issue is on the estimation of transition probability:
)|(
1−ii
ttP
or
),|( 21 −− iii tttP
where the particular bigram or
trigram in not available in the training corpus. It causes sparse-
data problem [2]. Here, we use smoothing method to solve this
problem. For trigram case, we use deleted linier interpolation
method [2] to estimate the zero trigram transition probability
of
),|( 21 −− iii tttP
such as shown in equation (3).
),|(')|('
)('),|(
21312
121
−−−
−−
+
+
=
iiiii
iiii
tttPttP
tPtttP
λλ
λ
..... (3)
In equation 3,
1
321
=
+
+
λ
λ
λ
;
P
is probability
distribution and
'
P
is the maximum likehood estimation of the
probability. For bigram case, we use Jelinec-Mercer smoothing
[21] to estimate
)|( 1−ii ttP
such as in equation (4).
)(')1()|(')|(
11 iiiii
tPttPttP
λ
λ
−
+
=
−−
..... (4)
B. Affix Tree for Handling OOV Words
We use affix tree to obtain emission probability vector of
OOV word like Schmid did [18, 19] in his research. We make
an adaptation of Schmid’s method for Bahasa Indonesia. We
build three types of tree which cover the capitalized words, the
uncapitalized words, and the cardinal words (ke-5, 100, etc).
For example, if the OOV word is a capitalized word, the
searching process only involves the tree constructed from
capitalized word in lexicon.
FIGURE 1. A SAMPLE PREFIX TREE OF LENGTH 2
As explanation in [18, 19], the affix tree is automatically
built from training corpus. Affix tree is constructed from the
affixes (prefix or suffix) of length 3 of all words in the corpus.
Tag frequencies are counted for all affixes and stored at the
corresponding tree nodes. Then, an information measure
)(SI
is calculated for each node of the tree such as in
equation (5).
∑
×−=
pos
sposPsposPSI )|(log)|()(
2
....... (5)
Here, S is the affix which correspond to the current node
and
)|( sposP
is the probability of tag
pos
given a word
with affix
s
. Using this information measure, the affix tree is
pruned. For each leaf, the weighted information gain
)(
aSG
is calculated such as in equation (6).
))()(()()( aSISIaSFaSG
−
×
=
............ (6)
Equation (6) is suitable for suffix case. To apply it to
prefix tree, the
Sa
is employed to replace the
aS
. Here,
S
is the affix (prefix or suffix) of the parent node,
aS
(
Sa
) is
the suffix (prefix) of the current node, and
)(aSF
(
)(
SaF
) is the frequency of suffix
aS
(prefix
Sa
).

If the information gain at some leaf of the affix tree is
below a given threshold
2
, the leaf is deleted. The deleted node
will become default node of its parent node (the shaded node
in figure 1 is the sample of default node). If the default node is
only the remaining node, the parent node becomes a leaf and
is also checked whether it can be deleted or not.
To illustrate the process, consider the following example
(prefix case), where m is the prefix of the parent node, mi is
the prefix of one child node, and me is the prefix of the other
child node. Sample tag frequencies of the nodes are given in
table 1.
TABLE 1. SAMPLE TAG FREQUENCIES
tag Prefix m Prefix me Prefix mi
VBT 75 75 0
VBI 20 19 1
NN 7 5 2
total 102 99 3
The information measure for parent node is calculated by
equation (7).
0523.1...
102
20
log
102
20
102
75
log
102
75
)(
22
≈−−−=mI
… (7)
The corresponding values for the child nodes are 0.97 for me
and 0.91 for mi. Then, we determine the weighted information
gain at each of child nodes such as in equation (8) and
equation (9).
92.7)97.005.1(99)(
=
−
=
meG
........................... (8)
42.0)91.005.1(3)(
=
−
=
miG
................................. (9)
Value of
)(meG
is above a threshold of 3. On the other hand,
value of
)(miG
is below a threshold. And therefore, node mi
should be deleted from the tree and becomes a default node.
The tree is searched during a lookup along the path, where
the nodes are annotated with the letters of the word affix in
reversed order. If a leaf is reached at the end of the path, the
corresponding tag probability vector is returned. If no
matching child node is found at some node on the path, tag
probability vector returned is from the summation of all tag
frequency vectors from all default nodes. If no default node
exists, our tagger system just return the current tag probability
vector.
In our tagger, there are 3 types of tree configuration used
for handling OOV case. Each of these configurations was
tested to find the best configuration for Indonesian POS tagger
system:
1. Prefix Tree
This configuration uses the first N character of word to
create the tree. Prefix of OOV word is used to obtain
emission probability vector of OOV word.
2. Suffix Tree
This configuration uses the last N character of word to
create the tree. Suffix of OOV word is used to obtain
emission probability vector of OOV word.
3. Prefix – Suffix Tree
This configuration combines prefix and suffix feature
of word. The result of this configuration is the
combination of emission probability vector produced
by prefix tree and suffix tree. Each corresponding
vector entry in both trees is summed and renormalized
in order to be combined.
C. Succeeding POS Tag
Basically, for a HMM based POS tagger, succeeding POS
tag is unknown when decoding process is being done (using
viterbi algorithm [2, 21]). When decoding process reaches the
n
th
word in the sentence, the tagger does not know the POS tag
of (n+1)
th
word (succeeding POS tag).
Nakagawa et al [13] tried to develop POS tagger system
using SVM and two-pass method. This two-pass method
enables us to use succeeding POS tag. Our work tried to use
this two-pass method for developing Indonesian POS tagger.
We did some adaptations in order this two-pass method is able
to be implemented in our HMM-based system. Below is the
process of our two-pass method :
1. first pass
In the first pass, POS tags of the input words are
predicted ordinarily without using the information of
succeeding POS tag. This process is done using
equation (1) or (2). Affix tree is used to predict the
emission probability vector of OOV words.
2. second pass
In the second pass, the tagging process is repeated from
the first input word. This time, the tagger uses
information from succeeding POS tag if current word is
OOV in decoding process. The tagger does not use
succeeding POS tag if the current word is known. The
POS tag of succeeding word (succeeding POS tag) is
obtained from the first pass.
In the first pass, the tagger just calculate the probability
)|(
1−ii
ttP
(bigram case) and
),|(
21 −− iii
tttP
(trigram
case) for all states in the decoding process. Meanwhile, in the
second pass, if the tagger meets with OOV word, the transition
probability
)|(
1−ii
ttP
is replaced with
),|(
11 +− iii
tttP
(
),|(
21 −− iii
tttP
with
),,|(
121 +−− iiii
ttttP
).
1+i
t
is obtained from the first pass. It
means that the transition probability is not only conditioned on
2
We used a gain of threshold 3

being preceding POS tag but also succeeding POS tag. The
tagger uses maximum likehood estimation to determine the
value of
),|(
11 +− iii
tttP
or
),,|(
121 +−− iiii
ttttP
.
D. Lexicon from KBBI-Kateglo
We use lexicon from KBBI - Kateglo [9, 10] for reducing
entry in emission probability vector produced by affix tree.
Different POS tag set between corpus and KBBI – Kateglo is
the main problem. POS tag set in tagged corpus is the
specialized form of POS tag set in KBBI and Kateglo.
Therefore, we make a table called category table that maps
POS in KBBI - Kateglo into POS in corpus. Table 2 shows the
example of category table for some POS tags in KBBI –
Kateglo.
TABLE 2. SAMPLE OF CATEGORY TABLE
NO POS tag
KBBI
POS tag corpus
(specialized)
1 NN NN NNP NNG
2 CD CDO CDC CDP CDI
3 PR PRP WP PRN PRL
If POS tag A (corpus) in emission probability vector is not
the specialized form from one of the POS tag (KBBI –
Kateglo) produced from lexicon for corresponding word, this
POS tag A will be deleted from the probability vector.
For example, there is OOV word bisa in input sequence.
Emission probability vector produced by affix tree for this
word is {MD:0.05, NN:0.1, VBI:0.6, CC:0.25}. POS tag set
produced from KBBI – Kateglo lexicon is {MD, NN}. VBI is
not specialized form for either MD or NN, CC is not either.
Therefore, VBI and CC are deleted from the probability
vector. Our current probability vector is {MD:0.05, NN:0,1}.
III. E
XPERIMENTS
A. Experimental Data
Experiments were performed using small hand tagged
Indonesian corpus which consists of 35 POS tags (table 3). The
tagset is modification from tagset used in [23, 7]. The corpus
was developed as the correction and modification of PAN
Localization corpus
3
for Bahasa Indonesia [23, 15]. Our small
tagged corpus was divided into 2 sub corpora, one for training
(~12000 words) and one for testing (~3000 words). We did
experiments in 3 different testing corpus (each corpus contains
~3000 words). First testing corpus contains 30% unknown
words. Second testing corpus contains 21% unknown words.
The last testing corpus contains 15% unknown words.
TABLE 3. TAGSET USED FOR BAHASA INDONESIA
NO POS POS Name Example
1 OP Open Parenthesis ({[
2 CP Close Parenthesis )}]
3 GM Slash /
4 ; Semicolon ;
5 : Colon :
6 “ Quotation “ ’
7 . Sentence Terminator . ! ?
8 , Comma ,
9 - Dash -
10 ... Ellipsis ...
11 JJ Adjective Kaya, Manis
12 RB Adverb Sementara, Nanti
13 NN Common Noun Mobil
14 NNP Proper Noun Bekasi, Indonesia
15 NNG Genitive Noun Bukunya
16 VBI Intransitive Verb Pergi
17 VBT Transitive Verb Membeli
18 IN Preposition Di, Ke, Dari
19 MD Modal Bisa
20 CC Coor-Conjunction Dan, Atau, Tetapi
21 SC Subor-Conjunction Jika, Ketika
22 DT Determiner Para, Ini, Itu
23 UH Interjection Wah, Aduh, Oi
24 CDO Ordinal Numerals Pertama, Kedua
25 CDC Collective Numerals Bertiga
26 CDP Primary Numerals Satu, Dua
27 CDI Irregular Numerals Beberapa
28 PRP Personal Pronouns Saya, Kamu
29 WP WH-Pronouns Apa, Siapa
30 PRN Number Pronouns Kedua-duanya
31 PRL Locative Pronouns Sini, Situ, Sana
32 NEG Negation Bukan, Tidak
33 SYM Symbols @#$%^&
34 RP Particles Pun, Kah
35 FW Foreign Words Foreign, Word
In the first experiment, the HMM tagger using Affix Tree
was compared to a baseline tagger which was trained and
tested on the same data. The baseline POS tagger employed the
HMM bigram tagger. This baseline method tagged tag NN
(Noun) for the OOV words.
In another experiment, we compared several configurations
for the HMM based POS tagger. The configuration consists of
the N-gram model, the Affix Tree (prefix, suffix, prefix-suffix),
and so on. It examined the effectiveness of the Succeeding
POS tag and the Lexicon from KBBI – Kateglo.
3
This PAN Localization project corpus is available at http://panl10n.net/english/OutputsIndonesia2.htm

B. Experimental Result
Table 4,5,6 show the performance of the POS tagger in 3
different types of testing corpus. The performance is showed
in format
)/( AccWordUnknownAccWordKnown
AccuracyOverall
.
The result shows that Affix Tree is successful in increasing
the performance of OOV words tagging. It improved the
accuracy of OOV words tagging by about 24%-38% from the
baseline. From the experimental result table, prefix
configuration gives better improvement than using suffix or
prefix-suffix configuration.
Lexicon from KBBI – Kateglo is an important resource for
handling OOV. It gives better tagging accuracy than only
using affix tree. The difference of OOV word tagging
accuracy between the tagger using only affix tree and affix
tree+lexicon is about 2%-4% (1%-2% for overall tagging
accuracy). Not all POS tags produced by affix tree appropriate
with the corresponding word. This phenomena happens
because affix tree method only uses information from prefix or
suffix of corresponding word. Therefore, we delete
inappropriate POS tags from the vector using this lexicon.
Intuitively, this approach can raise the performance and the
test had proved it.
On the other hand, information from succeeding POS tag
(bipass) does not give much improvement on the tagging
accuracy. Many cases in table 4,5,6 show that succeeding POS
tag decrease the tagging accuracy. But, there are some cases
where using information from succeeding POS tag
successfully increases the tagging accuracy. Word berdiri in
the sentence below is the example.
saya berdiri di jalan .
From the sentence, the POS tag of word berdiri has to be
VBI (intransitive verb). Furthermore, the tagger correctly tags
the known words {saya, di} by {PRP, IN} because there is no
ambiguity in those words. If the tagger just uses affix tree, the
tagger will choose VBT (transitive verb) as the POS tag of
word berdiri because the probability of bigram <PRP-VBT> is
bigger than bigram <PRP-VBI> in the training corpus. If the
tagger uses succeeding POS tag information, the computation
is focused on trigram <PRP-VBT-IN> and <PRP-VBI-IN>. In
this case, the POS tag IN is the succeeding POS tag. From the
training corpus, the probability of trigram <PRP-VBT-IN> is
less than trigram <PRP-VBI-IN>. Therefore, the tagger
chooses VBI as the POS tag.
TABLE 4. RESULT ON 15% OOV TESTING CORPUS
NO
Configuration
Affix Tree Configuration
PREFIX
SUFFIX
PRE
-
SUFF
1 Baseline
90
.
65
%
(99.42%/42.34%)
2 Bigram
95.67%
(99.43%/75.00%)
94.32%
(99.39%/66.44%)
95.36%
(99.43%/72.97%)
3 Trigram
95.29%
(99.18%/73.87%)
94.29%
(99.22%/67.12%)
95.01%
(99.22%/71.85%)
4 Bigram+succeeding POS
95.57%
(99.43%/74.32%)
94.56%
(99.35%/68.24%)
95.36%
(99.43%/72.97%)
5 Trigram+succeeding POS
94.94%
(99.02%/72.52%)
94.04%
(99.06%/66.44%)
94.70%
(99.02%/70.95%)
6 Bigram+Lexicon
96.30%
(99.43%/79.05%)
95.01%
(99.43%/70.72%)
96.23%
(99.43%/78.60%)
7 Trigram+Lexicon
95.98%
(99.18%/78.38%)
94.94%
(99.26%/71.17%)
95.95%
(99.26%/77.70%)
8 Bigram+succ+Lexicon
96.36%
(99.43%/79.50%)
95.36%
(99.43%/72.97%)
96.50%
(99.43%/80.41%)
9 Trigram+succ+Lexicon
95.78%
(99.02%/77.93%)
95.08%
(99.06%/73.20%)
95.91%
(99.06%/78.60%)

TABLE 5. RESULT ON 21% OOV TESTING CORPUS
NO
Configuration
Affix Tree
Configuration
PREFIX
SUFFIX
PRE
-
SUFF
1 Baseline
85
.
92
%
(99.10% / 37.72%)
2
Bigram
93.27%
(99.18%/71.64%)
92.40%
(99.15%/67.71%)
93.01%
(99.18%/70.42%)
3 Trigram
93.12%
(99.29%/70.56%)
92.31%
(99.07%/67.57%)
93.12%
(99.22%/70.83%)
4 Bigram+succeeding POS
93.41%
(99.22%/72.18%)
92.54%
(99.15%/68.39%)
93.36%
(99.18%/72.05%)
5
Trigram+succeeding POS
93.04%
(99.18%/70.56%)
92.22%
(99.15%/66.89%)
93.07%
(99.18%/70.69%)
6 Bigram+Lexicon
93.82%
(99.18%/74.22%)
93.27%
(99.18%/71.64%)
93.97%
(99.18%/74.90%)
7 Trigram+Lexicon
93.59%
(99.26%/72.86%)
93.33%
(99.18%/71.91%)
94.00%
(99.22%/74.90%)
8 Bigram+succ+Lexicon
93.97%
(99.18%/74.90%)
93.56%
(99.18%/73.00%)
94.46%
(99.18%/77.20%)
9 Trigram+succ+Lexicon
93.59%
(99.15%/73.27%)
93.24%
(99.11%/71.78%)
94.06%
(99.07%/75.71%)
TABLE 6. RESULT ON 30% OOV TESTING CORPUS
NO
Configuration
Affix Tree
Configuration
PREFIX
SUFFIX
PRE
-
SUFF
1 Baseline
83
.
28
%
(99.01%/47.21%)
2
Bigram
89.32%
(99.05%/67.01%)
88.98%
(99.05%/65.88%)
89.84%
(99.10%/68.61%)
3 Trigram
89.32%
(99.01%/67.11%)
89.06%
(98.97%/66.35%)
89.78%
(99.01%/68.61%)
4 Bigram+succeeding POS
88.86%
(99.01%/65.60%)
88.35%
(99.05%/63.81%)
89.55%
(99.10%/67.67%)
5
Trigram+succe
eding POS
88.66%
(98.89%/65.22%)
88.09%
(98.85%/63.43%)
88.92%
(98.97%/65.88%)
6 Bigram+Lexicon
90.32%
(99.01%/70.41%)
90.15%
(99.05%/69.75%)
91.18%
(99.01%/73.23%)
7 Trigram+Lexicon
90.32%
(99.05%/70.31%)
90.12%
(99.01%/69.75%)
91.30%
(99.05%/73.52%)
8 Bigram+succ+Lexicon
90.04%
(98.93%/69.65%)
89.81%
(99.05%/68.61%)
91.01%
(99.01%/72.67%)
9 Trigram+succ+Lexicon
90.15%
(98.85%/70.22%)
89.87%
(98.97%/68.99%)
90.78%
(98.93%/72.10%)
Unfortunately, this kind of case is very rare in the testing
corpus. In addition, there are many chances that succeeding
POS tag is incorrectly tagged in the first pass, especially when
the succeeding word is OOV word or the OOV words appear
consecutively in the testing data. These all will lead to tag the

current word incorrectly and decrease the tagging accuracy.
Roth and Zelenko also did similar experiment and reported the
same thing with us [22]. They used a dictionary (the most
frequent POS tag for each word) to get succeeding POS tag.
They reported that about 2% of accuracy decrease is caused by
incorrectly attached POS tags by their method [13, 22].
IV.
CONCLUSIONS
Our research focused on developing Indonesian POS tagger
using Hidden Markov Model and evaluating the performance
of the system for each configuration. The result showed that
HMM based POS Tagger for Bahasa Indonesia depends on the
OOV handling method used (~99.4% accuracy for non-OOV
words). If the OOV handling method is good at disambiguating
OOV words, the overall performance is very high.
Affix tree and lexicon from KBBI – Kateglo is useful to
improve the tagging accuracy for tagger on Indonesian data,
especially for OOV words. Prefix is the best configuration for
the affix tree since it gives better accuracy than other
configurations in many cases. The best tagging accuracy for
OOV words achieved ~80%. The overall tagging accuracy for
each testing data achieved 91.30%(30% OOV), 94.46%(21%
OOV), and 96.50%(15% OOV).
On the other hand, succeeding POS tag does not give much
improvement on OOV handling. Many cases showed that
succeeding POS tag decreases the tagging accuracy. Maybe,
we have to use succeeding POS tag in other strategy because
there are some cases where succeeding POS tag helps the
tagging performance. For example, we can add some rules to
decide when to use the succeeding POS tag.
In our next research, we will develop a larger Indonesian
corpus. We will also conduct study on Indonesian POS tagger
with another approach to investigate the best method for
Indonesian POS tagger.
A
CKNOWLEDGMENT
We would like to thank Dr. Hammam Riza from BPPT –
Indonesia for allowing and giving access to his training data,
and also for his useful guidance.
R
EFERENCES
[1] Azimizadeh, Ali. Mehdi, Mohammad. Rahati, Saeid. 2008. "Persian part
of speech tagger based on Hidden Markov Model". JADT 2008 : 9es
Journées internationales d’Analyse statistique des Données Textuelles.
[2] Brants, Thorsten. 2000. "TnT - A Statistical Part-of-Speech Tagger".
Proceedings of the sixth conference on Applied Natural Language
Processing (2000) 224.231.
[3] Brill, E."A simple rule-based part-of-speech taggCher". In: Proceedings
of the Third Conference on Applied Natural Language Processing
(ANLP ’92), Trento, Italy (1992) 152–155.
[4] Chandrawati, Triastuti. 2008. "Indonesian Part-of-Speech Tagger based
on Conditional Random Fields and Transformation based learning
methods".Undergraduate thesis. Depok:Fasilkom University of
Indonesia.2008.
[5] Cutting, Doug, et al. A Practical Part-of-speech Tagger. Xerox Palo
Alto Research Center. In Proceding of the third conference on applied
Natural Language Processing page 133-140. 1992.
[6] Dandapat, Sandipan., Sarkar Sudeshna. Part-of-Speech Tagging for
Bengali with Hidden Markov Model. In Proceedings of the NLPAI
Machine Learning Contest. Mumbai, India, 2006.
[7] Femphy Pisceldo, Manurung, R., Adriani, Mirna. Probabilistic Part-of-
Speech Tagging for bahasa Indonesia. Third International MALINDO
Workshop, colocated event ACL-IJCNLP 2009, Singapore, August 1,
2009.
[8] G. Rubin. B. Greene .1971. "Automatic Grammatical Tagging of
English". Technical Report, Department of Linguistics, Brown
University, Providence, Rhode Island.
[9] Kamus Besar Bahasa Indonesia. http://www.pusatbahasa.diknas.go.id,
access date 12 April 2010.
[10] Kateglo. Dictionary, Thesaurus, and Glosarium for bahasa Indonesia.
http://www.bahtera.org/kateglo. access date 12 April 2010.
[11] Lewis, M. Paul (ed.). 2009. "Ethnologue: Languages of the World,
Sixteenth edition". Dallas, Tex : SIL International.
[12] Manurung, Ruli. Adriani, Mirna. 2008. "A survey of bahasa Indonesia
NLP research conducted at the University of Indonesia".Second
MALINDO Workshop. Selangor, Malaysia: 12-13 June 2008.
[13] Nakagawa, T., Kudo, T., & Matsumoto, Y. Unknown word guessing and
part-of- speech tagging using support vector machines. In Proceedings
of the Sixth Natural Language Processing Pacific Rim Symposium.
2001.
[14] Nguyen, Nam., Guo, Yunsong. 2007. Comparisons of Sequence
Labeling Algorithms and Extensions. International Conference on
Machine Learning.
[15] PANL10N1(PAN Localization Project), http://www.panl10n.net, access
date 12 April 2010.
[16] Padr´o M. and Padr´o L. Developing Competitive HMM POS Taggers
Using Small Training Corpora. Estal. 2004
.
[17] Pajarskaite, Giedre, et al,. Designing HMM based Part-of-Speech
Tagging for Lithuanian Language. INFORMATICA vol 15, no 2, page
231-242. 2004.
[18] Schmid, Helmut. Probabilistic Part-of-Speech Tagging using Decision
Tree. Proceedings of International Conference on New Methods in
Language Processing. September 1994.
[19] Schmid, Helmut. 1995. “Improvements in Part-of-Speech Tagging with
an Application to German”. Proceedings of the ACL SIGDAT-
Workshop. March 1995.
[20] Scott M.T. M.P. Harper. 1999. "A second-order Hidden Markov Model
for part-of-speech tagging". Proceedings of the 37th Annual Meeting of
the Association for Computational Linguistics. pp: 175-182.
[21] Wibisono, Yudi. HMM for Sentece Compression. Graduate Thesis.
Institute of Technology Bandung. 2008.
[22] D. Roth and D. Zelenko. 1998. Part-of-Speech Tagging Using a
Network of Liniear Separators. In Proceedings of the joint 17th
International Conference on Computational Linguistics and 36th Annual
Meeting of the Association for Computational Linguistics
(ACL/COLING-98),pages 1136-1142.
[23] Adriani, Mirna. Riza, Hammam. Local Language Computing:
Development of Indonesian Language Resources and Translation
System. 2008.
[24] Sari, Syandra, Herika Hayurani, Mirna Adriani, and Stephane Bressan.
Developing Part-of-Speech Tagger for Bahasa Indonesia Using Brill
Tagger. The International Second MALINDO Workshop, 2008.