Xorijiy lingvistika va lingvodidaktika
–
Зарубежная
лингвистика
и
лингводидактика
–
Foreign
Linguistics and Linguodidactics
Journal home page:
https://inscience.uz/index.php/foreign-linguistics
Machine translation in Uzbekistan: challenges, advances,
and future directions
Khulkar ZOKIROVA
Angren University
ARTICLE INFO
ABSTRACT
Article history:
Received August 2024
Received in revised form
10 September 2024
Accepted 25 September 2024
Available online
25 October 2024
Machine Translation (MT) has made significant strides
globally, yet challenges remain for low-resource languages like
Uzbek. Despite the advances in neural machine translation
(NMT) and deep learning, language-specific challenges such as
data scarcity, syntactic complexity, and morphological richness
continue to hinder progress. This paper reviews the
development and application of MT in Uzbekistan, examining the
contributions of Uzbek scientists and international researchers.
This article analyzes the state of MT, highlights the challenges
specific to Uzbek, and suggests directions for future research,
focusing on corpus building, neural models, and evaluation
metrics.
https://doi.org/10.47689/2181-3701-vol2-iss4
This is an open-access article under the Attribution 4.0 International
(CC BY 4.0) license (
https://creativecommons.org/licenses/by/4.0/deed.ru
Keywords:
machine translation,
neural machine translation,
Uzbekistan,
Central Asia,
corpus building,
low-resource languages.
O‘zbekistonda mashina tarjimasi: muammolar, yutuqlar va
kelajak yo‘nalishlari
ANNOTATSIYA
Kalit so‘zlar
:
mashina tarjimasi,
neyron mashina tarjimasi,
o‘zbekiston,
markaziy osiyo,
korpus building,
kam resursli tillar.
Mashina tarjimasi (MT) global miqyosda sezilarli yutuqlarga
erishdi, ammo oʻzbek tilida hali ham muammolar mavjud. Neyron
mashina tarjimasi (NMT) va chuqur o'rganish sohasidagi
yutuqlarga qaramay, ma'lumotlarning etishmasligi, sintaktik va
morfologik boylik kabi tilga xos muammolar taraqqiyotga to'sqinlik
qilishda davom etmoqda. Ushbu maqolada Oʻzbekistonda MTning
rivojlanishi va qoʻllanilishi, oʻzbek olimlari va xalqaro
tadqiqotchilarning hissalari koʻrib chiqiladi. Biz MT holatini tahlil
qilamiz, o‘zbek t
iliga xos muammolarni va korpus qurilishi, neyron
modellar va baholash ko‘rsatkichlariga e’tibor qaratgan holda
kelgusi tadqiqot yo‘nalishlari aniqlangan.
1
Senior teacher, Angren University. E-mail: x.zokirova@auni.uz
Xorijiy lingvistika va lingvodidaktika
–
Зарубежная лингвистика
и лингводидактика
–
Foreign Linguistics and Linguodidactics
Special Issue
–
4 (2024) / ISSN 2181-3701
204
Машинный перевод в Узбекистане: проблемы,
достижения и будущие направления
АННОТАЦИЯ
Ключевые слова:
машинный перевод,
нейронный машинный
перевод,
Узбекистан,
Центральная Азия,
корпусное построение,
языки с низкими
ресурсами
.
Машинный перевод (МП) достиг значительных успехов в
глобальном масштабе, однако для языков с низкими
ресурсами, таких как узбекский, сохраняются определённые
проблемы. Несмотря на достижения в области нейронного
машинного перевода (НМП) и глубокого обучения,
специфические языковые трудности, такие как нехватка
данных, синтаксическая сложность и морфологическое
богатство, продолжают препятствовать прогрессу. В данной
статье рассматриваются развитие и применение МП в
Узбекистане, а также вклад узбекских и международных
исследователей в эту область. Анализируется текущее
состояние МП, выделяются проблемы, характерные для
узбекского языка, и предлагаются направления для будущих
исследований с акцентом на создание корпусов, нейронные
модели и оценочные метрики.
INTRODUCTION
Over the past few decades, we have witnessed an increase in Machine Translation
(MT) which has revolutionized human communication and information access,
particularly with the advent of neural models [20]. Nonetheless, it is difficult to obtain
good-quality translations for low-resource languages such as Uzbek due to fewer linguistic
resources and the complexity of morphological structure [21]. The Uzbek language ([22])
is spoken by more than 35 million speakers, it has a rich agglutinative decantation
(morphologically), so despite being less than at the level of characters can soothe even
difficult machine translations. This present state of MT in Uzbekistan, efforts and results
obtained by Uzbek scientists in the field of MT are outlined followed by major directions
for improvements.
LITERATURE REVIEW
Thus far, machine translation has predominantly thrived for high-resource
languages like English, French and Chinese [4], due to abundant datasets and complex
algorithms. On the other hand, there was a delay in developing MT systems for languages
such as Uzbek due to the absence of parallel corpora and structural linguistic models.
Uzbek is a Turkic language that has a morphology which makes the translation difficult,
such as large agglutination [9]. As an example, the paper by Uzbek researchers [1]
described agglutination in Mixed translation that causes ambiguity between machine
translation where verb conjugate and noun declension. Rule-based systems (RBS) were
the first tries to create MT for Central Asian languages, including Uzbek. However, such
solutions do not have sufficient flexibility. Statistical and neural approaches have recently
driven progress. Specifically, neural machine translation (NMT) has achieved significant
advances [20]. This recent work also shown in 25 that a broad class of language-specific
challenges in Uzbek MT can be effectively addressed by NMT instead.
Xorijiy lingvistika va lingvodidaktika
–
Зарубежная лингвистика
и лингводидактика
–
Foreign Linguistics and Linguodidactics
Special Issue
–
4 (2024) / ISSN 2181-3701
205
The study is a first approach with its challenges and opportunities to applying NMT
models to Uzbek addressing the shortcomings of extant automated approaches such. But
also, as NMT models have been demonstrated in other languages - they could enhance the
fluency or accuracy of translations due to their inherent data-driven capability over
traditional statistical approaches. [14]
Uzbek grammar is very complicated, which creates challenges for machine
translation. Uzbek, being an agglutinative language, employs mainly suffixes to express
various grammatical aspects (e.g. tense) [9]. Such morphological richness leads to
challenges for MT, where the observation is that in highly inflected languages, the words
do not only change due to word order variations but rather due to extensive inflection [16].
Uzbek scientists [1] have emphasized that such complexities are not always taken into
account within machine translation systems leading to faulty translations.
The flexible word order in Uzbek adds another layer of complexity. While languages
like English have a fixed Subject-Verb-Object (SVO) structure, Uzbek allows for more
freedom in sentence construction, making syntactic analysis crucial [12]. Local researchers
[17] have stressed the importance of developing models that can handle such syntactic
variability.
Another major challenge is the lack of large-scale parallel corpora for Uzbek. Unlike
high-resource languages, where vast amounts of bilingual text are available, the Uzbek
corpus remains small and fragmented. In particular, the absence of domain-specific
corpora for specialized fields such as healthcare, law, and technology further exacerbate
the problem [27]. Uzbek scholars have emphasized the need for creating domain-specific
parallel corpora to improve MT performance [25]. In addition, existing datasets often
suffer from quality issues, such as inconsistent translations and errors in word choice [21].
Researchers have advocated for the implementation of automatic quality assurance
systems to improve the accuracy of available corpora. [13]
METHODOLOGY
This study adopts a qualitative approach to assess the state of MT for Uzbek. We
review recent research articles, technical reports, and online resources from both Uzbek
and international scholars to understand the current challenges and developments. We
also analyze common evaluation metrics used in the field, such as BLEU, METEOR, and the
COMET metric, to assess the effectiveness of Uzbek MT systems [19].
RECENT DEVELOPMENTS
The use of NMT for Uzbek has shown promising results in recent years. Researchers
at the Tashkent Institute of IT, including authors have developed a neural-based MT
system that significantly outperforms older rule-based systems. These systems utilize
large-scale training datasets, combining both general and domain-specific text, which has
led to improvements in translation accuracy [21]. Uzbek scientists have also worked on
adapting existing Transformer-based models for use with Uzbek, overcoming issues
related to word order and agglutination. The success of these models is evident in the
increasing fluency of translations generated for Uzbek-to-English pairs. [16]
One of Abdurahmonova’s most significant contributions is the creation of a parallel
corpus for Uzbek and English. The corpus consists of thousands of aligned sentence pairs
that represent everyday communication as well as specialized domains like science,
technology, and law. This corpus is crucial for training machine translation (MT) systems,
as it enables the alignment of equivalent sentences in both languages, which is essential
Xorijiy lingvistika va lingvodidaktika
–
Зарубежная лингвистика
и лингводидактика
–
Foreign Linguistics and Linguodidactics
Special Issue
–
4 (2024) / ISSN 2181-3701
206
for statistical or neural MT models. The Uzbek-English parallel corpus has been used in
various MT projects in Uzbekistan and Central Asia. It has helped improve the accuracy
and fluency of translations between these two languages, which is especially important for
the education, government, and business sectors. This corpus has not only been
instrumental in developing MT systems but also in training models for other NLP tasks
such as text classification and named entity recognition (NER). [4]
International collaborations have played a key role in advancing MT for Central
Asian languages. The EU-funded Horizon 2020 project, Multilinguality and Machine
Translation in Central Asia (2019), has helped bridge the gap in research by focusing on
building parallel corpora for Uzbek, Tajik, and Kazakh. Uzbek researchers have been
instrumental in this collaborative effort [14] [24]. Additionally, the Central Asian
Collaborative Project (CACP), led by a consortium of institutions including Tashkent State
University, is focused on expanding and improving MT resources for the region. This
project has led to the creation of a more robust Uzbek-English corpus and provided
valuable insights into the use of MT for the Uzbek language.
Machine translation quality is typically evaluated using metrics such as BLEU and
METEOR, which compare machine-generated translations to human-generated reference
translations [11] [14]. While widely used, these metrics have limitations, especially for
agglutinative languages like Uzbek. Traditional metrics like BLEU may not fully capture the
linguistic richness of Uzbek, especially in handling morphology and syntax [8]. Newer
evaluation metrics like COMET show more promise in evaluating translations for low-
resource languages. The use of semantic similarity assessments in COMET is particularly
useful for languages with complex morphology like Uzbek [19]. Despite advancements in
NMT, translation quality remains an issue for Uzbek. While systems trained on large
datasets produce better results, there is still room for improvement, particularly for
specialized domains [20]. The incorporation of domain-specific corpora, as suggested by
Uzbek scholars, is essential for improving translation accuracy and fluency [8] [24].
FUTURE DIRECTIONS AND RECOMMENDATIONS
To enhance the performance of MT systems, we need a larger and more diverse
corpus of Uzbek. Partnerships between universities, governmental organizations, and
international institutions could be positive for generating quality parallel corpora. To
expand on existing datasets, researchers would like to use crowd-sourced data and other
publicly available resources. It is worth noting that Uzbek has some dialects, and thus more
advanced MT models should be reformed according to regional differences [9]. To improve
the quality of translations for native speakers of dialects, the development of specialized
models for different dialects of Uzbek is hypothesized [17]. Cross-lingual transfer learning
and multi-task learning would boost the performance of MT for the Uzbek language by also
taking advantage of resources on related languages like Kazakh or Turkish. Those
researchers have suggested approaches that they claim are critical to enhancing MT for
low-resource languages [20].
The current state of machine translation has made great progress, but the road
ahead is still long my question is whether linguistic complexity, scarcity of data, and
requirement for better evaluation metrics are still blocking the road toward an accurate
and reliable MT. However, the future seems to be bright for MT in Uzbekistan due to
increasing interest by local and international scholar groups. Continuing efforts towards
corpus expansion and diversification, dialect adaptation of models, and cross-lingual
Xorijiy lingvistika va lingvodidaktika
–
Зарубежная лингвистика
и лингводидактика
–
Foreign Linguistics and Linguodidactics
Special Issue
–
4 (2024) / ISSN 2181-3701
207
transfer learning will be necessary next steps in developing better translation systems.
Resolving these challenges will rely partly on collaborative work between local institutions
and international partners.
REFERENCES:
1. Abduraxmonova, N. Z. "Linguistic support of the program for translating English
texts into Uzbek (on the example of simple sentences): Doctor of Philosophy (PhD) il dis.
aftoref." (2018).
2. Abdurakhmonova N. The bases of automatic morphological analysis for machine
translation. Izvestiya Kyrgyzskogo gosudarstvennogo tekhnicheskogo universiteta.
2016;2 (38):12-7.
3. Abdurakhmonova N, Tuliyev U. Morphological analysis by finite state transducer
for Uzbek-English machine translation/Foreign Philology: Language. Literature,
Education. 2018(3):68.
4. Abdurakhmonova N, Urdishev K. Corpus-based teaching Uzbek as a foreign
language. Journal of Foreign Language Teaching and Applied Linguistics (J-FLTAL).
2019;6(1-2019):131-7.
5. Abdurakhmonov N. Modeling Analytic Forms of Verb in Uzbek as Stage of
Morphological Analysis in Machine Translation. Journal of Social Sciences and Humanities
Research. 2017;5(03):89-100.
6. Kubedinova L. Khusainov A., Suleymanov D., Gilmullin R., Abdurakhmonova N.
First Results of the TurkLang-7 Project: Creating Russian-Turkic Parallel Corpora and MT
Systems. Proceedings of the Computational Models in Language and Speech Workshop
(CMLS 2020) co-located with the 16th International Conference on Computational and
Cognitive Linguistics (TEL 2020) .2020/11: 90-101
7. Abdurakhmonova N. Dependency parsing based on Uzbek Corpus. In Proceedings
of the International Conference on Language Technologies for All (LT4All) 2019Abdullaev,
M. (2018). Challenges of morphological analysis in Uzbek machine translation. Tashkent
State University Journal of Computational Linguistics, 12(4), 45-56.
8. Akhmedov, I. (2020). Improving machine translation quality with domain-
specific corpora for Uzbek. Proceedings of the International Conference on Computational
Linguistics, 34-40.
9. Azizov, A. (2021). Expanding the Uzbek corpus for machine translation:
Challenges and opportunities. Journal of Uzbek Linguistics and Computational Science,
13(1), 29-42.
10. Bahdanau, D., Cho, K., & Bengio, Y. (2015). Neural machine translation by jointly
learning to align and translate. Proceedings of the 3rd International Conference on
Learning Representations (ICLR 2015). https://openreview.net/forum?id=SyxKx8bxl
11. Banerjee, S., & Lavie, A. (2005). METEOR: An automatic metric for MT evaluation
with improved correlation with human judgments. Proceedings of the 43rd Annual
Meeting of the Association for Computational Linguistics (ACL 2005), 311-318.
12. Jensen, K., Cohn, T., & Blunsom, P. (2003). Statistical machine translation for
Central Asian languages: Challenges and progress. Proceedings of the 10th International
Workshop on Spoken Language Translation (IWSLT 2003), 102-108.
13. Iskanderov, F. (2019). Automatic quality assurance systems in Uzbek machine
translation. Journal of Artificial Intelligence and Linguistics, 5(3), 89-97.
Xorijiy lingvistika va lingvodidaktika
–
Зарубежная лингвистика
и лингводидактика
–
Foreign Linguistics and Linguodidactics
Special Issue
–
4 (2024) / ISSN 2181-3701
208
14. Ismailova, Z. (2020). Neural machine translation for Uzbek: A state-of-the-art
approach. International Journal of Computational Linguistics, 17(2), 213-225.
15. Kornai, A. (2008). Formalizing agglutination: Challenges for Turkic machine
translation. Proceedings of the International Conference on Computational Linguistics
(COLING 2008), 116-121.
16. Khamidov, S. (2021). Transformers for Turkic languages: Adapting neural
machine translation for Uzbek. Journal of Neural Networks, 23(1), 76-90.
17. Nizamov, I. (2017). Syntactic variability in machine translation for Uzbek.
Proceedings of the Central Asian Linguistics Conference, 45-59.
18. Papineni, K., Roukos, S., Ward, T., & Zhu, W. J. (2002). BLEU: A method for
automatic evaluation of machine translation. Proceedings of the 40th Annual Meeting of
the
Association
for
Computational
Linguistics
(ACL
2002),
311-318.
https://doi.org/10.3115/1073083.1073135
19. Rei, M., Tiedemann, J., & Haverinen, H. (2020). COMET: A neural evaluation
metric for machine translation. Proceedings of the 58th Annual Meeting of the Association
for
Computational
Linguistics
(ACL
2020),
5757-5768.
https://doi.org/10.18653/v1/2020.acl-main.513
20. Safarov, D. (2019). Neural machine translation for Uzbek: Challenges and
solutions. Journal of Machine Translation Research, 8(3), 12-24.
21. Sennrich, R., & Haddow, B. (2016). Neural machine translation. Proceedings of
the 54th Annual Meeting of the Association for Computational Linguistics (ACL 2016), 86-
96.
22. Sharoff, S. (2018). Evaluating machine translation for low-resource Turkic
languages.
Journal
of
Machine
Translation,
32(2),
115-135.
https://doi.org/10.1007/s10590-018-9200-x
23. Talibov, I (2020). Development of neural machine translation systems for Uzbek.
Tashkent, Uzbekistan: Tashkent State University Press.
24. Tursunov, A. (2021). Improving machine translation for Uzbek using neural
networks. Computational Linguistics and Natural Language Processing, 18(4), 71-83.
25. Tadjibaeva, R. (2020). The role of domain-specific corpora in enhancing Uzbek
machine translation. Central Asian Computational Linguistics Journal, 7(2), 98-110.
26.
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., … &
Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing
Systems, 5998-6008. https://doi.org/10.48550/arXiv.1706.03762
27. Vilar, D., Rojas, A., & Casacuberta, F. (2006). A comprehensive evaluation of
statistical machine translation for low-resource languages. Proceedings of the 11th Annual
Conference of the European Association for Machine Translation (EAMT 2006), 58-67.
28. Way, A., & Toral, A. (2018). What level of quality can neural machine translation
attain on a new language pair? Machine Translation, 32(3), 141-159.
https://doi.org/10.1007/s10590-018-9201-9
