Professional Documents
Culture Documents
net/publication/228856053
Article
CITATIONS READS
9 5,879
2 authors:
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Abbas Raza Ali on 13 November 2014.
15
Proceedings of the Conference on Language & Technology 2009
accent i.e. the accent that is used in Pakistan while Some sounds have multiple realizations in Urdu
speaking English. orthography e.g. /s/ can be realized as ث، س، صetc
Out-Of-Vocabulary problem is resolved using
statistical techniques by first aligning English so in this case only one most commonly used alphabet
orthography to pronunciation sequences. Optimal is chosen which is سin this case. Similar is the case
pronunciation of an unknown word is computed by
picking maximum probable pronunciation and then with /z/ as it can be realized as ظ، ذ، ز، ضand /t ̪/
passing it for the same transliteration process.
The architecture of the English to Urdu which can be realized either as تor ط.
Transliteration system is shown in figure 1. Vowels in Urdu are represented using diacritics
i.e. zair, zaber and paish and four letters alif, wao,
English Text choti yeh and bari yeh. Combination of diacritics with
consonants forms short vowels while diacritics
Converted to Load
combined with alif, wao, choti yeh and bari yeh, form
transliteration
and language long vowels [3]. Same vowel is represented
English OOV
model differently in orthography, depending on whether it
applying Computing exists word initially, medially or finally.
Short vowels occurring word initially use alif as
َ َ
place holder e.g. “urban” is transliterated to +, ار
Syllabification Optimal
Pronunciation
applying Sequence /’ər.bən/ but when they occur word medially they are
represented only by the diacritics e.g. “justly” is
Urduization َ
ِ ð
transliterated to .012 , /’ʤəsʈ.li/.
Converted
Short vowels when occur word finally are
Urdu Script transformed into their corresponding long vowel i.e.
zabar is converted to alif e.g. “Andorra” /Æ n d ɔ ɹ ʌ/
َ َ َ
is transliterated to ورا45 ا/æn.’ɖɔ.rɑ/, similarly zair is
Figure 1: Architecture of English-to-Urdu
transliteration converted to choti yeh and pesh is converted to wao.
Hence there is no one-to-one correspondence
between English and Urdu vowels in most of the
2. English to Urdu mapping
cases and an English vowel is transliterated using
multiple Urdu characters depending on whether it
CMU pronouncing dictionary (v 0.7a) is used to
occurs word initially, medially or finally as shown
acquire pronunciation of English words. The
table 2.
dictionary comprises of 125,000 English words and
Table 2: English vowels mapping to Urdu
their corresponding transcription in Arpabet. The
orthography
pronunciation provided is based on American accent
[11]. Urdu
Arpabet IPA
The phonemic inventory of English comprises of Initial Middle Final
َ َ
24 consonants and 15 vowels. The phonemic inventory AA ɑ آ ◌ا ◌ا
of Urdu comprises of 37 consonants and 16 vowels َ َ
(Appendix B). English consonants can be easily AE Æ اى ◌ى ے
mapped to Urdu consonants and there is one-to-one َ َ
correspondence between them in all cases. There are AY Aɪ .:ِ آ ◌;◌ ِا .: ◌ا
some sounds in English e.g. dental fricatives, /Θ/ and َ َ
AW Aʊ آؤ ◌اؤ ◌اؤ
/Ð/ which are non-existent in Urdu and hence they are َ َ َ
AO ɔ او ◌و ◌و
mapped to their closest counterpart i.e. dental stops /t ̪ʰ/
and /d̪/ respectively. OY ɔɪ ◌;آ
ِ ◌;وا
ِ =ا
َ َ
EH ɛ اى ◌ى ے
16
Proceedings of the Conference on Language & Technology 2009
َ َ َ َ َ َ
ɝ ار ◌ر ◌ر ʌ . s oʊ . ʃ i .
?@Aا ? اBCAا
ER Associate
ʌt
Eɪ اى ى ے َ
وى.â اب
EY
َ َ
ʌb.lɪ.vi. ِ
ِا ◌ِ ◌ِ
Oblivious D Eِ , ا َ
IH ɪ ʌs اس
I ِاى ِ◌ى ِ◌ى ِڈى.I, او
َ
IY
oʊ . b i . d i . ِ َ
Obedient , او
?54
OW Oʊ او و و ʌnt ?Jا
ُ ُ ُ
UH ʊ ا ◌ ◌
َ َ َ
AH ʌ ا ◌ ◌ا 3.2. Special case
17
Proceedings of the Conference on Language & Technology 2009
^
4. Out-Of-Vocabulary problem Pr = arg max p( Pr | En ) = arg max p( En | Pr ) p ( Pr ) (1)
Pr Pr
Out-Of-Vocabulary is a very common problem in The trigram language model p( Pri −1 .Pri .Pri+1 ) and
various systems like text-to-speech, machine bigram transliteration model p( En .Pr ) is combined to
translation, cross language information retrieval i i
(CLIR), etc. To resolve this problem, English phoneme maximize the pronunciation probability Pr .
to orthography alignment has to be found out
probabilistically to get one-to-one mapping between 4.2. Computing optimal pronunciation
them as shown in table 6, and then train those aligned sequence
sequence to get most probable pronunciation for an
unknown word. Expectation maximization algorithm is used to
compute optimal alignment sequence. The algorithm
Table 6: English orthography to pronunciation is given below;
alignment Initialization
English Percentages For each English phoneme to orthography pair, assign
Pronunciatio equal weights to all possibilities generated from (1).
p . er . s . eh . n . t . ih . jh . ah . z
n repeat
p(p) . er(er) . c(s) . e(eh) . n(n) . t(t) .
Alignment Expectation-Step
a(ih) . g(jh) . e(ah) . s(z)
For each of the Arpabet phonemes, count up
The entire procedure consists of two steps; instances of its different mappings from the
• English orthography to pronunciation alignment. observations on all combinations produced in (1).
• Computing optimal pronunciation sequence. Normalize the score so that the mapping
After getting pronunciation of unknown text, it will probabilities sum to 1.
be passed through the same procedure like Maximization-Step
syllabification and then Urdu transliteration. The Recalculate the combination scores. Each
architecture of the OOV module is shown in figure 2. combination is scored with the product of the scores
of the symbol mappings it contains. Normalize the
CMU pronunciation dictionary scores so that the mapping probabilities sum to 1.
until convergence
Syllabification
18
Proceedings of the Conference on Language & Technology 2009
The System’s accuracy is recorded after maturity University of Computer and Emerging Sciences
of every independent module as mentioned in figure 3. (NUCES), Pakistan.
The lexicon of most frequently used words of English
(15,237 words from British national corpus (BNC)) 8. References
was transliterated into Urdu using the transliteration
system. Accuracy without applying syllabification and [1] W. Gao., K. F. Wong and W. Lam. “Phoneme-based
resolving unknown word problem is described in table Transliteration of Foreign Names for OOV Problem”. In
7 in detail. The results are generated by passing First International Joint Conference on Natural Language
transliterated text to Urdu text-to-speech system and Processing, Pages 374-381, 2004.
analyzing its output.
[2] Saleem, M. “Urdu Rasmulkhat ki Jaamiat”. Akhbar-i-
Urdu, Pages 6-10, Islamabad, Pakistan, 2002.
Table 7: English-to-Urdu mapping accuracy
[3] S. Hussain, “Letter-to-Sound Rules for Urdu Text to
Observations Total Size Speech System”. Proceedings of Workshop on
Correct Mapping (after applying rules) 12,940 Computational Approaches to Arabic Script-based
Incorrect Mapping (due to Syllabification) 173 Language, COLING-2004, Geneva, Switzerland, 2004.
Incorrect Mapping (due to OOV) 2,124
Total 15,237 [4] S. Hussain, “Phonological Processing for Urdu Text to
Accuracy (%) 84.92 Speech System”. Yadava, Y, Bhattarai, G, Lohani, RR,
Prasain, B and Parajuli, K (eds.) Contemporary issues in
Nepalese linguistics. Katmandu, Linguistic Society of
After applying syllabification technique; out of Nepal, 2005.
173 syllabication problems, 91% are resolved (manual
testing). The accuracy of OOV is evaluated [5] J. Kominek, and A. W. Black, “Learning Pronunciation
automatically by using automatic evaluation method Dictionaries: Language Complexity and Word Selection
Bilingual Evaluation Understudy BLEU [10] as shown Strategies”. In Proceedings of the Human Language
in table 8. Technology Conference of the NAACL, Pages 232-239.
New York City, USA, 2006.
Table 8: Overall system accuracy
[6] J. Lewis, , K. McGrath, and J. Reuppel, “Language
Identification and Language Specific Letter-to-Sound
Modules Correct Total Size Accuracy (%) Rules”. Colorado Research in Linguistics, Volume 17,
Mapping 12,940 12,940 100.00 Issue 1, June 2004.
Syllabification 158 173 91.31
OOV 1,518 2,124 72.46 [7] J. Martin, , R. Mihalcea, and T. Pedersen, “Word
Total 14,616 15,237 95.92 Alignment for Languages with Scarce Resources”. In
Proceedings of the ACL Workshop on Building and
6. Conclusion Exploiting Parallel Texts: Data Driven Machine
Translation and Beyond, Ann Arbor, MI, June 2005
Transliteration is a good technique which helps a [8] A. Sen, “Pronunciation Rules for Indian English TTS
system adding multi-lingual ability. It can be used in System”. Workshop on Spoken Language Processing,
various Systems, e.g. text-to-speech, information Mumbai, India, January 2003
retrieval, machine translation, English-to-Urdu parallel
corpus Consistency in Proper Names etc. Overall [9] R. Bokhari, and S. Pervez, “Syllabification and Re-
system’s accuracy is 96% which is quite promising. Syllabification in Urdu”. Akhbar-i-Urdu, Pages 63-67,
The System can be improved by training transliteration Islamabad, Pakistan, 2003.
model on Urdu accent instead of American.
[10] K. Papineni, S. Roukos, , T. Ward, , and W. J. Zhu,
“Bleu: a Method for Automatic Evaluation of Machine
7. Acknowledgements Translation”. Proceedings of the International Conference
on Spoken Language Processing (ICSLP), Pages 901–904,
The work on English to Urdu transliteration 2002.
system has been carried out in a project that involves
[11] CMU. “The CMU Pronunciation Dictionary”,
development of an open-source Urdu screen reader for www.speech.cs.cmu.edu/cgi-bin/cmudict, School of
visually impaired people funded by National Computer Science, Carnegie Mellon University, Pittsburgh,
USA, 2006.
19
Proceedings of the Conference on Language & Technology 2009
Urdu Urdu
Arpabet IPA Arpabet IPA
Initial Middle Final Initial Middle Final
َ َ
AA ɑ آ ◌ا ◌ا L L ل ل ل
َ َ
AE Æ اى ◌ى ے M M م م م
َ َ َ
AH ʌ ا ◌ ◌ا N N ن ن ن
َ َ َ
AO ɔ او ◌و ◌و NG Ŋ UJ UJ UJ
َ َ
AW Aʊ آؤ ◌اؤ ◌اؤ OW Oʊ او و و
َ َ
AY Aɪ .:ِ آ ◌;◌ ِا .: ◌ا OY ɔɪ ◌;آ
ِ ◌;وا
ِ =ا
20
Proceedings of the Conference on Language & Technology 2009
Vowels
Consonants
21
Proceedings of the Conference on Language & Technology 2009
Vowels
/i/
/ʊ/
/u/
/e/
/ɛ/ | /ɑɪ/
/o/
/ɑu/
Consonants
22
Proceedings of the Conference on Language & Technology 2009
Appendix D - Transliteration
23