Omnilingual SONAR: Cross-Lingual and Cross-Modal Sentence Embeddings Bridging Massively Multilingual Text and Speech

Alexandre Mourachko; Artyom Kozhevnikov; Belen Alastruey; Christophe Ropers; David Dale; Guillem Ram\'irez; Holger Schwenk; Ioannis Tsiamas; Jaehyeong Jo; Kevin Heffernan

OmniSONAR builds one semantic space for thousands of languages plus speech, code and math that halves cross-lingual search errors and still supports strong decoding and transfer.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-13 23:38 UTC pith:JRWYOJOS

load-bearing objection Large, carefully engineered omnilingual embedding space with real SOTA gains on FLORES/BIBLE and a clean English-only Spectrum transfer story; soft spots are long-tail domain and missing artifacts, not the core claims.

arxiv 2603.16606 v3 pith:JRWYOJOS submitted 2026-03-17 cs.CL

Omnilingual SONAR: Cross-Lingual and Cross-Modal Sentence Embeddings Bridging Massively Multilingual Text and Speech

Omnilingual SONAR Team: Jo\~ao Maria Janeiro , Pere-Llu\'is Huguet Cabot , Ioannis Tsiamas , Yen Meng , Vivek Iyer , Guillem Ram\'irez , Loic Barrault , Belen Alastruey

show 11 more authors

Xiang "Tony" Cao Yu-An Chung Marta R. Costa-Jussa David Dale Kevin Heffernan Jaehyeong Jo Artyom Kozhevnikov Alexandre Mourachko Christophe Ropers Holger Schwenk Paul-Ambroise Duquenne

This is my paper

classification cs.CL

keywords multilingual sentence embeddingscross-lingual alignmentcross-modal speechteacher-student distillationomnilingualcontrastive learninghard negativeslarge concept models

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Cross-lingual sentence encoders have been stuck covering only a few hundred languages and often lose downstream quality when they force languages into alignment. OmniSONAR claims that progressive training breaks that trade-off: first a strong 200-language foundation is learned with an LLM-initialized encoder-decoder, token-level decoding, a novel split-softmax contrastive loss and synthetic hard negatives; then thousands more language varieties are folded in by teacher-student distillation that mixes MSE and contrastive signals; finally 177 spoken languages are mapped into the same space from ASR data alone. The resulting embeddings cut similarity-search error in half on the 200-language FLORES set and by a factor of fifteen on a 1,560-language Bible benchmark, while the paired decoder beats specialized translation systems and much larger language models on massively multilingual into-English translation. Speech embeddings nearly match a strong speech-translation baseline despite never being trained on translation pairs. A language model trained only on English OmniSONAR sequences then transfers complex document tasks to thousands of languages and to speech, showing the space can serve as a language- and modality-agnostic foundation rather than just a retrieval tool.

Core claim

A progressive recipe—foundational 200-language encoder-decoder training with decoding plus split-softmax contrastive learning, followed by hybrid MSE-contrastive teacher-student expansion and speech distillation—produces a single fixed-size embedding space that natively covers thousands of text varieties, 177 spoken languages, code and math, while improving alignment and preserving or raising downstream and decoding quality instead of trading them away.

What carries the argument

Progressive training anchored by the split-softmax contrastive loss (separate in-batch margin term and hard-negative term) and by two-stage teacher-student distillation with language-type-dependent MSE and bidirectional contrastive weights. The foundation model acts as a frozen geometric teacher; new languages and speech are projected into its manifold without retraining the whole space from scratch.

Load-bearing premise

That a frozen 200-language teacher plus MSE-and-contrastive distillation on pairs that always include at least one foundational language will keep the geometry faithful for thousands of extremely low-resource varieties, even when most of their data are Bible-scale or synthetic.

What would settle it

Hold out a set of low-resource languages that have neither Bible data nor close relatives among the foundational 200; retrain or evaluate the omnilingual expansion on them. If similarity-search error and zero-shot decoding quality collapse relative to the reported Bible and FLORES numbers, the lossless-expansion claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

Cross-lingual retrieval and parallel-data mining become practical at 1,500+ languages with error rates previously seen only on high-resource pairs.
Fixed-size sentence vectors can drive massively multilingual text and speech translation without full-sequence encoder attention at decode time.
Document-level language models trained only on English OmniSONAR sequences obtain strong zero-shot coverage of thousands of languages and of speech.
Smaller distilled encoders remain drop-in replacements in the same 1024-dimensional space, so one task head works across compute budgets.
Code snippets and mathematical expressions sit natively in the same space as natural language for retrieval and explanation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same progressive distillation pattern could absorb further modalities (vision, sign language) without rebuilding the core text space.
Because transfer is strongest within linguistic families, future scaling will likely need family-aware or sparse routing rather than pure data volume.
Spectrum-style front-ends suggest multilingual and multimodal LLMs could largely offload coverage to a frozen omnilingual embedding encoder trained once.
Heavy reliance on Bible and synthetic long-tail data may overstate robustness on everyday domains for the lowest-resource languages.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

Large, carefully engineered omnilingual embedding space with real SOTA gains on FLORES/BIBLE and a clean English-only Spectrum transfer story; soft spots are long-tail domain and missing artifacts, not the core claims.

read the letter

This is the real systems paper on massively multilingual sentence embeddings. Progressive foundation (LLM-init encoder-decoder + split-softmax + hard negatives on 200 langs) then teacher-student expansion to 4k+ varieties, plus speech and Spectrum, is the actual advance over SONAR/LaBSE. The numbers are large and consistent: ~half the xsim++ error on FLORES-200, ~15× lower xsim on 1,560-language BIBLE, decoder beating NLLB-3B and much larger LLMs by ~15 chrF++ on BIBLE→En, speech cutting cross-modal error ~43% while hitting ~97% of SeamlessM4T zero-shot for ST. Spectrum trained only on English OmniSONAR sequences transferring to XBelebele/Speech-XBelebele is the cleanest demonstration that the space is usable for longer reasoning.

What they do well: ablations isolate each piece (Tables 11–16, Fig. 3–4, Table 14–15). Dynamic λ/τ by foundational vs new language, vocabulary warm-up, and the “always include a foundational language” distillation rule are justified, not hand-waved. Table 15 shows the expansion does not trash the 200-lang base. Relatedness-driven zero-shot transfer (Sec. 8.2) is honest about when omnilinguality helps vs hurts. Code/math and XLCoST results are a nice extra.

Soft spots, in proportion: long-tail data is heavily Bible/synthetic/mined; domain shift outside religious text is a residual risk the paper itself flags. No public code/checkpoints yet, so reimplementation cost is high. Hyperparameters (α/β/γ, m, τ, sampling) are free but standard for this genre. None of that undercuts the central similarity-search or decoding claims as stated; the stress-test concern about geometric collapse is already addressed by their own FLORES-vs-BIBLE comparison.

Who it is for: anyone doing bitext mining, low-resource MT, or multimodal LLM transfer. Cite it if you work in that space. Serious referee time is warranted; this is not desk-reject material. I would bring it to reading group and expect to cite the models and Spectrum recipe within a year once artifacts appear.

Referee Report

0 major / 5 minor

Summary. The paper introduces OmniSONAR, a progressive family of cross-lingual and cross-modal sentence embedding models that first builds a 200-language foundational space from an LLM-initialized encoder-decoder (token-level decoding + novel split-softmax contrastive loss with synthetic hard negatives), then expands via two-stage teacher-student distillation (MSE + bidirectional contrastive, with tokenizer warm-up) to 4,200+ language varieties, and finally maps 177 spoken languages into the same space via ASR-only MSE distillation. The central empirical claims are that this space halves FLORES-200 xsim++ error relative to prior SOTA, reduces BIBLE (1,560 languages) xsim error by ~15×, enables decoder-based X→En translation that beats NLLB-3B and multi-billion LLMs by ~15 chrF++ on BIBLE, reaches strong MTEB/XLCoST numbers, and yields a speech encoder with 43% lower cross-modal search error that attains ~97% of SeamlessM4T ST quality zero-shot. Spectrum further shows that an English-only encoder-decoder LM over OmniSONAR sequences transfers zero-shot to multilingual text and speech reasoning benchmarks.

Significance. If the reported numbers hold under independent re-evaluation, this is a substantial advance for massively multilingual representation learning: it simultaneously scales coverage by an order of magnitude, largely closes the historical alignment-vs-quality trade-off, and demonstrates that a frozen high-resource teacher plus carefully weighted MSE/contrastive distillation can expand without catastrophic forgetting (Table 15). The progressive recipe, split-softmax objective, omnilingual tokenizer warm-up, and Spectrum transfer results are concrete, reusable contributions. Extensive ablations (Tables 11–16, Figures 3–4) and multi-benchmark evaluation (FLORES, FLORES+, BOUQuET, AfroMT, BIBLE, MTEB, XLCoST, FLEURS) give the claims unusually strong empirical grounding for a systems paper of this scope.

minor comments (5)

Section 3.5 / Stage 4: the precise BLASER and early-encoder similarity thresholds (k values) and the exact composition of the 4,200 varieties are only partially tabulated; a short appendix table listing final thresholds and language-group sizes would improve reproducibility.
Equation (5) and surrounding text: the split-softmax formulation is clear, but a one-sentence intuition for why a single joint softmax with m=0 fails to converge (while the decoupled form succeeds) would help readers who skip the ablation.
Table 4 vs. Table 15: OmniSONAR-200 vs. final OmniSONAR numbers on FLORES are reported in slightly different places; a single consolidated row would make the “no forgetting” claim easier to verify at a glance.
Section 9 (Spectrum): the FLOPs analysis (Figure 8) is useful, but the exact sentence-segmentation model used at inference for non-English and speech inputs should be named in the main text rather than only in the appendix.
A few typographical inconsistencies remain (e.g., “Omninilingual”, “w/o Lang Tag” formatting, occasional missing spaces around ×). A light copy-edit pass would polish the camera-ready version.

Circularity Check

0 steps flagged

No significant circularity: SOTA error-rate and translation claims are measured on external held-out benchmarks against public baselines; progressive distillation targets frozen teacher embeddings, not the reported metrics.

full rationale

OmniSONAR is an empirical representation-learning paper whose central claims (halved FLORES xsim++ error, 15× BIBLE xsim reduction, decoder chrF++ gains, speech zero-shot ST quality) are evaluated on independent test sets (FLORES200/+, BIBLE, BOUQuET, AfroMT, FLEURS, MTEB, XLCoST) against external models (LaBSE, SONAR, MEXMA, NLLB, SeamlessM4T, LLMs). The progressive pipeline (Seq2Seq → contrastive+decoder → hard-negatives → teacher-student omnilingual extension → speech MSE) uses frozen OmniSONAR-200 teacher embeddings as distillation targets (Eqs. 7–12, dynamic λ/τ by language type); those targets are not the final reported metrics, and ablations (Tables 14–15, Fig. 3) plus zero-shot transfer analyses (Sec. 8) show the expansion is not forced by construction. Self-citations to prior SONAR/NLLB work supply architecture and data recipes but do not define or force the new error rates. No fitted parameter is renamed a prediction, no uniqueness theorem is imported to forbid alternatives, and no ansatz is smuggled via self-citation. The derivation chain is therefore self-contained against external evidence; residual domain-shift risk from Bible-scale data is a correctness concern, not circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 3 axioms · 3 invented entities

Empirical ML paper; load-bearing content is training recipe + data assumptions rather than mathematical axioms. Free parameters are the many loss weights and scales chosen by ablation. Domain assumptions are standard contrastive/teacher-student premises. Invented entities are the named losses, stages and Spectrum architecture.

free parameters (4)

contrastive weight α / translation weight β / hard-neg weight γ
Set to 0.05/1/0.8 (and stage-specific variants) by ablation; directly control the embedding geometry that produces the reported xsim gains.
additive margin m and logit scale τ
m=0.3, τ=100 (foundational) or 10/60 (omnilingual); chosen empirically and shown critical in Table 12.
MSE vs contrastive λ weights and language-drop probability
Different λ schedules for foundational vs new languages (Table 3); free choices that balance forgetting vs new-language learning.
temperature-based sampling β_L / β_D = 0.5 and BLASER/similarity thresholds
Control data mixture and filtering; affect which pairs enter the final space.

axioms (3)

domain assumption Token-level translation decoding plus contrastive alignment on pooled embeddings yields semantically rich fixed-size vectors that remain decodable.
Inherited from SONAR/LASER literature and used throughout Stages 1–3; not re-proved.
domain assumption A frozen high-resource teacher plus MSE+InfoNCE distillation on mixed pairs projects new languages into the same manifold without collapse.
Core premise of Stage 4 (Section 4.5); supported by ablations but remains an empirical assumption for the long tail.
ad hoc to paper Synthetic hard negatives generated by LLaMA3.3 and ASR-only speech pairs are sufficiently faithful for contrastive and MSE objectives.
Data-generation pipeline (Sections 3.6, A.3–A.4) is paper-specific; quality is validated only by downstream metrics.

invented entities (3)

split-softmax contrastive loss no independent evidence
purpose: Simultaneously train margin-based in-batch negatives and non-margin hard negatives without destabilizing either signal.
Novel objective (Eq. 5) introduced to make hard-negative training converge; independent evidence is the ablation gain on xsim++.
Spectrum encoder-decoder no independent evidence
purpose: Language-model over sequences of OmniSONAR embeddings, trained only on English, for zero-shot multilingual/speech transfer.
New architecture (Section 9); independent evidence is the reported XBelebele/Speech-SIB gains.
OmniSONAR progressive training stages + omnilingual tokenizer warm-up no independent evidence
purpose: Scale from 200 to 4k+ languages without representation collapse.
Named pipeline (Figure 1, Section 4); evidence is the before/after BIBLE and FLORES numbers.

pith-pipeline@v1.1.0-grok45 · 57437 in / 2954 out tokens · 33301 ms · 2026-07-13T23:38:02.407122+00:00 · methodology

0 comments

read the original abstract

Cross-lingual sentence encoders typically cover only a few hundred languages and often trade downstream quality for stronger alignment, limiting their adoption. We introduce OmniSONAR, a new family of omnilingual, cross-lingual and cross-modal sentence embedding models that natively embed text, speech, code, and mathematical expressions in a single semantic space, while delivering state-of-the-art downstream performance at the scale of thousands of languages, from high-resource to extremely low-resource varieties. To reach this scale without representation collapse, we use progressive training. We first learn a strong foundational space for 200 languages with an LLM-initialized encoder-decoder, combining token-level decoding with a novel split-softmax contrastive loss and synthetic hard negatives. Building on this foundation, we expand to several thousands language varieties via a two-stage teacher-student encoder distillation framework. Finally, we demonstrate the cross-modal extensibility of this space by seamlessly mapping 177 spoken languages into it. OmniSONAR halves cross-lingual similarity search error on the 200-language FLORES dataset and reduces error by a factor of 15 on the 1,560-language BIBLE benchmark. It also enables strong translation, outperforming NLLB-3B on multilingual benchmarks and exceeding prior models (including much larger LLMs) by 15 chrF++ points on 1,560 languages into English BIBLE translation. OmniSONAR also performs strongly on MTEB and XLCoST. For speech, OmniSONAR achieves a 43% lower similarity-search error and reaches 97% of SeamlessM4T speech-to-text quality, despite being zero-shot for translation (trained only on ASR data). Finally, by training an encoder-decoder LM, Spectrum, exclusively on English text processing OmniSONAR embedding sequences, we unlock high-performance transfer to thousands of languages and speech for complex downstream tasks.

discussion (0)

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Text Corpora as Concept Fields: Black-Box Hallucination and Novelty Measurement
cs.CL 2026-05 unverdicted novelty 7.0

Concept Fields model text corpora as local Gaussian drift fields in embedding space to score sentence transitions for hallucination detection and novelty via standardized deviation.
Text Corpora as Concept Fields: Black-Box Hallucination and Novelty Measurement
cs.CL 2026-05 unverdicted novelty 6.0

Concept Fields model text corpora as local Gaussian drift fields in embedding space to score sentence transitions for groundedness and novelty without model internals.

Reference graph

Works this paper leans on

31 extracted references · 2 canonical work pages · cited by 1 Pith paper · 1 internal anchor

[1]

https://openreview.net/forum?id=IW1PR7vEBf. BigScience Workshop, Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilić, Daniel Hesslow, Ro- man Castagné, Alexandra Sasha Luccioni, François Yvon, Matthias Gallé, Jonathan Tow, Alexander M. Rush, Stella Biderman, Albert Webson, Pawan Sasanka Ammanamanchi, Thomas Wang, Benoît Sagot, Niklas ...

Pith/arXiv arXiv doi:10.18653/v1/2024 2023
[2]

https://aclanthology.org/2022.tacl-1.30/

doi: 10.1162/tacl_a_00474. https://aclanthology.org/2022.tacl-1.30/. Nuno M. Guerreiro, Ricardo Rei, Daan van Stigt, Luisa Coheur, Pierre Colombo, and André F. T. Martins. xCOMET: Transparent Machine Translation Evaluation through Fine-grained Error Detection. �� ...

Pith/arXiv arXiv doi:10.1162/tacl_a_00474 2022
[3]

58 Oddur Kjartansson, Supheakmungkol Sarin, Knot Pipatsrisawat, Martin Jansche, and Linne Ha

doi: 10.1109/JSTSP.2022.3192714. 58 Oddur Kjartansson, Supheakmungkol Sarin, Knot Pipatsrisawat, Martin Jansche, and Linne Ha. Crowd-Sourced Speech Corpora for Javanese, Sundanese, Sinhala, Nepali, and Bangladeshi Bengali. In �� ...

work page doi:10.1109/jstsp.2022.3192714 2022
[4]

ISBN 979-8-89176-341-8

Association for Computational Linguistics. ISBN 979-8-89176-341-8. doi: 10.18653/v1/2025.wmt-1.22. https: //aclanthology.org/2025.wmt-1.22/. Sneha Kudugunta, Isaac Caswell, Biao Zhang, Xavier Garcia, Derrick Xin, Aditya Kusupati, Romi Stella, Ankur Bapna, and Orhan Firat. MADLAD-400: A Multilingual And Document-Level Large Audited Dataset. In A. Oh, T. Na...

Pith/arXiv arXiv doi:10.18653/v1/2025.wmt-1.22 2025
[5]

Jean Maillard, Cynthia Gao, Elahe Kalbassi, Kaushik Ram Sadagopan, Vedanuj Goswami, Philipp Koehn, Angela Fan, and Francisco Guzmán

https://aclanthology.org/2025.naacl-short.36/. Jean Maillard, Cynthia Gao, Elahe Kalbassi, Kaushik Ram Sadagopan, Vedanuj Goswami, Philipp Koehn, Angela Fan, and Francisco Guzmán. Small data, big impact: Leveraging minimal data for effective machine translation. In �� ...

doi:10.18653/v1/2024.emnlp-main.380 2025
[6]

ISBN 979-8-89176-195-7

Association for Computational Linguistics. ISBN 979-8-89176-195-7. doi: 10.18653/v1/2025.findings-naacl

doi:10.18653/v1/2025.findings-naacl 2025
[7]

Nasrin Mostafazadeh, Nathanael Chambers, Xiaodong He, Devi Parikh, Dhruv Batra, Lucy Vanderwende, Pushmeet Kohli, and James Allen

https://aclanthology.org/2025.findings-naacl.371/. Nasrin Mostafazadeh, Nathanael Chambers, Xiaodong He, Devi Parikh, Dhruv Batra, Lucy Vanderwende, Pushmeet Kohli, and James Allen. A corpus and cloze evaluation for deeper understanding of commonsense stories. In Kevin Knight, Ani Nenkova, and Owen Rambow, editors, �� ...

Pith/arXiv arXiv doi:10.18653/v1/n16-1098 2025
[8]

ISBN 979-8-89176-251-0

Association for Computational Linguistics. ISBN 979-8-89176-251-0. doi: 10.18653/v1/2025.acl-long.919. https://aclanthology.org/2025.acl-long.919/. Keshan Sodimana, Knot Pipatsrisawat, Linne Ha, Martin Jansche, Oddur Kjartansson, Pasindu De Silva, and Supheakmungkol Sarin. A Step-by-Step Process for Building TTS Voices Using Open Source Data and Frame- wo...

doi:10.18653/v1/2025.acl-long.919 2025
[9]

63 Per Erik Solberg and Pablo Ortiz

https://arxiv.org/abs/2402.16829. 63 Per Erik Solberg and Pablo Ortiz. The Norwegian parliamentary speech corpus. In �� , 2022. Saba Sturua, Isabelle Mohr, Mohammad Kalim Akram, Michael Günther, Bo Wang, Markus Krimmel, ...

Pith/arXiv arXiv doi:10.1007/978-3-031-88720-8_21 2022
[10]

XNLI 2.0: Improving XNLI dataset and performance on Cross Lingual Understanding (XLU)

Association for Computational Linguistics. doi: 10.18653/v1/2023.eacl-main.108. https://aclanthology. org/2023.eacl-main.108/. Jörg Tiedemann. The Tatoeba Translation Challenge – Realistic Data Sets for Low Resource and Multilingual MT. In Loïc Barrault, Ondřej Bojar, Fethi Bougares, Rajen Chatterjee, Marta R. Costa-jussà, Christian Federmann, Mark Fishel...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.18653/v1/2023.eacl-main.108 2023
[11]

https://doi.org/10.24963/ijcai.2019/746

doi: 10.24963/ijcai.2019/746. https://doi.org/10.24963/ijcai.2019/746. Biao Zhang, Ankur Bapna, Rico Sennrich, and Orhan Firat. Share or Not? Learning to Schedule Language-Specific Capacity for Multilingual Translation. In �� , 2021. https: //openreview.net/fo...

Pith/arXiv arXiv doi:10.24963/ijcai.2019/746 2019
[12]

enable_points

https://openreview.net/forum?id=vfzRRjumpX. Dong Zhang, Shimin Li, Xin Zhang, Jun Zhan, Pengyu Wang, Yaqian Zhou, and Xipeng Qiu. Speechgpt: Empow- ering large language models with intrinsic cross-modal conversational abilities. In �� ...

doi:10.1162/tacl_a_00595 2023
[13]

The script checks whether 'ExitSignal' exists within the 'Actions' list of the dictionary linked to the 'session' key in the 'raw_action_logs' dictionary, and if so, assigns the result of a numpy operation (after squeezing) to the variable 'exit_signals_processed'
[14]

The script verifies if the event_label is 'Position' and the 'Anchored' value in the entry dictionary is True, then updates self.disable_anchor to True
[15]

The script checks if the event_category property of the trip_event object matches TripEventCategory.POSITION. JavaScript Example Source: const DataHandler = require('./lib/dataHandler.js'); let windowRef; const dataHandler = new DataHandler({}); Target: �� ...
[16]

The script imports a function or variable called buildHandler from a file at '../lib/dataHandler'
[17]

The script imports HandlerData from a file named dataHandler and uses it to instantiate a new Vuex Handler, assigning the result to a constant called handler
[18]

The script imports the initializeHandler function from the 'redux' package and also imports the reducer from './handlerReducer' to set up a handler. Math Example Source: G� → X Target: �� G �� n �� ...
[19]

The function G� maps every element of the set E to a corresponding element in the set S
[20]

There exists a mapping or function from the set or space G to the set or space H
[21]

Figure 14 Examples of hard negatives generation across Python code (blue), JavaScript code (orange), and mathe- matical expressions (green)

A function G, parameterized by λ, maps to or transforms into a space or set denoted as X, also parameterized by λ. Figure 14 Examples of hard negatives generation across Python code (blue), JavaScript code (orange), and mathe- matical expressions (green). 73 Hard Negatives Generation You are a text transformation specialist. Generate ONLY valid xsim++ tra...
[22]

ENTITY REPLACEMENT: • Swap proper nouns (people, locations, organizations) • Replace pronouns ( ��→ �� → ��)
[23]

Transformed sentence 1

NUMBER ALTERATION: • Change quantities ( ��→ ��) • Modify dates/times ( ��→ ��) • Alter percentages ( ��→ ��) Follow these patterns from training examples: {few-shot examples} Now transform THIS SPECIFIC INPUT SENTENCE using the above patterns. Output ONLY a Python list of 1-5 modified sentences in this exact format: [ "Transformed s...
[24]

Create 1-5 unique modified sentences
[25]

Maximize difference from original text
[26]

Mix transformation types where possible
[27]

Maintain grammatical correctness
[28]

Do NOT generate paraphrases, or synonyms
[29]

NEVER output empty strings
[30]

Output ONLY a Python list of strings
[31]

��

No explanations, headers, or additional text Input sentence to transform: {example} Figure 15 Prompt for generating xsim++ transformations with clear instructions and structure. 74 A.5 Language code correspondence �� ...

2024

[1] [1]

https://openreview.net/forum?id=IW1PR7vEBf. BigScience Workshop, Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilić, Daniel Hesslow, Ro- man Castagné, Alexandra Sasha Luccioni, François Yvon, Matthias Gallé, Jonathan Tow, Alexander M. Rush, Stella Biderman, Albert Webson, Pawan Sasanka Ammanamanchi, Thomas Wang, Benoît Sagot, Niklas ...

Pith/arXiv arXiv doi:10.18653/v1/2024 2023

[2] [2]

https://aclanthology.org/2022.tacl-1.30/

doi: 10.1162/tacl_a_00474. https://aclanthology.org/2022.tacl-1.30/. Nuno M. Guerreiro, Ricardo Rei, Daan van Stigt, Luisa Coheur, Pierre Colombo, and André F. T. Martins. xCOMET: Transparent Machine Translation Evaluation through Fine-grained Error Detection. �� ...

Pith/arXiv arXiv doi:10.1162/tacl_a_00474 2022

[3] [3]

58 Oddur Kjartansson, Supheakmungkol Sarin, Knot Pipatsrisawat, Martin Jansche, and Linne Ha

doi: 10.1109/JSTSP.2022.3192714. 58 Oddur Kjartansson, Supheakmungkol Sarin, Knot Pipatsrisawat, Martin Jansche, and Linne Ha. Crowd-Sourced Speech Corpora for Javanese, Sundanese, Sinhala, Nepali, and Bangladeshi Bengali. In �� ...

work page doi:10.1109/jstsp.2022.3192714 2022

[4] [4]

ISBN 979-8-89176-341-8

Association for Computational Linguistics. ISBN 979-8-89176-341-8. doi: 10.18653/v1/2025.wmt-1.22. https: //aclanthology.org/2025.wmt-1.22/. Sneha Kudugunta, Isaac Caswell, Biao Zhang, Xavier Garcia, Derrick Xin, Aditya Kusupati, Romi Stella, Ankur Bapna, and Orhan Firat. MADLAD-400: A Multilingual And Document-Level Large Audited Dataset. In A. Oh, T. Na...

Pith/arXiv arXiv doi:10.18653/v1/2025.wmt-1.22 2025

[5] [5]

Jean Maillard, Cynthia Gao, Elahe Kalbassi, Kaushik Ram Sadagopan, Vedanuj Goswami, Philipp Koehn, Angela Fan, and Francisco Guzmán

https://aclanthology.org/2025.naacl-short.36/. Jean Maillard, Cynthia Gao, Elahe Kalbassi, Kaushik Ram Sadagopan, Vedanuj Goswami, Philipp Koehn, Angela Fan, and Francisco Guzmán. Small data, big impact: Leveraging minimal data for effective machine translation. In �� ...

doi:10.18653/v1/2024.emnlp-main.380 2025

[6] [6]

ISBN 979-8-89176-195-7

Association for Computational Linguistics. ISBN 979-8-89176-195-7. doi: 10.18653/v1/2025.findings-naacl

doi:10.18653/v1/2025.findings-naacl 2025

[7] [7]

Nasrin Mostafazadeh, Nathanael Chambers, Xiaodong He, Devi Parikh, Dhruv Batra, Lucy Vanderwende, Pushmeet Kohli, and James Allen

https://aclanthology.org/2025.findings-naacl.371/. Nasrin Mostafazadeh, Nathanael Chambers, Xiaodong He, Devi Parikh, Dhruv Batra, Lucy Vanderwende, Pushmeet Kohli, and James Allen. A corpus and cloze evaluation for deeper understanding of commonsense stories. In Kevin Knight, Ani Nenkova, and Owen Rambow, editors, �� ...

Pith/arXiv arXiv doi:10.18653/v1/n16-1098 2025

[8] [8]

ISBN 979-8-89176-251-0

Association for Computational Linguistics. ISBN 979-8-89176-251-0. doi: 10.18653/v1/2025.acl-long.919. https://aclanthology.org/2025.acl-long.919/. Keshan Sodimana, Knot Pipatsrisawat, Linne Ha, Martin Jansche, Oddur Kjartansson, Pasindu De Silva, and Supheakmungkol Sarin. A Step-by-Step Process for Building TTS Voices Using Open Source Data and Frame- wo...

doi:10.18653/v1/2025.acl-long.919 2025

[9] [9]

63 Per Erik Solberg and Pablo Ortiz

https://arxiv.org/abs/2402.16829. 63 Per Erik Solberg and Pablo Ortiz. The Norwegian parliamentary speech corpus. In �� , 2022. Saba Sturua, Isabelle Mohr, Mohammad Kalim Akram, Michael Günther, Bo Wang, Markus Krimmel, ...

Pith/arXiv arXiv doi:10.1007/978-3-031-88720-8_21 2022

[10] [10]

XNLI 2.0: Improving XNLI dataset and performance on Cross Lingual Understanding (XLU)

Association for Computational Linguistics. doi: 10.18653/v1/2023.eacl-main.108. https://aclanthology. org/2023.eacl-main.108/. Jörg Tiedemann. The Tatoeba Translation Challenge – Realistic Data Sets for Low Resource and Multilingual MT. In Loïc Barrault, Ondřej Bojar, Fethi Bougares, Rajen Chatterjee, Marta R. Costa-jussà, Christian Federmann, Mark Fishel...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.18653/v1/2023.eacl-main.108 2023

[11] [11]

https://doi.org/10.24963/ijcai.2019/746

doi: 10.24963/ijcai.2019/746. https://doi.org/10.24963/ijcai.2019/746. Biao Zhang, Ankur Bapna, Rico Sennrich, and Orhan Firat. Share or Not? Learning to Schedule Language-Specific Capacity for Multilingual Translation. In �� , 2021. https: //openreview.net/fo...

Pith/arXiv arXiv doi:10.24963/ijcai.2019/746 2019

[12] [12]

enable_points

https://openreview.net/forum?id=vfzRRjumpX. Dong Zhang, Shimin Li, Xin Zhang, Jun Zhan, Pengyu Wang, Yaqian Zhou, and Xipeng Qiu. Speechgpt: Empow- ering large language models with intrinsic cross-modal conversational abilities. In �� ...

doi:10.1162/tacl_a_00595 2023

[13] [13]

The script checks whether 'ExitSignal' exists within the 'Actions' list of the dictionary linked to the 'session' key in the 'raw_action_logs' dictionary, and if so, assigns the result of a numpy operation (after squeezing) to the variable 'exit_signals_processed'

[14] [14]

The script verifies if the event_label is 'Position' and the 'Anchored' value in the entry dictionary is True, then updates self.disable_anchor to True

[15] [15]

The script checks if the event_category property of the trip_event object matches TripEventCategory.POSITION. JavaScript Example Source: const DataHandler = require('./lib/dataHandler.js'); let windowRef; const dataHandler = new DataHandler({}); Target: �� ...

[16] [16]

The script imports a function or variable called buildHandler from a file at '../lib/dataHandler'

[17] [17]

The script imports HandlerData from a file named dataHandler and uses it to instantiate a new Vuex Handler, assigning the result to a constant called handler

[18] [18]

The script imports the initializeHandler function from the 'redux' package and also imports the reducer from './handlerReducer' to set up a handler. Math Example Source: G� → X Target: �� G �� n �� ...

[19] [19]

The function G� maps every element of the set E to a corresponding element in the set S

[20] [20]

There exists a mapping or function from the set or space G to the set or space H

[21] [21]

Figure 14 Examples of hard negatives generation across Python code (blue), JavaScript code (orange), and mathe- matical expressions (green)

A function G, parameterized by λ, maps to or transforms into a space or set denoted as X, also parameterized by λ. Figure 14 Examples of hard negatives generation across Python code (blue), JavaScript code (orange), and mathe- matical expressions (green). 73 Hard Negatives Generation You are a text transformation specialist. Generate ONLY valid xsim++ tra...

[22] [22]

ENTITY REPLACEMENT: • Swap proper nouns (people, locations, organizations) • Replace pronouns ( ��→ �� → ��)

[23] [23]

Transformed sentence 1

NUMBER ALTERATION: • Change quantities ( ��→ ��) • Modify dates/times ( ��→ ��) • Alter percentages ( ��→ ��) Follow these patterns from training examples: {few-shot examples} Now transform THIS SPECIFIC INPUT SENTENCE using the above patterns. Output ONLY a Python list of 1-5 modified sentences in this exact format: [ "Transformed s...

[24] [24]

Create 1-5 unique modified sentences

[25] [25]

Maximize difference from original text

[26] [26]

Mix transformation types where possible

[27] [27]

Maintain grammatical correctness

[28] [28]

Do NOT generate paraphrases, or synonyms

[29] [29]

NEVER output empty strings

[30] [30]

Output ONLY a Python list of strings

[31] [31]

��

No explanations, headers, or additional text Input sentence to transform: {example} Figure 15 Prompt for generating xsim++ transformations with clear instructions and structure. 74 A.5 Language code correspondence �� ...

2024