Towards the Roof of the World: NLP Progress in Tibetan Languages

AI summary

Overview

• Research area: Natural Language Processing (NLP) — specifically a bibliometric survey of Tibetan-language NLP research.
• Technical level: Intermediate. The paper is readable without deep technical background, but familiarity with NLP tasks (segmentation, machine translation, LLMs, benchmarks) helps.
• Scope (one sentence): A systematic bibliometric survey of 140 Tibetan NLP publications from 2004 to 2026, drawn from the ACL Anthology and DBLP, that maps research trends, research areas, contributing countries, and resource release practices.

What This Paper Is About

Tibetan, spoken by some seven million people on and around the Qinghai-Xizang Plateau, remains underrepresented in NLP. Its research literature is fragmented across communities and venues, and no systematic peer-reviewed survey of the field existed. The authors address this by building and annotating a corpus of 140 Tibetan NLP publications, then analyzing how the field has evolved, which areas are served and underserved, and what stands in the way of progress in the large language model (LLM) era.

Key Contributions

1. A first-of-its-kind bibliometric survey. The authors present what they describe as, to their knowledge, the first peer-reviewed systematic bibliometric survey of Tibetan NLP, covering 140 publications from 2004 through 00:00 UTC, September 25, 2026.

2. A hybrid, annotated corpus. Publications were retrieved from the ACL Anthology (37 papers: 25 conference, 12 workshop) and from DBLP (103 papers: 29 conference, 74 journal), then annotated for venue type, first-author affiliation country, research area (using up to two of the 29 ACL Rolling Review areas), and whether authors release data, code, or models.

3. The "full-pipeline paradigm" concept. The authors characterize Tibetan NLP as having grown bottom-up along the entire NLP pipeline — from corpora and word segmentation, through script and speech processing, to applications and, since 2022, pre-trained and large language models — while continuing to build resources, and they propose this as a reference model for other low-resource languages of the Global South.

4. An audit of released resources and a research agenda. The survey records that only 41 of 140 publications (29.3%) release data, code, or models at a public URL, and it closes with recommendations on open science, shared standards, native evaluation, community participation, and responsible NLP. The full bibliography is released at https://github.com/metaphors/TibetanNLPSurvey.

Main Findings

• Rapid but recent growth: Output was sporadic for nearly two decades — three papers in 2004–2006, none in 2007–2009, and 19 in 2010–2017 — before rising sharply. 118 publications (84.3%) date from 2018 onward, and 61 (43.6%) from 2024–2026 alone (2026 only partially covered).

• Venue clustering drives annual peaks: Five of the nine papers of 2018 appeared at PRCV 2018, seven of the 13 of 2019 in IEEE Access, eight of the ten of 2021 in TALLIP (six in one special issue), and five of the 21 of 2026 in one issue of Data Intelligence. Journal articles are the largest share (74; 52.9%), followed by conference papers (54; 38.6%) and workshop papers (12; 8.6%), spread across 10 workshops, 22 conferences, and 25 journals. TALLIP (19 articles) and IEEE Access (14 articles) lead.

• Most work sits outside the ACL Anthology: 103 of 140 publications (74%) lie outside it. Only 37 publications (26.4%) appear in the Anthology, but 17 of these date from 2024–2026, including ACL, EMNLP, AACL, COLING, and LREC.

• Skewed research areas: The 255 area assignments cover 25 of the 29 ARR areas. Multilingualism and Cross-Lingual NLP (41 publications) and Resources and Evaluation (40) account for 31.8% of assignments; the five largest areas account for 56.1%. Eleven areas have one to four publications, and four have none: Dialogue and Interactive Systems; Ethics, Bias, and Fairness; LLM Agents; and NLP and Code Models.

• Layered, bottom-up expansion: Resources and Evaluation appeared in 2004 and remained active in 16 of the 20 years with any Tibetan NLP publication; segmentation, script, speech, and cross-lingual work appeared in 2010–2012; web mining in 2014; language modelling, machine translation, question answering, generation, and information extraction in 2017–2019; and efficient methods, interpretability, summarization, clinical applications, and discourse only in 2023–2025. Only Syntax (14 publications) has declined, with no publications after 2023.

• Strong concentration in China: 122 publications (87.1%) are first-authored in China; the remaining 18 come from the USA (5), the UK (4), Germany (3), Israel (3), and Australia, France, and India (one each). China-affiliated work covers all 25 represented areas, 15 of them exclusively. Work from other countries concentrates on Old or Classical Tibetan and Buddhist texts, mainly in Resources and Evaluation (9), Syntax (5), NLP Applications (5), and Multilingualism (5).

• Two partly separate communities: Two-thirds of publications from outside China (12 of 18) appear in the ACL Anthology, seven of them at workshops, whereas 97 of the 122 China-affiliated publications (79.5%) appear in journals or conferences outside it.

• Only partial resource release: 41 of 140 publications (29.3%) release data, code, or models publicly. Release has increased from none of the 31 publications of 2004–2018 to 27 of the 61 of 2024–2026 (44.3%).

• A large measured LLM gap: GPT-3.5 reaches 3.6% F1 on Tibetan WikiAnn named-entity recognition, the lowest of ten low-resource Sino-Tibetan languages; on TLUE most LLMs perform below a random baseline, and Qwen2.5-72B drops from 84.7% on the Chinese CMMLU to 16.5% on its Tibetan counterpart; general multimodal LLMs recognize 49–66% of characters on TibOCR-Bench against 92% for a 7B model tuned on 4.5K Tibetan images.

• Evaluation is itself a bottleneck: Encoders are compared chiefly on one news-classification corpus (TNCC); task data, alignment data, and many benchmarks are increasingly translated through other languages. Machine-translated test items depress scores, most of all for Tibetan, and tasks built from native minority-language knowledge are 11.4–13.9 points harder than translated ones. The single gold segmentation in shared tasks covers only 22% of valid granularities.

• Dialects are underserved: Only eight publications name Lhasa, Amdo, or a multi-dialect setting in their titles, all on speech recognition, synthesis, or translation. Open speech data for all three dialect groups total about 80–100 hours, against over 10,000 for Mandarin or English. The main multi-dialect corpus, TIBMD@MUC, was released with 8.7 hours of Kham speech against 30 of Lhasa speech, and the best recognizer trained on it errs on 23% of Kham characters versus 3–5% for the other groups.

• Classical and modern traditions transfer asymmetrically: Old Tibetan has been normalized to Classical orthography to reuse tools, and one pipeline annotates texts from the 11th to the 21st century, but pre-training on Classical text did not help segment spoken Standard Tibetan, and general-domain Tibetan–Chinese MT drops from about 50 BLEU to at most 5 on Buddhist text.

• Script properties shape the research: Tibetan is an abugida with 30 consonant letters and 4 vowel signs; the tsheg delimits syllables but not words; Chinese national extension standards encode whole stacks, 7,198 in all, as single characters. Decomposing stacks into components reduces 562 handwritten character classes to 120 components.

Methodology in Plain English

The authors built their corpus in two stages. First, an automated retrieval step: they pulled every entry with "Tibetan" in the title or abstract from the full ACL Anthology BibTeX export, and queried the DBLP Search API with the same keyword, keeping only Web of Science–indexed journal articles and papers from venues recommended in the 7th edition of the China Computer Federation catalogue. Second, four researchers independently screened the merged, deduplicated candidates against inclusion criteria, considering only English-language papers and resolving disagreements by discussion.

Inclusion required peer-reviewed work in which Tibetan, modern or classical, is the primary language of study or a principal experimental object, with a contribution concerning computational processing of Tibetan text, speech, or script; multilingual studies qualified when Tibetan was among the principal experimental languages. Each publication was then annotated for venue type, the first author's affiliation country, and research area (up to two of the 29 ARR areas). Because 115 publications received two areas and 25 received one, the 255 area assignments sum to more than 140. The authors also recorded whether code, data, or models were released at a public URL.

Why This Matters

Impact on research. The survey consolidates a literature that is scattered across communities: 74% of the identified publications lie outside the ACL Anthology, the source used by the closest existing survey of Sino-Tibetan languages, and the closest Tibetan-specific overview was a brief editorial. By annotating venues, areas, countries, and release practices, it gives researchers a baseline for measuring progress and exposes gaps — including four ARR areas with no Tibetan publications at all and a near absence of ethics, bias, and safety work.

Real-world applications (drawn from the surveyed work and the recommendations):

• Document digitization and OCR: handwriting, woodblock, and manuscript recognition, scene text, interlinearized ancient texts, and OCR spelling correction, with a shared archival base in Buddhist Digital Resource Center scans.
• Speech technology: recognition for Lhasa and Amdo and in multi-dialect settings, speech synthesis, speech translation from Amdo or all three dialect groups, and word-level evaluation.
• Translation and language services: syllable- and subword-based neural machine translation, noisy parallel-data filtering, and comparison of dedicated MT engines against a prompted LLM.
• Content and safety moderation: offensive-language and rumour detection, information retrieval over Tibetan web and microblog text, and medical named-entity recognition (where F1 falls to 71% from 94% on news).

Industry relevance. The paper documents concrete obstacles to building Tibetan products: tokenization is unresolved, with a syllable vocabulary of 35K types still leaving 15% of tokens unknown, and multilingual tokenizers fragment Tibetan so that a 27-token English sentence takes 100 or more tokens; expanding byte-level BPE vocabularies helps only beyond about 3,000 new tokens per language. Data quality and availability are also constrained: 42% of sampled Tibetan CulturaX pages retain headers or footers, CINO's pre-training corpus is not public, and 41 GB of the 72 GB corpus used for dense and mixture-of-experts models is private. Safety alignment transfers unreliably across minority languages, and most LLMs fall below chance on TLUE's safety set — relevant for anyone planning deployment in education, healthcare, or public services.

Future Directions

• Close the dialect gap. Only eight publications name Lhasa, Amdo, or a multi-dialect setting in their titles. The authors call for Kham and Amdo resources to be prioritized, publicly released, and evaluated per variety, and note that only about 80–100 hours of open speech data exist across all three groups.
• Fix evaluation. The authors urge LLM benchmarks built increasingly from native material with expert involvement, such as CMiLBench's minority-knowledge tasks and DharmaBench, rather than by translation alone, plus shared conventions for encoding normalization, segmentation, and unit terminology, public test sets, and reporting of splits, metrics, and the varieties and scripts covered.
• Expand resource release and cross-community exchange. With only 29.3% of publications releasing data, code, or models, the authors recommend public release under documented licences, naming the varieties studied in titles and abstracts with venues recording them as metadata, and joint workshops, special issues, and shared tasks linking China-affiliated research with research elsewhere and the classical with the modern tradition.
• Study responsible NLP and involve communities. No surveyed publication assesses Tibetan speakers' needs; bias is unmeasured; safety is evaluated only within recent benchmarks. The authors call for consent and privacy safeguards, community participation across all three major dialect groups including diaspora communities, and evaluation of whether tools should take spoken dialect input rather than produce dialect output.
• Test transfer to related languages. Dzongkha and Ladakhi use Tibetan script, so OCR and segmentation tools may transfer directly, and Burmese could share low-level resources.

Target Audience

Researchers and graduate students in low-resource NLP, especially those working on Tibetan, Sino-Tibetan languages, or other underrepresented languages of the Global South; computational linguists and digital humanists working with Classical Tibetan and Buddhist texts; research funders and policy makers who need an evidence base on where Tibetan language technology stands; and NLP practitioners, product managers, or engineers evaluating whether Tibetan-language capabilities are ready for deployment, including those assessing tokenization, data availability, and benchmark reliability. Language community members and organizations interested in how their language is represented in computational research will also find the gap analysis directly relevant.

Authors’ abstract

Tibetan, spoken by some seven million people on and around the Qinghai-Xizang Plateau, remains underrepresented in natural language processing (NLP). We present, to our knowledge, the first peer-reviewed systematic bibliometric survey of Tibetan NLP: 140 publications from 2004 to 2026 in the ACL Anthology and DBLP, annotated for venue type, research area, first-affiliation country, and resource release. Output has grown rapidly, with 84% of the publications appearing since 2018, and has expanded bottom-up along the NLP pipeline, from corpora and word segmentation through script, speech, and applications to large language models (LLMs). Researchers based in China first-authored 87% of the publications, which span all 25 represented research areas, whereas work from other countries concentrates on Classical Tibetan. We argue that this full-pipeline paradigm, realized under the compounded constraints of a complex script, dialectal diversity, and geographic remoteness, offers a reference model for other low-resource languages of the Global South. We also identify gaps in dialect coverage, responsible NLP, native and comparable evaluation, cross-community visibility, and resource release (only 29% of the publications release data, code, or models), and recommend open science, shared standards, and community participation for the LLM era.