💬 NLP

Recent Advancements and Challenges of Turkic Central Asian Language Processing

This paper reviews the current state of Natural Language Processing for Central Asian Turkic languages (Kazakh, Uzbek, Kyrgyz, and Turkmen), highlighting recent progress in dataset collection and model development while addressing persistent low-resource challenges and outlining future research directions.

Original authors: Yana Veitsman, Mareike Hartmann

Published 2026-02-17

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Yana Veitsman, Mareike Hartmann

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the world of language technology (like Siri, Google Translate, or spell-checkers) as a giant, bustling library. In this library, some languages are like best-selling blockbusters with thousands of copies, movie adaptations, and fan clubs. English is the ultimate blockbuster.

Then, there are the Turkic languages of Central Asia (Kazakh, Uzbek, Kyrgyz, and Turkmen). In this library, they are like rare, hand-written manuscripts. They are spoken by about 80 million people, but they have very few copies, no movie adaptations, and the librarians (the AI researchers) are struggling to find enough pages to teach the computers how to read them.

This paper is a status report on these rare manuscripts. It asks: What do we have? What is missing? And how can we build a better library for these languages?

Here is the breakdown of their findings, using some everyday analogies:

1. The "Family Resemblance" (Linguistic Features)

Think of these four languages as siblings in a large family. They all speak with the same accent and follow the same family rules (grammar).

The Rules: They all put the verb at the end of the sentence (like saying "I the apple eat" instead of "I eat the apple").
The Complexity: They are "agglutinative." Imagine building with LEGO bricks. In English, you might have a separate word for "un-happi-ness." In these languages, you snap many small bricks (suffixes) onto one big block (the root word) to change the meaning. This makes it hard for computers because they have to figure out how to snap the bricks apart to understand the meaning.
The Script Issue: Some siblings write in Latin letters (like Uzbek), while others still use Cyrillic (Russian letters, like Kazakh). It's like trying to teach a child to read when one sibling uses a red marker and the other uses a blue one. It confuses the computer.

2. The "Resource Gap" (Data Availability)

The authors looked at how much "training data" (text, audio, and images) exists for each language. They ranked them like a video game leaderboard:

🥇 Kazakh (The "Rising Star"): This language has the most data. It's like a growing startup that just got a big investment. They have huge libraries of text, audio recordings, and even datasets for sign language and handwriting. They are close to having a full toolkit.
🥈 Uzbek (The "Hopeful"): Uzbek is the runner-up. They have a decent amount of data, especially for things like analyzing feelings in reviews (sentiment analysis), but they are still playing catch-up to Kazakh.
🥉 Kyrgyz & Turkmen (The "Stragglers"): These two are in trouble. It's like trying to bake a cake with only a few crumbs of flour. There is very little data available. For Turkmen, it's almost non-existent. Researchers are mostly just "scraping" (scooping up) whatever random text they can find on the internet, which isn't always high quality.

3. Why is the Library So Empty? (Reasons for Scarcity)

Why don't these languages have more data? The paper points to three main culprits:

The Russian Shadow: For decades, Russian was the dominant language in schools, government, and media in this region. It's like if everyone in a town spoke English at the dinner table, even though their native language was something else. People didn't write much in their native tongues online, so there's no digital history to learn from.
The Internet Gap: You can't build a digital library if people can't get online. In some of these countries, internet access is still limited. If you can't get online, you can't contribute to Wikipedia, write a blog, or upload a video.
The Funding Void: Building AI is expensive. It requires money and specialized schools. While there are some tech hubs popping up, there aren't enough government programs dedicated specifically to teaching computers these languages.

4. The "Cheat Codes" (How Researchers are Fixing It)

Since they can't just wait for millions of people to start typing in these languages, researchers are using clever tricks:

Transfer Learning (The "Big Brother" Strategy): Since Turkish (a related language) has a massive library of data, researchers are teaching the computer using Turkish first, then asking it to "translate" that knowledge to Kazakh or Uzbek. It's like teaching a child to drive a truck by letting them practice in a smaller car first.
Data Augmentation (The "Photocopier"): If you have a small pile of text, you can use AI to rewrite it in different ways to make it look like you have more. It's like taking one photo and using filters to create 100 slightly different versions to train the computer.
Transliteration: Since some languages use different scripts (Latin vs. Cyrillic), researchers are teaching the computer to read both as if they were the same language, smoothing over the script differences.

5. The Current Tech Landscape

Kazakh: Has working tools for speech recognition (listening), translation, and finding names in text. It's the most advanced.
Uzbek: Has some speech tools and translation, but they aren't as polished as Kazakh's yet.
Kyrgyz & Turkmen: Mostly have basic tools. If you ask a computer to translate a sentence or recognize speech in these languages today, it will likely give you a nonsense answer.

The Bottom Line

The paper concludes that while we have made great progress with Kazakh, the other three languages are still in the "construction zone."

To fix this, we need:

More Data: People need to write, speak, and post online in their native languages.
Better Tools: We need to stop treating these languages as "afterthoughts" and build specific models for them.
Teamwork: Since these languages are so similar, helping one (like Kazakh) can actually help the others (like Kyrgyz) through those "Big Brother" transfer learning tricks.

In short: The technology exists to speak these languages, but right now, the computers are mostly mute. We need to feed them more data and teach them the family rules so they can finally join the conversation.

1. Problem Statement

The paper addresses the significant data scarcity and resource limitations facing Natural Language Processing (NLP) for four Central Asian Turkic languages: Kazakh, Uzbek, Kyrgyz, and Turkmen. While these languages are spoken by approximately 80 million people, they suffer from:

Lack of high-quality, annotated datasets for downstream tasks (e.g., Named Entity Recognition, Machine Translation).
Inconsistent orthography (e.g., Uzbek uses Latin script while Kazakh remains primarily Cyrillic), hindering cross-lingual transfer learning.
Morphological complexity, specifically their highly agglutinative nature, which complicates tokenization and machine translation.
Socio-technical barriers, including limited internet penetration in the region, the historical dominance of Russian (limiting local digital content), and a lack of dedicated AI initiatives.

2. Methodology

The authors conducted a comprehensive survey of the current state of NLP resources and technologies for these languages. Their methodology involved:

Literature Review: Systematic search of academic venues (ACL, EMNLP, LREC, IEEE), pre-print servers (arXiv), and code repositories (GitHub, Hugging Face) up to November 2024.
Linguistic Typology Analysis: Comparison of grammatical features (vowel harmony, case systems, word order) across the four languages and Turkish to assess potential for transfer learning.
Resource Inventory: Categorization of available datasets (unannotated corpora, linguistic features, task-specific, multimodal) and evaluation of their quality and accessibility.
Technology Assessment: Review of existing models (Rule-based, Statistical, and Deep Learning/LLMs) for tasks like Automatic Speech Recognition (ASR), Machine Translation (MT), and Text Classification.
Classification Framework: Application of the "Joshi et al. (2020)" framework to categorize languages based on resource availability (e.g., "The Rising Star," "The Scraping-Bys").

3. Key Contributions

A. Linguistic and Typological Insights

Agglutinative Challenges: The paper highlights that the high number of suffixes (e.g., Kazakh has 12 plural suffixes vs. Turkish's 2) and vowel harmony rules create unique challenges for tokenization and morphological analysis.
Transfer Learning Potential: While Turkish is the most resource-rich Turkic language, the paper argues that intra-regional transfer (e.g., Kazakh $\to$ Kyrgyz) may be more effective than Turkish $\to$ Central Asian due to closer grammatical proximity and shared script issues (Cyrillic vs. Latin).

B. Comprehensive Resource Inventory

The authors provide the first unified overview of data availability, revealing a stark disparity:

Kazakh ("The Rising Star"): Possesses the most diverse resources, including large-scale speech corpora (KSC2, 1,200+ hours), multimodal datasets (sign language, emotional TTS), and task-specific datasets for NER, QA, and Sentiment Analysis.
Uzbek ("The Hopeful"): Second in resources, with strong datasets for sentiment analysis and text classification, but lagging in speech and multimodal data compared to Kazakh.
Kyrgyz & Turkmen ("The Scraping-Bys"): Severely under-resourced. Kyrgyz has some news corpora and handwritten datasets, while Turkmen relies almost entirely on web-scraped data and small poetry collections.

C. Technology Landscape Assessment

Kazakh: Leads in technology with state-of-the-art ASR (Word Error Rate ~7.2%), custom MT systems (Tilmash) outperforming Google/Yandex, and fine-tuned models (XLM-RoBERTa) achieving high precision in NER.
Uzbek: Has pre-trained models (UzBERT, BERTBek) and functional ASR (WER ~14.3%), but lacks advanced downstream task models.
Kyrgyz & Turkmen: Technology is nascent. Most efforts are limited to basic text classification or comparative studies; no dedicated pre-trained models or robust ASR systems exist.

D. Identification of Data Scarcity Drivers

The paper identifies three root causes for the resource gap:

Russian Dominance: High reliance on Russian for education and media limits the generation of native digital content.
Internet Access: Low connectivity in Kyrgyzstan (38%) and Turkmenistan (21%) restricts user-generated content (blogs, wikis).
Funding Gaps: Lack of government-funded, NLP-specific initiatives; existing AI institutes cover broad topics (robotics, IoT) rather than focusing on linguistics.

4. Results and Findings

Data Imbalance: There is a massive gap between Kazakh and the other three languages. Kazakh has over 1,200 hours of speech data, whereas Turkmen has virtually none.
Script Inconsistency: The mix of Cyrillic (Kazakh, Kyrgyz) and Latin (Uzbek, Turkmen) scripts creates a barrier for unified model training, requiring complex preprocessing or transliteration strategies.
Effectiveness of Augmentation: Data augmentation (using LLMs for sentence generation) and R-Drop regularization show promise but are less effective in extremely low-resource settings compared to high-resource languages.
Transfer Learning Success: Studies show that transferring from Kazakh to Kyrgyz is more viable than from Turkish, suggesting a "Central Asian cluster" approach for model training.

5. Significance and Future Directions

Foundation for Future Research: This paper serves as a critical roadmap, identifying exactly where data is missing and which tasks are feasible.
Strategic Recommendations:
- For Kazakh: Focus on advancing text generation and Question Answering (QA) using existing multimodal data.
- For Uzbek: Prioritize large-scale data collection to match Kazakh's infrastructure.
- For Kyrgyz/Turkmen: Immediate focus must be on data collection and aggregation. Pre-training large models is currently infeasible; researchers should focus on transfer learning from Kazakh and utilizing statistical algorithms.
Policy Implications: The authors emphasize that achieving "high-resource" status requires government-backed initiatives to digitize local content and improve internet access, not just academic efforts.

In conclusion, the paper establishes that while Kazakh is rapidly approaching a state-of-the-art NLP ecosystem, Uzbek, Kyrgyz, and Turkmen remain in the early stages, requiring urgent, coordinated efforts in data collection and infrastructure development to prevent further digital marginalization.