Multilingual NLP: Low-Resource and Morphologically Rich Languages

Most of the world's 7,000 languages have almost no data, and many build words by gluing morphemes together. This module measures the data gap and its compounding, builds morphological analyzers (finite-state transducers) and unsupervised segmenters (MDL), evaluates tokenizers fairly across languages (fertility, parity, boundary F1, bits per byte), reproduces the curse of multilinguality in a controlled toy, aligns embedding spaces across languages, and audits and cleans noisy multilingual data. Machine Translation for Low-Resource Languages: NLLB, chrF++ With Error Bars, and What the Metric Cannot See-End to End: A Language Pipeline From Raw Data to an Evaluated Model on 16 GB (partner) cover low-resource MT, adapting LLMs to a new language, speech, fair evaluation and an end-to-end capstone pipeline on 16 GB.

Objectives

Episodes

1. 7,000 Languages and the Data Gap18 min · 25 cues

   Typology (WALS features: word order, morphology, tone, scripts); morphological types (isolating, agglutinative, fusional, polysynthetic) with real examples (Turkish, Finnish, isiZulu, Swahili, Quechua, Inuktitut); resource taxonomies (Joshi et al. 2020); how data gaps compound through every pipeline stage; why this matters scientifically (generalization, linguistic diversity) and socially.
2. Computational Morphology: From Finite-State Transducers to Neural Inflection19 min · 29 cues

   Morphemes, stems, affixes, templatic morphology; finite-state transducers (two-level morphology, Koskenniemi) as composable string functions; morphological analyzers; UniMorph schema; SIGMORPHON shared tasks; neural inflection/reinflection as seq2seq; unsupervised segmentation (Morfessor).
3. Tokenization for Morphologically Rich Languages18 min · 26 cues

   Why BPE splits morphology badly; tokenizer fertility and parity (Petrov et al. 2023, Rust et al. 2021); unigram LM tokenization (SentencePiece, Kudo 2018); morphology-aware tokenizers; byte-level and tokenizer-free models (ByT5, CANINE, MEGABYTE-style patches, Byte Latent Transformer, 2024) and the latest 2025–2026 work; vocabulary allocation across languages in multilingual models.
4. Multilingual Pretraining and the Curse of Multilinguality17 min · 26 cues

   mBERT, XLM-R (Conneau et al. 2020), mT5, BLOOM, Aya-101/Aya Expanse, Qwen/Gemma multilingual coverage; language sampling with temperature (alpha); curse of multilinguality and capacity; shared vs language-specific parameters (adapters, MAD-X, X-MOD, language-specific experts); cross-lingual transfer and why it works at all.
5. Cross-Lingual Alignment: Shared Spaces Between Languages17 min · 27 cues

   Why multilingual models align representations without parallel data (anchors, shared subwords, structural similarity); cross-lingual word embedding alignment via Procrustes (link Procrustes Alignment: Comparing and Aligning Representation Spaces) and unsupervised methods (MUSE); sentence embeddings for bitext (LASER, LaBSE); language-neutral vs language-specific subspaces; the semantic hub / English-pivot findings in LLMs (2024–2026 research).
6. Low-Resource Data: Finding, Cleaning, and Creating It16 min · 26 cues

   Participatory research (Masakhane, Nekoto et al. 2020); web-crawl quality audits (Kreutzer et al. 2022); language identification at scale (GlotLID, OpenLID, link Language ID, Toxicity, and PII); bitext mining; dictionaries and grammars as data (e.g., MTOB: translating from a grammar book); synthetic data and back-translation; data sovereignty and community ownership (e.g., Te Hiku Media).
7. Machine Translation for Low-Resource Languages: NLLB, chrF++ With Error Bars, and What the Metric Cannot See18 min · 26 cues

   NLLB-200 (2022) design: data mining, distillation, MoE, evaluation on FLORES-200; MADLAD-400; LLM-based MT for low-resource languages and its failure modes; pivoting and multilingual transfer; chrF++ vs neural metrics on low-resource; human evaluation protocols.
8. Adapting an LLM to a New Language: Vocabulary Extension, Embedding Initialization, and Forgetting17 min · 24 cues

   Vocabulary extension and embedding initialization (WECHSEL, FOCUS, and 2025–2026 methods); continued pretraining and data mixing to limit forgetting; LoRA/QLoRA adapters; instruction data (Aya Dataset/Collection) and cultural adaptation; evaluation of fluency vs knowledge; current best practice.
9. Speech for Low-Resource and Unwritten Languages: Self-Supervised Features, Tone, and a Thousand-Language Recognizer17 min · 25 cues

   Massively Multilingual Speech (MMS, Pratap et al. 2023), Whisper's coverage limits, self-supervised speech representations for low-resource ASR, tone and tonal languages, languages without standard orthography, speech-to-speech translation (SeamlessM4T); latest multilingual speech models (2025–2026).
10. Evaluating Multilingual Models Fairly: Per What, Translated From What, and Safe in Which Language14 min · 21 cues

    Benchmarks (XTREME, Belebele, Global-MMLU, INCLUDE, FLORES) and their pitfalls (translationese, cultural bias, contamination); per-language reporting; tokenization-normalized metrics (bits per byte/character instead of perplexity per token); safety and toxicity across languages (multilingual jailbreak gaps, link Red Teaming: Measuring Robustness and the Cost of Defenses); cultural competence evaluation.
11. End to End: A Language Pipeline From Raw Data to an Evaluated Model on 16 GB15 min · 22 cues

    Capstone: choose a morphologically rich, lower-resource language; collect and audit data, language-ID and clean, train and compare tokenizers, adapt a small LLM (continued pretraining + LoRA SFT), evaluate (bits-per-byte, chrF++, task accuracy, human spot checks), error analysis by morphology, safety checks, and a data/model card; every stage visualized.

Exercises

Labs

• 7,000 Languages and the Data Gap browser L.01 · 40 min
• Computational Morphology: From Finite-State Transducers to Neural Inflection browser L.02 · 60 min
• Tokenization for Morphologically Rich Languages browser L.03 · 60 min
• Multilingual Pretraining and the Curse of Multilinguality browser L.04 · 50 min
• Cross-Lingual Alignment: Shared Spaces Between Languages browser L.05 · 55 min
• Low-Resource Data: Finding, Cleaning, and Creating It browser L.06 · 55 min
• Machine Translation for Low-Resource Languages: NLLB, chrF++ With Error Bars, and What the Metric Cannot See browser L.07 · 75 min
• Adapting an LLM to a New Language: Vocabulary Extension, Embedding Initialization, and Forgetting browser L.08 · 90 min
• Speech for Low-Resource and Unwritten Languages: Self-Supervised Features, Tone, and a Thousand-Language Recognizer browser L.09 · 75 min
• Evaluating Multilingual Models Fairly: Per What, Translated From What, and Safe in Which Language browser L.10 · 60 min
• End to End: A Language Pipeline From Raw Data to an Evaluated Model on 16 GB browser L.11 · 120 min

Library contributions

Files in nlpcore this module teaches you to write.

• nlpcore/multilingual/align.py
• nlpcore/multilingual/fst.py
• nlpcore/multilingual/morfessor.py
• nlpcore/multilingual/sampling.py
• nlpcore/multilingual/tokenizer_eval.py
• nlpcore/multilingual/typology.py
• nlpcore/multilingual/vocab_extend.py

Sources

1. Wang et al., AfriMTE and AfriCOMET: Enhancing COMET to Embrace Under-resourced African Languages, arXiv 2023/2024 · arXiv:2311.09828
2. Wikipedia, Agglutination: Turkish evlerinizden; Finnish talossa / taloissa; Swahili noun classes · link
3. Ahia et al., Do All Languages Cost the Same? Tokenization in the Era of Commercial Language Models, EMNLP 2023 · arXiv:2305.13707
4. Apples to Apples? Towards Comparable Crosslingual Language Model Evaluation, arXiv preprint 2026 (authors likely Yang, Arnett, Wilcox; NOT verified) · arXiv:2608.25089
5. Arnett & Bergen, Why do language models perform worse for morphologically complex languages?, COLING 2025 · arXiv:2411.14198
6. Arnett, Hudspeth & O'Connor, Evaluating Morphological Alignment of Tokenizers in 70 Languages, Tokenization Workshop @ ICML 2025 · link
7. Artetxe, Labaka & Agirre, Learning bilingual word embeddings with (almost) no bilingual data, ACL 2017
8. Singh et al., Aya Dataset: An Open-Access Collection for Multilingual Instruction Tuning, ACL 2024 · arXiv:2402.06619
9. Dang et al., Aya Expanse: Combining Research Breakthroughs for a New Multilingual Frontier, Dec 2024 · arXiv:2412.04261
10. Üstün et al., Aya Model: An Instruction Finetuned Open-Access Multilingual Language Model, ACL 2024 · arXiv:2402.07827
11. Beesley & Karttunen, Finite State Morphology, CSLI Publications, 2003
12. Bandarkar et al., The Belebele Benchmark: a Parallel Reading Comprehension Dataset in 122 Language Variants, ACL 2024 · arXiv:2308.16884
13. BigScience Workshop, BLOOM: A 176B-Parameter Open-Access Multilingual Language Model, 2022 · arXiv:2211.05100
14. Pagnoni et al., Byte Latent Transformer: Patches Scale Better Than Tokens, arXiv Dec 2024 · arXiv:2412.09871
15. Bostrom & Durrett, Byte Pair Encoding is Suboptimal for Language Model Pretraining, Findings of EMNLP 2020 · arXiv:2004.03720
16. Xue et al., ByT5: Towards a Token-Free Future with Pre-trained Byte-to-Byte Models, TACL 2022
17. Clark et al., CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation, TACL 2022
18. Chang, Arnett, Tu & Bergen, When Is Multilinguality a Curse? Language Modeling for 250 High- and Low-Resource Languages, EMNLP 2024 · arXiv:2311.09205
19. Chen, Adeleke, Downey, Linguistically Informed Evaluation of Multilingual ASR for African Languages, arXiv 2026 · arXiv:2602.04716
20. Cui, Yang, Yao, Efficient and Effective Text Encoding for Chinese LLaMA and Alpaca, arXiv 2023 · arXiv:2304.08177
21. Conneau, Lample, Ranzato, Denoyer & Jégou, Word Translation Without Parallel Data, ICLR 2018 · arXiv:1710.04087
22. Conneau et al., Unsupervised Cross-lingual Representation Learning at Scale (XLM-R), ACL 2020 · arXiv:1911.02116
23. Cotterell et al., The SIGMORPHON 2016 Shared Task: Morphological Reinflection, SIGMORPHON 2016 · link
24. Cotterell et al., CoNLL-SIGMORPHON 2017 Shared Task: Universal Morphological Reinflection in 52 Languages · arXiv:1706.09031
25. Cotterell et al., The CoNLL-SIGMORPHON 2018 Shared Task: Universal Morphological Reinflection · arXiv:1810.07125
26. Creutz & Lagus, Unsupervised Discovery of Morphemes, ACL 2002 Workshop on Morphological and Phonological Learning · arXiv:cs/0205057
27. Creutz & Lagus, Unsupervised Models for Morpheme Segmentation and Morphology Learning, ACM TSLP 4(1), 2007
28. Wikipedia, Cusco Quechua: wasi, wasikuna, wasiy, wasipi, wasimanta · link
29. Deng, Zhang, Pan, Bing, Multilingual Jailbreak Challenges in Large Language Models, arXiv 2023 (ICLR 2024 venue not verified) · arXiv:2310.06474
30. Eberhard, Simons & Fennig (eds.), Ethnologue, 27th edition (2024): 7,164 living languages · link
31. Penedo et al., FineWeb2: One Pipeline to Scale Them All, arXiv 2025 · arXiv:2506.20920
32. openlanguagedata/flores_plus dataset card: dev 997 / devtest 1,012 sentences; CC BY-SA 4.0; gated; not to be used as training data · link
33. Dobler, de Melo, FOCUS: Effective Embedding Initialization for Monolingual Specialization of Multilingual Models, EMNLP 2023 · arXiv:2305.14481
34. Gemma Team, Gemma 3 Technical Report, 2025 · arXiv:2503.19786
35. Singh et al., Global MMLU: Understanding and Addressing Cultural and Linguistic Biases in Multilingual Evaluation, arXiv 2024 · arXiv:2412.03304
36. Kargaran, Imani, Yvon & Schütze, GlotLID: Language Identification for Low-Resource Languages, Findings of EMNLP 2023 · arXiv:2310.16248
37. Hewitt, Initializing New Word Embeddings for Pretrained Language Models, 2021-12-06 · link
38. Hwang, Wang & Gu, Dynamic Chunking for End-to-End Hierarchical Sequence Modeling (H-Net), arXiv July 2025 · arXiv:2507.07955
39. Chen et al., Speech-to-Speech Translation For A Real-world Unwritten Language, arXiv 2022 · arXiv:2211.06474
40. Hudspeth, Burns & O'Connor, Contextual morphologically-guided tokenization for Latin encoder models, EACL 2026 · link
41. Ibrahim et al., Simple and Scalable Strategies to Continually Pre-train Large Language Models, TMLR 2024 · arXiv:2403.08763
42. Romanou et al., INCLUDE: Evaluating Multilingual Language Understanding with Regional Knowledge, arXiv 2024 · arXiv:2411.19799
43. Wikipedia, Inuit grammar: tusaatsiarunnanngittualuujunga 'I can't hear very well' with morpheme glosses · link
44. Joshi, Santy, Budhiraja, Bali & Choudhury, The State and Fate of Linguistic Diversity and Inclusion in the NLP World, ACL 2020 · arXiv:2004.09095
45. K, Wang, Mayhew & Roth, Cross-Lingual Ability of Multilingual BERT: An Empirical Study, ICLR 2020 · arXiv:1912.07840
46. Kann & Schütze, Single-Model Encoder-Decoder with Explicit Morphological Representation for Reinflection (MED), ACL 2016 · arXiv:1606.00589
47. Kaplan & Kay, Regular Models of Phonological Rule Systems, Computational Linguistics 20(3), 1994
48. Kirov et al., UniMorph 2.0: Universal Morphology, LREC 2018
49. Koehn, Statistical Significance Tests for Machine Translation Evaluation, EMNLP 2004 (paired bootstrap)
50. Koskenniemi, Two-level morphology: a general computational model for word-form recognition and production, University of Helsinki, Dept. of General Linguistics, Publications No. 11, 1983 · link
51. Kreutzer et al., Quality at a Glance: An Audit of Web-Crawled Multilingual Datasets, TACL 2022 · arXiv:2103.12028
52. Kudo & Richardson, SentencePiece, EMNLP 2018 (system demonstrations) · arXiv:1808.06226
53. Kudo, Subword Regularization: Improving Neural Network Translation Models with Multiple Subword Candidates, ACL 2018 · arXiv:1804.10959
54. Feng, Yang, Cer, Arivazhagan & Wang, Language-agnostic BERT Sentence Embedding, ACL 2022 (arXiv 2020) · arXiv:2007.01852
55. Lample & Conneau, Cross-lingual Language Model Pretraining (XLM), NeurIPS 2019 · arXiv:1901.07291
56. Artetxe & Schwenk, Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond, TACL 2019
57. Libovický, Rosa & Fraser, On the Language Neutrality of Pre-trained Multilingual Representations, Findings of EMNLP 2020 · arXiv:2004.05160
58. Kudugunta et al., MADLAD-400: A Multilingual And Document-Level Large Audited Dataset, NeurIPS 2023 Datasets & Benchmarks · arXiv:2309.04662
59. Kudugunta et al., MADLAD-400: A Multilingual And Document-Level Large Audited Dataset, NeurIPS 2023 Datasets and Benchmarks (arXiv 2023-09-09) · arXiv:2309.04662
60. Pfeiffer, Vulić, Gurevych & Ruder, MAD-X: An Adapter-Based Framework for Multi-Task Cross-Lingual Transfer, EMNLP 2020 · arXiv:2005.00052
61. Martins, Astudillo, From Softmax to Sparsemax, ICML 2016
62. google-research/bert, multilingual.md (mBERT) · link
63. McCarthy et al., UniMorph 3.0: Universal Morphology, LREC 2020
64. Yu et al., MEGABYTE: Predicting Million-byte Sequences with Multiscale Transformers, 2023
65. Mikolov, Le & Sutskever, Exploiting Similarities among Languages for Machine Translation, arXiv 2013 · arXiv:1309.4168
66. Mitchell et al., Model Cards for Model Reporting, FAT* 2019 · arXiv:1810.03993
67. Zhao, Pratap, Auli, Scaling A Simple Approach to Zero-Shot Speech Recognition, arXiv 2024 · arXiv:2407.17852
68. Pratap et al., Scaling Speech Technology to 1,000+ Languages, arXiv 2023 (JMLR 2024) · arXiv:2305.13516
69. Asgari, El Kheir & Sadraei Javaheri, MorphBPE: A Morpho-Aware Tokenizer Bridging Linguistic Complexity for Efficient LLM Training Across Morphologies, arXiv Feb 2025 · arXiv:2502.00894
70. Xue et al., mT5: A Massively Multilingual Pre-trained Text-to-Text Transformer, NAACL 2021 · arXiv:2010.11934
71. Tanzer, Suzgun, Visser, Jurafsky & Melas-Kyriazi, A Benchmark for Learning to Translate a New Language from One Grammar Book, ICLR 2024 · arXiv:2309.16575
72. Mundra et al., An Empirical Comparison of Vocabulary Expansion and Initialization Approaches for Language Models, CoNLL 2024 · arXiv:2407.05841
73. Limisiewicz, Blevins, Gonen, Ahia & Zettlemoyer, MYTE: Morphology-Driven Byte Encoding for Better and Fairer Multilingual Language Modeling, ACL 2024 · arXiv:2403.10691
74. Nekoto et al., Participatory Research for Low-resourced Machine Translation: A Case Study in African Languages, Findings of EMNLP 2020 · link
75. NLLB Team, No Language Left Behind: Scaling Human-Centered Machine Translation, arXiv 2022 (posted 2022-07-11) · arXiv:2207.04672
76. NLLB Team, Scaling neural machine translation to 200 languages, Nature 2024 · link
77. facebook/nllb-200-distilled-600M model card (Hugging Face): license CC-BY-NC-4.0; research model, not for production; trained on inputs up to 512 tokens · link
78. Olusanya, Tone in Yoruba ASR: Evaluating the Impact of Tone Recognition on Transformer-Based ASR Models, LoResLM 2026, pp. 149-156 · link
79. Omnilingual ASR team (Meta), Omnilingual ASR: Open-Source Multilingual Speech Recognition for 1600+ Languages, arXiv 2025; Meta AI blog 2025-11-10 · arXiv:2511.09690
80. Burchell, Birch, Bogoychev & Heafield, An Open Dataset and Model for Language Identification, ACL 2023 · arXiv:2305.13820
81. Foroutan, Meister, Paul, Niklaus, Ahmadi, Bosselut & Sennrich, Parity-Aware Byte-Pair Encoding: Improving Cross-lingual Fairness in Tokenization, arXiv Aug 2025 · arXiv:2508.04796
82. Petrov, La Malfa, Torr & Bibi, Language Model Tokenizers Introduce Unfairness Between Languages, NeurIPS 2023 · arXiv:2305.15425
83. Gao et al., The Pile: An 800GB Dataset of Diverse Text for Language Modeling, arXiv 2020 · arXiv:2101.00027
84. Popović, chrF: character n-gram F-score for automatic MT evaluation, WMT 2015
85. Popović, chrF++: words helping character n-grams, Proceedings of WMT 2017, pp. 612-618 · link
86. Qwen Team, Qwen3 Technical Report, May 2025 · arXiv:2505.09388
87. Robinson, Ogayo, Mortensen, Neubig, ChatGPT MT: Competitive for High- (but not Low-) Resource Languages, arXiv 2023 · arXiv:2309.07423
88. Rust, Pfeiffer, Vulić, Ruder & Gurevych, How Good is Your Tokenizer? On the Monolingual Performance of Multilingual Language Models, ACL 2021 · arXiv:2012.15613
89. Csaki et al., SambaLingo: Teaching Large Language Models New Languages, arXiv 2024 · arXiv:2404.05829
90. Schut, Gal & Farquhar, Do Multilingual LLMs Think In English?, arXiv Feb 2025 · arXiv:2502.15603
91. Seamless Communication et al., SeamlessM4T: Massively Multilingual & Multimodal Machine Translation, arXiv 2023 · arXiv:2308.11596
92. Jurafsky & Martin, SLP3 draft (Aug 19, 2026), Ch. 13 Machine Translation (13.1 typology, 13.5 low-resource, multilingual models, sociotechnical issues) · link
93. Jurafsky & Martin, SLP3 draft (Aug 19, 2026), Ch. 2 Words and Tokens (2.1 Heaps' law, 2.2 morpheme typology, 2.3 Unicode, 2.4 BPE) · link
94. Smith, Turban, Hamblin & Hammerla, Offline bilingual word vectors, orthogonal transformations and the inverted softmax, ICLR 2017
95. Caswell et al., SMOL: Professionally translated parallel data for under-represented languages, arXiv 2025; google/smol, CC-BY-4.0 · arXiv:2502.12301
96. HuggingFaceTB/SmolLM2-360M model card: Apache-2.0; 4T training tokens; primarily English · link
97. Søgaard, Ruder & Vulić, On the Limitations of Unsupervised Bilingual Dictionary Induction, ACL 2018 · arXiv:1805.03620
98. Song et al., Is LLM the Silver Bullet to Low-Resource Languages Machine Translation?, arXiv 2025 · arXiv:2503.24102
99. Wiktionary, ninakupenda: ni- (I) -na- (present) -ku- (you) penda (love) · link
100. ITU, AI for indigenous languages: te reo Māori (Aug 9, 2022) · link
101. NVIDIA blog: Māori Speech AI Model Helps Preserve and Promote New Zealand Indigenous Language (Te Hiku Media) · link
102. Universal Declaration of Human Rights (UN, 1948): English (un.org) and Kiswahili (transcription at lexilogos.com), retrieved 2026-10-29 · link
103. Universal Declaration of Human Rights, Article 1, via the Unicode UDHR project / udhr@6.0.0 npm package · link
104. Chung et al., UniMax: Fairer and More Effective Language Sampling for Large-Scale Multilingual Pretraining, ICLR 2023 · arXiv:2304.09151
105. Batsuren et al., UniMorph 4.0: Universal Morphology, LREC 2022 · arXiv:2205.03608
106. Dryer & Haspelmath (eds.), The World Atlas of Language Structures Online, 2013 edition, MPI-EVA; CC BY 4.0; first online April 2008 · link
107. Baevski et al., wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations, NeurIPS 2020 · arXiv:2006.11477
108. Minixhofer, Paischer, Rekabsaz, WECHSEL: Effective initialization of subword embeddings for cross-lingual transfer of monolingual language models, NAACL 2022 · arXiv:2112.06598
109. Wendler, Veselovsky, Monea & West, Do Llamas Work in English? On the Latent Language of Multilingual Transformers, ACL 2024 · arXiv:2402.10588
110. Radford et al., Robust Speech Recognition via Large-Scale Weak Supervision, ICML 2023 (arXiv Dec 2022) · arXiv:2212.04356
111. Deutsch et al., WMT24++: Expanding the Language Coverage of WMT24 to 55 Languages & Dialects, Findings of ACL 2025 · arXiv:2502.12404
112. Wu, Yu, Yogatama, Lu & Kim, The Semantic Hub Hypothesis: Language Models Share Semantic Representations Across Languages and Modalities, ICLR 2025 · arXiv:2411.04986
113. Xing, Wang, Liu & Lin, Normalized Word Embedding and Orthogonal Transform for Bilingual Word Translation, NAACL 2015
114. Babu et al., XLS-R: Self-supervised Cross-lingual Speech Representation Learning at Scale, arXiv 2021 · arXiv:2111.09296
115. facebook/wav2vec2-xls-r-300m (Apache-2.0); facebook/mms-300m (CC-BY-NC-4.0); google/fleurs (CC-BY-4.0, 16 kHz)
116. Pfeiffer et al., Lifting the Curse of Multilinguality by Pre-training Modular Transformers (X-MOD), NAACL 2022 · link
117. Hu et al., XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization, ICML 2020 · arXiv:2003.11080
118. Yong, Menghini, Bach, Low-Resource Languages Jailbreak GPT-4, arXiv 2023 · arXiv:2310.02446
119. 16 bytes per parameter for mixed-precision Adam (2 + 2 + 4 + 8) is the common accounting popularized by the ZeRO paper (Rajbhandari et al. 2020)
120. Minixhofer, Ponti, Vulić, Zero-Shot Tokenizer Transfer, arXiv 2024 · arXiv:2405.07883
121. Zhu et al., Multilingual Machine Translation with Large Language Models: Empirical Results and Analysis, arXiv 2023 · arXiv:2304.04675
122. Wikipedia, Zulu grammar: noun classes 1/2 (umuntu/abantu), 7/8 (isi-/izi-), Ngiyazigeza · link