cs.CLOct 9, 2026

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

Authors: Xi Cao, Yijie Li, Yuan Sun, Shiwei Dong, Zeguang Huang, Nuo Qun

Organizations: School of Philosophy and Religious Studies, Minzu University of China · Institute of National Security, Minzu University of China · National Language Resource Monitoring & Research Center | Minority Languages Branch · School of Information Engineering, Minzu University of China · 

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

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.

Figures & tables

Figure 1: The Tibetan word བསྒྲིགས་བཟོ་ ( programming ), illustrating the three-level structure of Tibetan script: letters (30 consonants and 4 vowel signs, each encoded by a Unicode code point, shown at the bottom) combine into syllables according to strict orthographic rules, and syllables, each followed by the inter-syllable mark tsheg (་), form words whose boundaries are not marked.
FeatureOur SurveyNAACL Survey ( Liu and Best, 2025 )TALLIP Editorial ( Long and Hill, 2021 )
ScopeTibetan onlySino-Tibetan family (49 low-resource languages)Tibetan only
Coverage Period2004–2026 : from the earliest identified paper through 00:00 UTC, September 25, 20261952–2022: the ACL OCL corpus ( Rohatgi et al., 2023 ) ; venue trends for 2000–20212016–2021: editorial for a special issue of ACM TALLIP
Corpus140 publications , annotated for venue type, research area, country, and resource release253 papers (274 language–paper matches for 23 of the 49 languages), annotated for topicNo fixed count: narrative overview of key developments
SourcesHybrid : ACL Anthology BibTeX and DBLP Search APIACL Anthology only, via the ACL OCL corpusSelected by the special-issue editors
Table 1: Comparison of our survey with the two most closely related overviews of Tibetan NLP.
Figure 2: Two-step construction of the survey corpus. Step 1 (automated retrieval): entries with Tibetan in the title or abstract are retrieved from the full ACL Anthology BibTeX file, and publications matching the keyword Tibetan are retrieved through the DBLP Search API, retaining journal articles verified by Web of Science and conference papers from venues recommended in the 7th edition of the CCF catalogue. Step 2 (manual screening): the merged and deduplicated candidate pool is reviewed for relevance, yielding the final set of 140 publications.
Figure 3: Annual number of Tibetan NLP publications by venue type (left; stacked bars) and overall share of each venue type (right; counts in the legend), 2004–2026. The year 2026 is only partially covered. The full-scale version is Figure 7 in Appendix E .
Figure 4: Number of publications per conference (top left), workshop (top right), and journal (bottom). Papers at joint conferences (ACL-IJCNLP 2015, IJCNLP-AACL 2025) are counted under each constituent venue, so the conference counts sum to 57 for 54 papers. The full-scale version is Figure 8 in Appendix E .
Figure 5: Number of publications per ARR research area (left; descending order) and per country of the first author’s affiliation (right; ascending order). As a publication can belong to more than one area, area counts sum to 255. The full-scale version is Figure 9 in Appendix E .
Figure 6: Number of publications per ARR research area and year, 2004–2026; darker cells indicate more publications. Areas are ordered by total count (descending). The full-scale version is Figure 10 in Appendix E .
PhasePublicationsAreasChina
2004–2017221395.5%
Top areas: Resources (7), Phon./Morph./Seg. (5), Multilingual (5), Syntax (4), Multimodality (4)
2018–2021401882.5%
Top areas: Resources (9), Multilingual (9), Multimodality (7), Speech (7)
2022–2026782387.2%
Top areas: Multilingual (27), Resources (24), Language Modeling (13), Multimodality (13)
Table 2: Three phases of Tibetan NLP. Areas : ARR areas represented; China : share first-authored in China; Top areas : four most frequent areas, with ties. Boundaries: the 2018 output rise and the first BERT-style model, TiBERT ( Liu et al., 2022 ) , in 2022.
Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: Full-scale version of Figure 3 : distribution of the 140 Tibetan NLP publications by year and venue type, 2004–2026. Left: stacked bars give the annual number of conference papers (blue), workshop papers (green), and journal articles (tan); numbers inside the segments are per-type counts, and numbers above the bars are annual totals. No publications were identified for 2007–2009. Output rises from three papers in 2017 to nine in 2018 and peaks at 22 in 2024; 2026 is covered only up to 00:00 UTC, September 25, 2026. Right: overall share of each venue type: journal articles 52.9% (74), conference papers 38.6% (54), and workshop papers 8.6% (12).
Figure 8: Full-scale version of Figure 4 : number of Tibetan NLP publications per venue; within each panel, venues are sorted by count in descending order, with ties listed alphabetically. Top left: 22 conferences, led by ACL (8, including Findings of ACL), NLPCC (6), PRCV (6), and COLING (5). Papers at joint conferences (ACL-IJCNLP 2015 and IJCNLP-AACL 2025) are counted under each constituent venue, so the bars sum to 57 attributions for 54 conference papers. Top right: 10 workshops with 12 papers, led by ALP and EURALI (2 each). Bottom: 25 journals with 74 articles, led by ACM Transactions on Asian and Low-Resource Language Information Processing (19), IEEE Access (14), and Data Intelligence (8); journal names follow DBLP abbreviations. Full venue names are listed in Appendices A and B.
Figure 9: Full-scale version of Figure 5 : number of Tibetan NLP publications by ACL Rolling Review (ARR) research area and by country of the first author’s primary affiliation. Left: the 29 ARR research areas, sorted in descending order. Each publication was assigned up to three areas (in practice, 115 received two and 25 received one), so the counts sum to 255 rather than 140. Multilingualism and Cross-Lingual NLP (41) and Resources and Evaluation (40) lead, whereas Dialogue and Interactive Systems; Ethics, Bias, and Fairness; LLM Agents; and NLP and Code Models have no publications. Right: countries, sorted in ascending order: China accounts for 122 (87.1%), followed by the USA (5), the UK (4), Germany (3), Israel (3), and Australia, France, and India (one each). Per-country and per-area publication lists are given in Appendices C and D, respectively.
Figure 10: Full-scale version of Figure 6 : number of Tibetan NLP publications in each of the 29 ARR research areas (rows) and each year (columns), 2004–2026. Rows follow the order of Figure 9 (descending total count). Cell shading darkens with the number of publications (scale 0–10); each non-empty cell shows its count, and blank cells indicate no publications. Because a publication can belong to more than one area, the cells sum to 255. The largest values are Resources and Evaluation in 2026 (10) and Multilingualism and Cross-Lingual NLP in 2024 (9); the columns for 2007–2009 are empty, and 2026 is covered only up to 00:00 UTC, September 25, 2026.
More in this topic

Low-Resource Language Processing

Low-resource language processing builds and evaluates language technologies for languages with limited data, tools, or digital representation.

See all 565 papers

Explore similar work

CardsList

  1. A Systematic Review of NLP for Ghanaian Languages: Datasets, Models, and a Research Roadmap

    May 10, 2024Sheriff Issaka, Erick Rosas Gonzalez, Colene Agbo +18Low-Resource Language ProcessingAfrican Language Processing

    African Languages Lab

  2. Measuring Research Progress and Linguistic Inclusion in Low-Resource Languages

    Oct 21, 2026Atnafu Lambebo Tonja, Tadesse Destaw Belay, Jesujoba Oluwadara Alabi +6Low-Resource Language Processing

    University College London, UK · Instituto Politécnico Nacional, Mexico · Saarland University, Germany +2

  3. A Comprehensive Analysis of Arabic Natural Language Processing Research: Trends, Topic Evolution, and Research Gaps -- A Bibliometric and Topic-Based Study

    Aug 24, 2026Mullosharaf K. ArabovArabic NLPLow-Resource Language Processing

    Kazan Federal University, Institute of Computational Mathematics and Information Technologies, Kazan, Russia