Multilingual & Cross-Lingual Transfer
A Resource-Floor Benchmark: Measuring the Minimum Data, Lexical, and Pretrained-Model Prerequisites Below Which Multilingual NLP Methods Fail
Generated automatically from the limitations stated in 119 papers (ACL, EMNLP, NAACL), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
Across transfer, adaptation, evaluation, and data-construction work, methods repeatedly assume some minimal target-language resource — unlabeled text, a bilingual dictionary, a working MT system, fastText embeddings, a pretrained multilingual encoder, or native-speaker references — and each assumption silently excludes the world's least-served languages. Because every paper states this as a single-sentence caveat, nobody knows which prerequisite is binding, how much of it is actually needed, or where the failure floor lies. The consequence is a systematic blind spot: methods are reported as "multilingual" while being untested, and likely unusable, for languages with only tens of thousands of sentences, no dictionary, and no MT coverage. A quantified floor would turn these caveats into design constraints and tell funders and corpus-builders exactly which resources to create first.
Why it matters
Method designers could stop claiming general multilingual applicability and instead state verified operating ranges; corpus-building efforts could be prioritized by which single missing resource unblocks the most methods per language.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Prerequisite ablation grid. Pick 3–5 representative tasks (cross-lingual transfer, instruction tuning, toxicity/generation eval, embedding retrieval) and 6–8 languages spanning a controlled resource gradient (from ~10k unlabeled sentences and zero parallel data upward). For each task, systematically remove or degrade each prerequisite (dictionary, MT, embeddings, encoder coverage, unlabeled text) and measure the performance cliff. Output: a per-task "resource floor" curve, e.g., accuracy vs. hours of unlabeled text at fixed model scale.
- 2
Sparse-resource stress test on a real floor language. Select 2–3 languages with genuinely minimal coverage (e.g., an Indigenous American or Philippine language with <50k sentences available) and run the best-performing methods from each task family at that floor, measuring how far below the synthetic-floor predictions real performance lands. This validates whether the ablation grid predicts reality.
- 3
Floor-conditional method routing. Using the measured curves, produce and release a decision table — for a language with X tokens, Y dictionary entries, Z MT quality, which methods are viable — validated by leave-one-language-out prediction of method success on held-out languages.
Have a different approach?
Describe how you would tackle this problem and we'll look for papers that already do it. Free; your text stays private.
Why it might fail
The floors may turn out to be method-specific to the point of uselessness — no generalizable table emerges, and the paper reduces to "everything needs data," reproducing the existing caveats with numbers attached.
Sub-problems
A Systematic Multi-Tier Robustness Benchmark of Cross-Lingual Transfer Across Truly Low-Resource and Non-Latin Regimes
Current multilingual transfer methods are overwhelmingly evaluated on high- and mid-resource Indo-European languages or synthetic machine-translated benchmarks. As a result, reported transfer gains consistently fail to generalize to truly low-resource languages, typologically distant scripts, and settings lacking abundant English parallel corpora. Without a unified, multi-tier evaluation across standardized architectures and genuine native test sets, researchers cannot determine which cross-lingual adaptation methods remain effective when foundational multilingual embeddings and parallel supervision degrade.
Cross-Lingual Transfer Under Complete Target-Resource Absence
Existing cross-lingual transfer pipelines require target-language preconditions: monolingual corpora for masked language model adaptation, bilingual lexicons for embedding alignment, or machine translation systems for data projection. For the vast majority of the world's ~7,000 languages, none of these digital artifacts exist. As a result, current methods cannot operate in true zero-resource scenarios, structurally excluding thousands of under-documented, endangered, or primarily spoken languages from modern language technologies.
Cross-Benchmark Robustness and Evaluation Reliability of Multilingual Models on Micro-Scale Low-Resource Test Sets
Evaluation in low-resource and dialectal NLP routinely relies on test sets of only 100–300 annotated examples per language, often spanning severe class imbalances or single-speaker recordings. Because no unified empirical study has systematically evaluated standard multilingual models and metrics across these diverse micro-scale settings, it is unknown whether reported performance gains reflect genuine linguistic generalization or high sampling variance. Consequently, researchers cannot reliably determine whether multilingual models transfer effectively across low-resource typologies or if standard evaluation protocols fail under extreme data scarcity.
Benchmarking and Developing NLP Pipelines Under Zero-Upstream-Resource Conditions
Current NLP and alignment pipelines presuppose the availability of upstream linguistic tools, such as reliable word tokenizers, dependency parsers, morphological reference databases, or pretrained language models. When applied to extremely low-resource languages and dialects where these tools do not exist, systems are forced to fall back on surface-level n-gram overlap or fail entirely. This structural dependency prevents modern NLP methods from being deployed or evaluated across thousands of under-resourced languages.
Disentangling Dataset Size and Granularity Imbalances in Multilingual Benchmark Evaluations
Existing multilingual benchmarks exhibit severe disparities in sample size, relation coverage, and proficiency levels across languages, frequently varying by more than an order of magnitude. Because raw evaluation scores are highly sensitive to sample counts and slice granularities, current cross-lingual comparisons conflate dataset volume with intrinsic linguistic difficulty or model capability. This prevents researchers from determining whether performance drops in low-resource languages reflect true modeling failures or are artifacts of unbalanced test sets.
Evaluation and Probing of Low-Resource Language Models Without Expert-Curated Linguistic Resources
Current linguistic probing, template-based evaluation, and diagnostic benchmarks for multilingual language models strictly depend on expert-designed grammar templates, native reference corpora, and mature upstream tooling like dependency parsers. In truly low-resource and endangered language varieties where neither computational parsers nor accessible native linguists exist, these evaluation pipelines cannot be instantiated at all. As a result, model performance and syntactic competence on thousands of world languages remain entirely unmeasured and unverifiable.
Evidence
Each paper's own statement of the limitation, verbatim.
- When Does Mixing Help? Analyzing Query Embedding Interpolation in Multilingual Dense RetrievalACL 2026
Restricted to the 14 languages with parallel translations in mMARCO, limiting coverage of low-resource pairs; language-pair count grows quickly and parallel data for low-resource languages requires substantial resources
- Toward Robust Multilingual Adaptation of LLMs for Low-Resource LanguagesICML 2026
Requires generating and scoring multiple translation candidates using auxiliary LLMs/MT models during offline preprocessing or training, which introduces compute overhead.
- The GaoYao Benchmark: A Comprehensive Framework for Evaluating Multilingual and Multicultural Abilities of Large Language ModelsACL 2026
Non-uniform language and task coverage across integrated sub-benchmarks (e.g., SAGE and CultureScope cover only 2 languages/cultures).
- SiLP: Enhancing Non-Dominant Language Capabilities with a Selective Bidirectional Language Projection FrameworkACL 2026
Only a single non-dominant language is handled per model; cross-lingual interactions when aligning multiple languages simultaneously are untested
- SEA-BED: How Do Embedding Models Represent Southeast Asian Languages?ACL 2026
Coverage is uneven across the 10 languages; Tetum has only 4 datasets, undermining reliability of conclusions for extremely low-resource languages
- Post-Training Language Models for Crosslingual ConsistencyICML 2026
Excluding low-resource languages in the 12-language setup (Bengali for input length; Swahili/Yoruba at random chance) and English accuracy can drop substantially with some β settings (−11.74 with SW-anchored strengths)
- NiuTrans.LMT: Toward Inclusive and Scalable Multilingual Machine Translation with LLMsACL 2026
Coverage restricted to 60 languages with a bi-centric (Zh/En) design; scaling to more languages is constrained by parallel data availability and training cost.
- Multilingual Routing in Mixture-of-ExpertsICLR 2026
Ineffective for models with extremely poor multilingual capabilities (like OLMoE) where the prerequisite cross-lingual routing alignment is absent.
Show all 119 papers
- MultiFinBen: Benchmarking Large Language Models for Multilingual and Multimodal Financial ApplicationACL 2026
Newly introduced datasets are small (e.g., PolyFiQA has 204 questions per tier; some Spanish/Greek datasets have 50-225 examples), reducing statistical robustness
- ML-Embed: Inclusive and Efficient Embeddings for a Multilingual WorldICML 2026
Underperforms high-resource-specialized baselines on English (73.26 vs 75.97 top-1), Chinese (67.22 vs 78.52 top-1), and MTEB-Multilingual benchmarks due to balancing broad multilingual coverage.
- Less Languages, Less Tokens: An Efficient Unified Logic Cross-lingual Chain-of-Thought Reasoning FrameworkACL 2026
Requires a validation set with parallel translations/prompt renditions across all languages to estimate language centers, plus hyperparameter sensitivity to pruning ratio (accuracy drops from 81.8 to 63.2 as rho goes to 0.85)
- Language-Aware Token Boosting: LLM Language Confusion Reduction Without TuningACL 2026
Requires per-language hyperparameter tuning (α, β); vanilla LATB needs careful search to allow necessary non-target-language output
- LCMA-SRT: Language-Conditional Mixture-of-Experts Adapters for Joint Multilingual Speech Recognition and TranslationACL 2026
Only tested on Europarl-ST with limited domain and language-family coverage; results may not generalize to other domains
- Knowledge Beyond Language: Bridging the Gap in Multilingual Machine Unlearning EvaluationACL 2026
Results rely on synthetic profiles and LLM-translated data verified by back-translation, which may not reflect naturally memorized multilingual knowledge.
- ChiKhaPo: A Large-Scale Multilingual Benchmark for Evaluating Lexical Comprehension and Generation in Large Language ModelsACL 2026
Heavy reliance on public lexicons, especially PanLex, which is noisy for many languages; lexicons lack word-sense annotations, and undocumented synonyms/variants cause false negatives in WT
- CTC-DRO: Robust Optimization for Reducing Language Disparities in Speech RecognitionICLR 2026
Only evaluated with small training data (1–9 hours per language-corpus pair) on six-language sets (plus one 18-language experiment); results on much larger-scale or autoregressive (e.g., Whisper-style) training are untested
- Beyond English-Centric Training: How Reinforcement Learning Improves Cross-Lingual Reasoning in LLMsICLR 2026
Only 8K samples per language (GSM8K) and 45K (LUFFY) math-centric training data are used, so the non-English superiority finding may not hold for other domains or data scales
Training languages tested are limited (En, Zh, De, Fr, Ja) and are all high-resource languages with strong Qwen pre-training coverage; truly low-resource training languages are untested
- When Life Gives You Samples: The Benefits of Scaling up Inference Compute for Multilingual LLMsEMNLP 2025
Only high-resource, well-represented languages are tested; generalization to underrepresented languages unsupported by the model is unknown and expected to degrade sample quality and judge precision
- When Less Language is More: Language-Reasoning Disentanglement Makes LLMs Better Multilingual ReasonersNeurIPS 2025
Interventions in the uppermost layers cause severe degradation in linguistic fidelity, forcing models to default to English regardless of the input language.
- UniversalCEFR: Enabling Open Multilingual Research on Language Proficiency AssessmentEMNLP 2025
Severe data imbalance across languages and granularities: Spanish has zero C2 instances, Welsh has no B2/C-level data, Hindi/Russian/Welsh are sentence-only, and most texts are English (324k of 505k), so cross-language comparisons conflate data quantity with language difficulty (e.g., Portuguese tops out near 33.5 weighted F1).
- Qwen2.5-xCoder: Multi-Agent Collaboration for Multilingual Code Instruction TuningACL 2025
Only applied to code instruction tuning; the paradigm is untested on other multilingual tasks
- Preserving Multilingual Quality While Tuning Query Encoder on English OnlyNAACL 2025
Results shown only for one main multilingual model (multilingual-e5-small) with limited replication on MiniLM-L12-v2 and a monolingual e5-small-v2; other model families and sizes untested
- Parrot: Multilingual Visual Instruction TuningICML 2025
MoE experts are fixed at six, one per benchmark language, so extension to new languages requires new experts and multilingual training data
- PROM: Pivoted and Regulated Optimization for Multilingual Instruction LearningNAACL 2025
Absolute scores are very low (e.g., MGSM accuracy 0.4–5.6, QA EM mostly under 6), so gains are relative to a weak baseline rather than strong multilingual performance
- Multilinguality Does not Make Sense: Investigating Factors Behind Zero-Shot Cross-Lingual Transfer in Sense-Aware TasksEMNLP 2025
Training languages are limited and highly imbalanced (German, French, English overrepresented); low-resource languages lack sufficient fine-tuning data, and MONO Hindi/Chinese weakness may stem from small dataset size rather than the claimed effects.
- Multilingual vs Crosslingual Retrieval of Fact-Checked Claims: A Tale of Two ApproachesEMNLP 2025
Skewed language representation (English paired with nearly all languages; few crosslingual pairs for e.g. Dutch/Romanian) means findings may change with a different language distribution
- Multilingual Encoder Knows more than You Realize: Shared Weights Pretraining for Extremely Low-Resource LanguagesACL 2025
Single-language fine-tuning experiments are restricted to Tibetan due to lack of datasets for the other three languages, limiting scope of conclusions
- Multilingual Arbitration: Optimizing Data Pools to Accelerate Multilingual ProgressACL 2025
Only mid- and high-resource languages are evaluated; extremely low-resource languages lacking suitable teacher experts remain infeasible due to routing-expert scarcity and poor synthetic data quality
- MultiSocial: Multilingual Benchmark of Machine-Generated Text Detection of Social-Media TextsACL 2025
Many language-platform combinations lack sufficient samples (e.g., several languages have no WhatsApp/Discord data), limiting fully balanced evaluation; dataset is biased toward Indo-European languages (18 of 22)
- Multi-LMentry: Can Multilingual LLMs Solve Elementary Tasks Across Languages?EMNLP 2025
Nine languages only; not all 25 tasks port to every language (e.g., rhyming, homophones), limiting cross-language task comparability
- MuRating: A High Quality Data Selecting Approach to Multilingual Large Language Model PretrainingNeurIPS 2025
Covers only 17 target languages; no language-specific or culturally adapted scoring, and translation quality constrains applicability to languages with good machine translation
- McEval: Massively Multilingual Code EvaluationICLR 2025
Only ~50 generation/explanation samples per language, so per-language estimates are coarse
- MPO: Multilingual Safety Alignment via Reward Gap OptimizationACL 2025
Requires a model that is already well safety-aligned in a dominant language; Qwen2.5's weaker foundational low-resource language ability caps MPO's absolute gains (e.g., Swahili ASR remains high on AdvBench-X)
- MEXMA: Token-level objectives improve sentence representationsACL 2025
Covers only 81 languages versus SONAR's 200, limiting multilingual usability; only BERT-like MLM-pretrained backbones were explored.
- Learn and Unlearn: Addressing Misinformation in Multilingual LLMsEMNLP 2025
Only one fake-data language per training run is tested; simultaneous multilingual contamination is not evaluated.
- Language Fusion for Parameter-Efficient Cross-lingual TransferACL 2025
Relies on the availability of a high-quality machine translation system or pre-existing parallel corpora for the target language
- LUSIFER: Language Universal Space Integration for Enhanced Representation in Multilingual Text Embedding ModelsSIGIR 2025
English performance (57.20) is below English-centric baselines like E5-Mistral (66.64) and BGE-en-1.5 (63.27)
- Is Translation All You Need? A Study on Solving Multilingual Tasks with Large Language ModelsNAACL 2025
External translation strategies require additional computational overhead and dependency on third-party APIs or secondary models like NLLB.
- IndiGEC: Multilingual Grammar Error Correction for Low-Resource Indian LanguagesEMNLP 2025
Depends on availability of pretrained translation (IndicTrans2) and masked language models; for extremely low-resource languages lacking these, data quality degrades
- Improving Informally Romanized Language IdentificationEMNLP 2025
Only 9 of 20 languages have any harvested natural romanized data (Bodo, Kashmiri, Maithili and Sanskrit have none in MADLAD-400 or GlotCC), limiting the harvested-data component
- How do Multimodal Foundation Models Encode Text and Speech? An Analysis of Cross-Lingual and Cross-Modal RepresentationsNAACL 2025
Only three models analyzed and prompts not varied for SALMONN; six of 30 languages are not shared between SONAR and the other models, making cross-model comparisons imperfect.
- Explaining and Mitigating Crosslingual Tokenizer InequitiesNeurIPS 2025
Training on parallel data (NLLB) reduces CTC by only ~1% on average, and comparisons were only against an English baseline for 7 languages due to parallel data availability.
- EnAnchored-X2X: English-Anchored Optimization for Many-to-Many TranslationEMNLP 2025
Not tested on low-resource languages, where the method fails because it depends on abundant English-centric parallel data and adequate model en2x ability; synthetic quality is bounded by en2x proficiency
- Efficiently Democratizing Medical LLMs for 50 Languages via a Mixture of Language Family ExpertsICLR 2025
The 2.5B token training set is heavily skewed toward English and Chinese, potentially limiting the depth of medical knowledge for the rarest languages.
- Editing Across Languages: A Survey of Multilingual Knowledge EditingEMNLP 2025
Lack of evaluation coverage for low-resource, morphologically complex, and culturally specific languages (e.g., Swahili, Urdu); benchmarks largely rely on machine translation and assume knowledge is language-invariant.
- Do Large Language Models have an English Accent? Evaluating and Improving the Naturalness of Multilingual LLMsACL 2025
Requires native, human-written reference corpora and reliable word tokenizers and dependency parsers, so the approach does not extend to low-resource languages where such resources are unavailable
- Cross-lingual Transfer of Reward Models in Multilingual AlignmentNAACL 2025
Experiments use only 3B RMs with 86k instances and four target languages, so effects of scale, data size, and mixture are untested (though 3B RMs matched ArmoRM trained on 550k+ instances)
- Conan-Embedding-v2: Training an LLM from Scratch for Text EmbeddingsEMNLP 2025
Performance is highly uneven across languages: Arabic (65.2) and Korean (67.5) score far below Germanic/Slavic languages (~74-75.5), suggesting linguistic similarity to English, not data volume, drives effectiveness
- CC-Tuning: A Cross-Lingual Connection Mechanism for Improving Joint Multilingual Supervised Fine-TuningACL 2025
Requires parallel (non-English, English) training pairs, so it cannot be applied to monolingual-only instruction data
- Brain-Informed Fine-Tuning for Improved Multilingual Understanding in Language ModelsNeurIPS 2025
Small data volume per participant (only 2756 TRs of fine-tuning data from 6 bilingual participants), which constrains the magnitude of performance gains
- Beyond WER: Probing Whisper’s Sub‐token Decoder Across Diverse Language Resource LevelsEMNLP 2025
Only ~10 minutes of Common Voice audio per language (validated to cover ~73% of unique sub-tokens that appear in 60 minutes), a small snapshot that may miss phonological and domain diversity.
- AlignX: Advancing Multilingual Large Language Models with Multilingual Representation AlignmentEMNLP 2025
Performance gains are most pronounced in non-English directions, with smaller relative improvements for English-centric tasks.
- AlignFreeze: Navigating the Impact of Realignment on the Layers of Multilingual Models Across Diverse LanguagesNAACL 2025
Language coverage is skewed toward high-resource, mostly European languages (34-35 languages), and only encoder-only models and two coarse freezing strategies were tested
- When Is Multilinguality a Curse? Language Modeling for 250 High- and Low-Resource LanguagesEMNLP 2024
Evaluation is limited to raw language modeling log-likelihood; results may not transfer to downstream tasks, fine-tuned cross-lingual transfer, or machine translation
- Wav2Gloss: Generating Interlinear Glossed Text from SpeechACL 2024
Dataset is small and unevenly distributed: 41.8 training hours across 22 seen languages (several languages have <0.2 hours), and speaker overlap between splits exists because many corpora have only 1-2 speakers.
- Using Language Models to Disambiguate Lexical Choices in TranslationEMNLP 2024
Test sets are small—only 120–180 expert-curated examples per language and 1,377 overall—because human annotation of lexical selection is labor-intensive
- The Hidden Space of Transformer Language AdaptersACL 2024
Most experiments use models trained from scratch on a relatively small corpus; extension to existing large multilingual models like mGPT is left to future work, so findings may not transfer to production-scale multilingual LMs.
- The Effect of Data Partitioning Strategy on Model Generalizability: A Case Study of Morphological SegmentationNAACL 2024
No validation/tuning set was used (indigenous-language data too small after test-set carve-outs), so models may be under-tuned, which the authors note may itself signal poor generalization.
- Teaching LLMs to Abstain across Languages via Multilingual FeedbackEMNLP 2024
Performance remains weakest for Tamil and Malayalam (0.484 and 0.492 average vs. 0.520 low-resource average), and abstain decisions transfer poorly across unrelated or low-resource languages (only 48.4% overlap for low-resource language groups)
- TEMA: Token Embeddings Mapping for Enriching Low-Resource Language ModelsEMNLP 2024
Requires high-resource L1 models which may be computationally expensive to pre-train if not already available.
- RLHF Can Speak Many Languages: Unlocking Multilingual Preference Optimization for LLMsEMNLP 2024
Coverage limited to 23 languages (no dialects distinguished), and single runs take ~5 days on 8x H100 80GB GPUs, restricting experiments to 8B scale
- MultiParaDetox: Extending Text Detoxification with Parallel Data to New LanguagesNAACL 2024
Dataset sizes are highly uneven across languages (8,500 unique inputs Russian vs 337 Spanish), causing a marked drop in Spanish results (J=0.239, near human references at 0.224)
- Modularized Multilingual NMT with Fine-grained InterlinguaNAACL 2024
Only 5 languages (10 English-centric pairs) on MultiUN with simulated balanced/unbalanced data ratios; no low-resource languages and no external benchmarks like FLORES-101 or WMT
- MMTE: Corpus and Metrics for Evaluating Machine Translation Quality of Metaphorical LanguageEMNLP 2024
Corpus is small (647 annotated translations per language pair from 647 source sentences) and only covers EN-ZH and EN-IT; language typology effects remain a working hypothesis rather than a verified conclusion
- Language-Specific Neurons: The Key to Multilingual Capabilities in Large Language ModelsACL 2024
Study covers only 7 languages on pre-trained foundation models (excluding instruction-tuned variants from primary analysis), so scaling to many languages and differences among language groups remain unexamined
- Language Concept Erasure for Language-invariant Dense RetrievalEMNLP 2024
Only evaluated in zero-shot settings (no fine-tuning on evaluation training splits), and covers only 16 training languages out of 7,000+ spoken languages.
- LANDeRMT: Dectecting and Routing Language-Aware Neurons for Selectively Finetuning LLMs to Machine TranslationACL 2024
Low absolute BLEU scores overall (e.g., 16.97 for en-ar), suggesting the setting uses limited 200k-sentence-per-direction WMT data and small/comparable-capacity baselines
- Key ingredients for effective zero-shot cross-lingual knowledge transfer in generative tasksNAACL 2024
Coverage is limited: only three encoder-decoder models, no full per-combination LR grid search, no decoder-only models, and mmT5 excluded; findings may not generalize beyond the two tested tasks (summarization, QA) and the five target languages per task
- How Good is Zero-Shot MT Evaluation for Low Resource Indian Languages?ACL 2024
Small evaluation dataset size limited to 250 human-annotated segments per target language.
- Global Gallery: The Fine Art of Painting Culture Portraits through Multilingual Instruction TuningNAACL 2024
The study relies on machine-translated data for instruction tuning, which may miss subtle linguistic-cultural pragmatics despite high reference-free quality scores.
- Getting The Most Out of Your Training Data: Exploring Unsupervised Tasks for Morphological InflectionEMNLP 2024
Results are sensitive to vocabulary and data sizes, and language-level disparities in gains are not analyzed; the low-resource setting uses only 1k training samples per language
- Getting More from Less: Large Language Models are Good Spontaneous Multilingual LearnersEMNLP 2024
The alignment stability is sensitive to the choice of source languages, with low-resource languages providing less consistent generalization.
- Enhancing Multilingual Capabilities of Large Language Models through Self-Distillation from Resource-Rich LanguagesACL 2024
Relies on an external machine translation system (NLLB-200-3.3B) plus quality filtering, so gains depend on MT quality and remain limited for extremely low-resource languages; the authors note vocabulary expansion or continued pre-training may be needed there
Experiments use a single source language and one target language at a time; multi-source/multi-target interactions are untested
- Distillation for Multilingual Information RetrievalSIGIR 2024
Training on languages beyond the target collection's (or a union of both collections' languages) generally hurts effectiveness versus training only on evaluation languages, suggesting model capability or translation-artifact limits
- Democratizing LLMs for Low-Resource Languages by Leveraging their English Dominant Abilities with Linguistically-Diverse PromptsACL 2024
Requires unlabeled text in the target low-resource language in the context (back-translation prompts) for En→X tasks to work effectively, so it does not eliminate the need for target-language monolingual data
- Decoupled Vocabulary Learning Enables Zero-Shot Translation from Unseen LanguagesACL 2024
Requires high-quality fastText monolingual embeddings and bilingual dictionaries for alignment, unavailable for endangered, primarily spoken, or orthography-less languages; performance drops when fastText needs special tokenization (Chinese, Japanese, Vietnamese, Korean)
- DM-BLI: Dynamic Multiple Subspaces Alignment for Unsupervised Bilingual Lexicon InductionACL 2024
Evaluated only on relatively low-resource languages (Finnish, Hindi, Turkish, etc.), not truly low-resource languages where embedding quality itself is poor
- Cross-lingual Back-Parsing: Utterance Synthesis from Meaning Representation for Zero-Resource Semantic ParsingEMNLP 2024
Evaluated with only one mPLM (mT5-large) and no actual low-resource languages, due to dataset language coverage
- Code-Switching Can be Better Aligners: Advancing Cross-Lingual SLU through Representation-Level and Prediction-Level AlignmentACL 2024
Effectiveness tied to the underlying mPLM, which may not represent all languages equally well
- Can Watermarks Survive Translation? On the Cross-lingual Consistency of Text Watermark for Large Language ModelsACL 2024
Initial results are presented as 'preliminary,' suggesting the need for broader validation across more language pairs and model architectures.
- A Comparison of Language Modeling and Translation as Multilingual Pretraining ObjectivesEMNLP 2024
Small fixed model scale (12 layers, 512 hidden) and restricted data (only languages in both UNPC and OpenSubtitles), so findings may not transfer to larger models or other data regimes
- 1+1>2: Can Large Language Models Serve as Cross-Lingual Knowledge Aggregators?EMNLP 2024
Requires constructing a dedicated training dataset and training an external low-resource classifier for each supported language.
Relies on external machine translation APIs, which introduces potential translation error propagation and additional inference latency.
- Zero-shot Cross-lingual Transfer With Learned Projections Using Unlabeled Target-Language DataACL 2023
Requires target languages to be known a priori and needs unlabeled target text plus a precomputed SVD per language per layer.
- XSemPLR: Cross-Lingual Semantic Parsing in Multiple Natural Languages and Meaning RepresentationsACL 2023
All non-English data are translations from English datasets, so cultural variation and native-speaker information needs are not captured; coverage of tasks, languages, and MRs is necessarily partial
- Vicinal Risk Minimization for Few-Shot Cross-lingual Transfer in Abusive Language DetectionEMNLP 2023
Only the test set is multilingual; training/development data is English-only, forcing a partitioned test set for few-shot experiments (10-100 instances), which limits comparability with the original dataset paper's results
- Unifying Cross-Lingual and Cross-Modal Modeling Towards Weakly Supervised Multilingual Vision-Language Pre-trainingACL 2023
Requires parallel translation pairs (19M WikiMatrix pairs) in addition to unpaired text; languages lacking such resources may benefit less
- Towards a Better Understanding of Variations in Zero-Shot Neural Machine Translation PerformanceEMNLP 2023
High-resource languages are capped at 5M sentence pairs, so the resource gap between high and medium tiers is small relative to lower tiers, weakening resource-level conclusions
- Towards Leaving No Indic Language Behind: Building Monolingual Corpora, Benchmark and Models for Indic LanguagesACL 2023
Performance severely degrades on extremely low-resource 'left-behind' languages (e.g., Santhali and Sindhi score 25.9% and 17.7% lower than Hindi on IndicCOPA) due to scarce pretraining data and lack of shared script/linguistic relatives.
- The Best of Both Worlds: Combining Human and Machine Translations for Multilingual Semantic Parsing with Active LearningACL 2023
Hyperparameters of the acquisition functions are tuned on a separate existing bilingual language pair (e.g. English-Thai for English-German), which may not transfer since such auxiliary datasets are often unavailable and languages differ
- Multilingual LLMs are Better Cross-lingual In-context Learners with AlignmentACL 2023
Relies on manually engineered task-aligning statements, which may not scale efficiently to all possible language-task combinations.
- Multilingual Holistic Bias: Extending Descriptors and Patterns to Unveil Demographic Biases in Languages at ScaleEMNLP 2023
The dataset is small per language (a few hundred sentences, ~20,500 total) relative to standard MT benchmarks, too small for per-language toxicity analysis, and the 'friends' template may confound results since masculine plural can denote mixed-gender groups.
- Massively Multilingual Lexical Specialization of Multilingual TransformersACL 2023
Coverage restricted to 50 languages and two base-size MMTs (mBERT, XLM-R Base); results may differ substantially for larger MMTs like XLM-R Large or mT5
- MULTITuDE: Large-Scale Multilingual Machine-Generated Text Detection BenchmarkEMNLP 2023
Data is limited to news articles in a single domain with only ~1k human texts per training language; detectors may fail on other domains, writing styles, dialects, slang, or code-switching
- MRN: Multiplexed Routing Network for Incremental Multilingual Text RecognitionICCV 2023
Performance depends on language introduction order: with the Eastern-Asia languages introduced first (O1), MRN averages 78.4% on MLT17 vs 82.9% with them grouped last (O3), so naive ordering loses up to ~4.5 points
- MAD-TSC: A Multilingual Aligned News Dataset for Target-dependent Sentiment ClassificationACL 2023
Cross-language comparison is confounded by variable quality of available pretrained language models, so score differences do not purely reflect task difficulty.
- Learning Language-Specific Layers for Multilingual Machine TranslationACL 2023
LSLs were only explored in the encoder; decoder placement gave no improvements and adding new languages by training only LSLs was not validated
- Hybrid Knowledge Transfer for Improved Cross-Lingual Event Detection via Hierarchical Sample SelectionACL 2023
Depends on a pre-trained multilingual encoder covering both source and target languages, excluding the vast majority of the world's ~7000 languages
Requires unlabeled target-language data, which can be unavailable for extremely low-resource languages
- GradSim: Gradient-Based Language Grouping for Effective Multilingual TrainingEMNLP 2023
Language coverage is limited (12 African languages, 15 WikiAnn, 27 UD languages), so results may not represent all languages; gradients come from a short initial joint training run, which is not analyzed for sensitivity to that run's length or data amount
- GlobalBench: A Benchmark for Global Progress in Natural Language ProcessingEMNLP 2023
Datasets from different sources/genres within a task are not difficulty-normalized, so cross-language comparisons may not be valid; utility normalization by empirical maximum is approximate
- FOCUS: Effective Embedding Initialization for Monolingual Specialization of Multilingual ModelsEMNLP 2023
Requires monolingual target-language text for both auxiliary fastText training and LAPT, unavailable for many very low-resource languages
- Exploring Linguistic Probes for Morphological GeneralizationEMNLP 2023
Designing language-specific probes requires expert linguistic and domain knowledge per language, limiting scalability; the authors propose community/shared-task extension as future work.
- Enhancing Code-Switching for Cross-lingual SLU: A Unified View of Semantic and Grammatical CoherenceEMNLP 2023
Performance relies heavily on the correctness/quality of bilingual dictionaries, which may not exist or be accurate for many language pairs
- Empowering Cross-lingual Behavioral Testing of NLP Models with Typological FeaturesACL 2023
Template creation requires native speaker expertise and knowledge of the language's grammar, so it does not scale without human effort
- Don’t Trust ChatGPT when your Question is not in English: A Study of Multilingual Abilities and Types of LLMsEMNLP 2023
Analysis is heavily centered on ChatGPT, which may not represent the behavior of smaller or differently pre-trained multilingual models.
- Dialect-robust Evaluation of Generated TextACL 2023
Evaluation corpora are small — thousands of examples for Mandarin and Portuguese and only hundreds for English — insufficient to capture the full extent of these languages
- Crosslingual Generalization through Multitask FinetuningACL 2023
xP3 covers only 46 languages (mT5 pretrains on 101); performance correlates with pretraining language share, so low-resource languages see much smaller gains.
- Cross-lingual Transfer Can Worsen Bias in Sentiment AnalysisEMNLP 2023
Bias benchmarks exist in only five languages, so interactions between source/target language pairs (e.g., shared alphabets) cannot be tested
- Cross-lingual Prompting: Improving Zero-shot Chain-of-Thought Reasoning across LanguagesEMNLP 2023
Performance remains significantly lower on low-resource languages (e.g., 38.8% accuracy for Telugu vs 82.4% for Spanish)
- CoLaDa: A Collaborative Label Denoising Framework for Cross-lingual Named Entity RecognitionACL 2023
Requires a machine translation system and target-language unlabeled text; inapplicable to languages lacking either
- Clustering Pseudo Language Family in Multilingual Translation Models with Fisher Information MatrixEMNLP 2023
Sometimes selects linguistically divergent auxiliary languages (e.g., Korean for Farsi→en) whose benefit is explained only post-hoc via data similarity, not validated causally
- CS2W: A Chinese Spoken-to-Written Style Conversion Dataset with Multiple Conversion TypesEMNLP 2023
The dataset is small (7,237 sentences, 724 test) and Chinese-Mandarin telephone-conversation specific, so results may not generalize to other domains or languages
- Beyond English-Centric Bitexts for Better Multilingual Language Representation LearningACL 2023
Scaling does not improve performance on unseen extremely low-resource languages: on AmericasNLI (10 Indigenous languages), XY-LENTBase (45.5) and XY-LENTXL (47.2) are nearly equal and far below XNLI performance
- BasahaCorpus: An Expanded Linguistic Resource for Readability Assessment in Central Philippine LanguagesEMNLP 2023
Datasets are small and imbalanced (e.g., 22 Hiligaynon L2 documents, 27 Minasbate L1 vs 36 Karay-a L2), which may make accuracy estimates unreliable
- BLOOM+1: Adding Language Support to BLOOM for Zero-Shot PromptingACL 2023
Effective adaptation requires ~100M tokens of the new language, so truly low-resource languages (e.g., Guarani with 30K sentences) benefit only when models exceed 3B parameters and may lack sufficient unlabeled data entirely.
- A Systematic Study of Performance Disparities in Multilingual Task-Oriented Dialogue SystemsEMNLP 2023
Analysis restricted to one dataset (MULTI3WOZ) and four languages, text-only (no speech/ASR), and assumes uniform annotation cost across languages
- MAD-X: An Adapter-Based Framework for Multi-Task Cross-Lingual TransferEMNLP 2020
Requires access to unlabeled text in the target language to perform MLM-based language adaptation.
- What Matters for Neural Cross-Lingual Named Entity Recognition: An Empirical AnalysisEMNLP 2019
Performance is heavily dependent on the coverage and quality of external bilingual dictionaries and pre-trained word embeddings.
- Unsupervised Multilingual Word Embedding with Limited Resources using Neural Language ModelsACL 2019
Currently lacks subword or character-level information, limiting performance on morphologically rich languages.
- Cheap Translation for Cross-Lingual Named Entity RecognitionEMNLP 2017
Significant sensitivity to dictionary quality and coverage, particularly for languages with complex morphology or non-Latin scripts.
Requires external language-specific tools like stemmers or name-substitution gazetteers to achieve peak performance in extremely low-resource settings.
Nearest existing work
- Expanding Pretrained Models to Thousands More Languages via Lexicon-based AdaptationACL 2022
- Translation or Recitation? Calibrating Evaluation Scores for Machine Translation of Extremely Low-Resource LanguagesACL 2026
- To Translate or Not to Translate: A Systematic Investigation of Translation-Based Cross-Lingual Transfer to Low-Resource LanguagesNAACL 2024
- From Zero to Hero: On the Limitations of Zero-Shot Language Transfer with Multilingual TransformersEMNLP 2020
- Universal Neural Machine Translation for Extremely Low Resource LanguagesNAACL 2018
- Match the Script, Adapt if Multilingual: Analyzing the Effect of Multilingual Pretraining on Cross-lingual TransferabilityACL 2022
- UniBridge: A Unified Approach to Cross-Lingual Transfer Learning for Low-Resource LanguagesACL 2024
- Unifying Cross-Lingual Transfer across Scenarios of Resource ScarcityEMNLP 2023
- P-MMEval: A Parallel Multilingual Multitask Benchmark for Consistent Evaluation of LLMsEMNLP 2025
- Subword Evenness (SuE) as a Predictor of Cross-lingual Transfer to Low-resource LanguagesEMNLP 2022
- Languages Still Left Behind: Toward a Better Multilingual Machine Translation BenchmarkEMNLP 2025
- Small Data, Big Impact: Leveraging Minimal Data for Effective Machine TranslationACL 2023
- GradSim: Gradient-Based Language Grouping for Effective Multilingual TrainingEMNLP 2023
- Harnessing Multilinguality in Unsupervised Machine Translation for Rare LanguagesNAACL 2021
- Make the Best of Cross-lingual Transfer: Evidence from POS Tagging with over 100 LanguagesACL 2022