Organizations: University College London, UK · Instituto Politécnico Nacional, Mexico · Saarland University, Germany · University of the Witwatersrand, South Africa · MBZUAI, UAE
Large Language Models (LLMs) have accelerated research on low-resource and multilingual NLP, but their effectiveness in creating an equitable landscape remains unmapped. We analyze 1,158 peer-reviewed open-access papers published between 2020 and 2025, examining research attention to low-resource languages (LRLs) and representation across language and speaker populations. To assess linguistic inclusion beyond publication counts, we introduce the Linguistic Inclusion Framework (LIF), a multidimensional framework that measures linguistic inclusion in four dimensions: research growth, representation expansion, representation continuity, and distributional equity. While annual publications grew from 52 to 415, a 5.4X increase in their ACL Anthology publication share, this growth largely concentrated on represented languages rather than expanding coverage of underrepresented languages. Further analysis shows that languages spoken in Asia account for 64.1% of speakers across LRLs we analyze but receive only 46.7% of the language representation share (0.7X their proportional share), while European languages receive 5.3X their proportional share. At the country level, linguistically diverse regions such as sub-Saharan Africa and Southeast Asia show the lowest language coverage ratios. Applying the framework at the continental level, we find that no region scores well on all four dimensions, showing that research growth alone is not a reliable metric of progress toward equitable and inclusive language technology\footnote{We will release all collected and annotated data for research purposes}.
Figures & tables
Figure 1: Overall paper retrieval and annotation pipeline.
Figure 2: Year-by-year growth in LRL research publications from 2020 to 2025, aggregated by publication venue.
Figure 3: Overview of publication trends and research topics.
Figure 4: Country-level distribution of language representation in our dataset. (a) Language coverage ratio across all languages, regardless of the number of papers in which they are represented; (b) language coverage ratio considering only languages represented in more than one paper; and (c) language representation per one million speakers.
Figure 5: Year-to-year language transition by categories and continent.
Figure 6: Year-by-year growth in newly released HuggingFace models (left) and datasets (right).
Dimension
What it shows
Measurement
A higher score indicates
Research growth
Whether research activity involving LRLs increases over time
mean annual instancesleast-squares slope of annual instances
Stronger growth relative to the average research activity
Representation expansion
Whether research extends beyond the languages represented at the start of the study
all languages, 2020–2025languages first seen after 2020
More new languages entered the literature after the baseline year
Representation continuity
Whether represented languages remain visible rather than appearing once
repeated+droppedrepeated
More sustained representation, a lower score means languages disappear after first appearing
Distributional equity
Whether attention is evenly shared among LRLs or concentrated in a subset
1−Gini of instances per language, within continent
A more even distribution
Table 1: The four dimensions of the LIF. Instances are language-publication instances, and repeated and dropped are counted over all year-to-year transitions.
Continent
Langs
Inst.
Growth
Expan.
Cont.
1 − G
Top-10
Africa
97
1,856
0.348
0.814
0.810
0.335
50.5%
Asia
138
3,520
0.328
0.674
0.816
0.311
43.0%
Europe
91
1,713
0.264
0.571
0.821
0.391
45.5%
North America
27
248
0.319
0.852
0.785
0.464
75.8%
South America
22
186
0.306
0.909
0.692
0.404
84.4%
Oceania
10
65
0.382
0.900
0.739
0.605
100.0%
Table 2: LIF results per continent, for dimensions defined in Table 1 . We added top-10 shares because equity is inflated where the represented languages are few.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Term group
Exact search terms
Language-model terms
"large language model" ; "LLM" ; "pre-trained language model" ; "foundation model" ; and "generative AI" .
"multilingual language model" "low-resource" ; "multilingual model" "under-resourced" ; and "multilingual NLP" "low-resource language" .
Appendix
Table 4: Supplementary literature-search queries.
Figure 7: Geographic distribution of low-resource NLP papers by the countries of all author affiliations. Each paper is counted once per distinct country appearing among its affiliations. Darker shades indicate more papers; grey denotes countries with no mapped affiliations. The United States (135), Germany (82), China (77), India (72), and the United Kingdom (61) contribute the most papers.
Category
# Papers
%
Low-resource languages
1,158
82.54
Low-resource domain
104
7.41
Simulation only
93
6.63
Not low-resource
48
3.42
Appendix
Table 5: Distribution of papers across the scope-classification categories.
Agreement
Language(s)
Low-resource categorization
Contribution
NLP topic
H1 & H2
93.3
96.01
98.66
92.4
H1 & H3
91.97
90.26
98.66
93.33
H2 & H3
96.49
91.17
98.67
93.01
Human & LLM
97.32
96.68
99.10
99.11
Appendix
Table 6: Pairwise percentage agreement between H1–H3 and the majority-vote human and LLM-generated annotations across the five annotation dimensions.
Year
Matched
Anthology
Share
2020
31
7,163
0.43%
2021
45
6,996
0.64%
2022
75
8,456
0.89%
2023
119
8,936
1.33%
2024
204
11,937
1.71%
2025
342
14,686
2.33%
Appendix
Table 7: Papers from our corpus indexed in the ACL Anthology, against total Anthology output per year.
Figure 8: Composition of data modalities by publication year.
Figure 9: Continental shares of L1&L2 speakers and language representation.
Figure 10: Language representation with respect to the speaker population.
Criteria
Weight
Language code in repository tags
+5
Language code or name in card language field
+5
Language name in card README text ∗
+5
Language name in tags or card metadata
+3
Language name in repository identifier
+2
Language name in other descriptive fields
+1
Appendix
Table 8: Relevance criteria used to validate Hugging Face candidates. ∗ Evaluated only when the card’s language field did not match.
L3S Research Center, Leibniz University Hannover, Appelstraße 9A, 30167 Hannover, Germany · Technische Universität Darmstadt, Residenzschloss 1, 64283 Darmstadt, Germany