PaperlayerView exact revision on arXiv

Source-linked AI summary

Hallucinations in Large Multilingual Translation Models

Nuno M. Guerreiro, Duarte Alves, Jonas Waldendorf, Barry Haddow, Alexandra Birch, Pierre Colombo, André F. T. Martins

arXiv:2303.16104v1cs.CL

TL;DR

Large-scale multilingual translation models can produce hallucinated translations that damage user trust and pose safety concerns, while prior research focused mainly on small bilingual, English-centric high-resource models. This work investigates hallucinations in M2M-100 models and ChatGPT across varied translation settings, finding resource- and direction-dependent hallucinations, toxicity, model-family effects, and differing fallback behavior.

  • Problem

    Prior hallucination research mainly studied small bilingual models trained on a single English-centric high-resource language pair, leaving a knowledge gap about large-scale multilingual models across translation directions, domains, and data conditions.

  • Method

    The study investigates hallucinations in the M2M-100 family, including its 12B-parameter model, and in ChatGPT, using translation benchmarks including FLORES-101, WMT, and TICO.

  • Results

    Hallucinations are concentrated in low-resource pairs and translation out of English, can become toxic, and differ between ChatGPT and conventional translation models; diverse external fallback models improve quality and eliminate oscillatory hallucinations.

  • Takeaways & Limitations

    Hallucination prevalence and properties vary across model scale, translation direction, and data conditions, while external fallback models offer a mitigation strategy within the studied settings.

  • Takeaways & Limitations

    The study mainly examines the M2M family, so it is unclear how its findings generalize to other multilingual model families such as NLLB.

Abstract

from arXiv · show

Large-scale multilingual machine translation systems have demonstrated remarkable ability to translate directly between numerous languages, making them increasingly appealing for real-world applications. However, when deployed in the wild, these models may generate hallucinated translations which have the potential to severely undermine user trust and raise safety concerns. Existing research on hallucinations has primarily focused on small bilingual models trained on high-resource languages, leaving a gap in our understanding of hallucinations in massively multilingual models across diverse translation scenarios. In this work, we fill this gap by conducting a comprehensive analysis on both the M2M family of conventional neural machine translation models and ChatGPT, a general-purpose large language model~(LLM) that can be prompted for translation. Our investigation covers a broad spectrum of conditions, spanning over 100 translation directions across various resource levels and going beyond English-centric language pairs. We provide key insights regarding the prevalence, properties, and mitigation of hallucinations, paving the way towards more responsible and reliable machine translation systems.

1 Introduction

Hallucinations threaten the safety and reliability of increasingly capable multilingual translation systems, yet prior work largely examined small, English-centric bilingual models. This study analyzes hallucinations across massively multilingual M2M models and ChatGPT, finding resource-, direction-, and model-specific patterns and mitigation challenges.

  • Motivation: Hallucination research has largely focused on small bilingual models trained on a single English-centric high-resource pair, leaving multilingual evidence limited.The paper addresses prevalence and properties across translation directions, domains, and data conditions.
  • Approach: The study investigates M2M-100 models, including a 12B-parameter open-source model, alongside ChatGPT for translation.These represent conventional massively multilingual NMT and prompted generative LLM approaches.
  • Findings: Low-resource hallucinations can contain toxic patterns traceable to training data, creating serious safety concerns.This extends the analysis beyond translation quality to the content and provenance of hallucinated outputs.
  • Findings: ChatGPT hallucinations differ qualitatively from conventional translation-model hallucinations, including off-target translations, overgeneration, and failed translation attempts.The finding concerns hallucination forms rather than overall translation quality.
  • Mitigation: Hallucinations are difficult to reverse with same-family models, whereas more diverse fallback models can improve quality and remove oscillatory pathologies.The paper also releases code and over one million translations spanning more than 100 directions.

2 Background

Massively multilingual translation systems use one model across many language pairs, while LLMs offer prompted translation capabilities. Machine-translation hallucinations are rare on clean inputs, can be induced by perturbations, and include outputs detached from the source.

  • Multilingual translation: Massively multilingual NMT translates directly among many languages and language pairs without relying on a pivot language.These systems are trained on large quantities of parallel data, often using data mining and backtranslation.
  • LLM translation: LLMs can produce fluent, adequate translations, especially for high-resource English-centric pairs, sometimes competing with dedicated supervised systems.Their general-purpose prompting enables translation alongside many other NLP tasks.
  • Hallucination background: Translation hallucinations are unusually rare and difficult to observe in clean, unperturbed data because the task is relatively closed-ended.This motivates constructing perturbed inputs to study their properties.
  • Hallucination types: A perturbation hallucination is a substantially lower-quality translation caused by a slightly altered source, exposing limited robustness to source perturbations.Examples include misspellings and capitalization errors.
  • Hallucination types: Natural hallucinations occur without perturbation and contain content detached from the source, including fluent detached or oscillatory outputs.Oscillatory hallucinations contain erroneous repetitions of words and phrases.

3 Experimental Suite

The study evaluates conventional supervised multilingual NMT models and prompted LLM translation across established multilingual benchmarks. It uses M2M-100 variants, SMaLL100, and ChatGPT with sentence-level metrics, while noting data-access and API limitations.

  • Models: The experiments compare conventional supervised multilingual NMT models with LLMs prompted for translation.The conventional models are from the M2M-100 family, alongside SMaLL100 and ChatGPT.
  • Models: M2M-100 variants range from 418M to 12B parameters and support 100 languages and thousands of translation directions.The models were trained on 7.5B web-crawled parallel sentences.
  • Models: ChatGPT is evaluated as gpt-3.5-turbo, a 175B-parameter GPT-family model fine-tuned with human feedback and prompted in zero-shot translation.The study used the API in March 2023.
  • Datasets: The evaluation uses FLORES-101, WMT, and TICO to cover broad multilingual and specialized medical-domain translation benchmarks.Dataset selection emphasized researcher familiarity and avoiding train/test overlap for M2M models.
  • Evaluation Metrics: The study focuses mainly on sentence-level evaluation and uses spBLEU alongside reference-based and reference-free neural metrics.spBLEU is selected for its use in massively multilingual translation and fairer low-resource evaluation than BLEU.
  • Evaluation Metrics: ChatGPT translation experiments encountered API or server errors especially for low-resource languages and lower-coverage scripts.Default parameters alleviated, but did not entirely eliminate, these errors.

4 Hallucinations under Perturbation

The study tests hallucinations induced by minimal source perturbations across 31 translation pairs, using threshold-based detection and multilingual comparisons. Hallucinations remain concentrated in lower-resource settings, vary across models, and differ qualitatively for ChatGPT.

  • Evaluation Setting: The experiments construct perturbed sources with misspellings, frequent-token insertions, and capitalization errors across 31 language pairs.The evaluation uses FLORES and focuses on translation out of English, including bridge and additional low-resource languages.
  • Detection: A model is counted as hallucinating when an originally adequate translation falls below a very low quality threshold after perturbation.The multi-model procedure requires original spBLEU > 9, selects the top 20% by average quality, and uses perturbed spBLEU < 3.
  • Results: Perturbations can trigger hallucinations even in larger models, with aggregated results reported in Table 1 and language-specific results in Figure 1.The detection procedure is designed to identify severe quality drops while excluding already low-quality unperturbed translations.
  • Resource Trends: Hallucination rates generally decrease as resource levels increase, with high-resource languages showing almost no hallucinations under perturbation.The results also vary across languages and among models trained within the same family.
  • Limitations: The perturbation detector does not explicitly target detachment from the source text, distinguishing these cases from natural hallucinations.Low-quality unperturbed translations are reserved for the subsequent natural-hallucination analysis.
  • Model Comparisons: SMaLL100 hallucinates less and for fewer low- and mid-resource languages than its larger M2M teacher model.The authors connect this pattern to SMaLL100’s uniform sampling across language pairs.
  • ChatGPT: ChatGPT produces more hallucinations for mid-resource than low-resource languages and shows off-target translations, overgeneration, and failed translation attempts.Unlike traditional NMT models, it produces no oscillatory character hallucinations under perturbation.
  • Reversal: The vast majority of hallucinations can be reversed through further sampling, suggesting that some arise from generation variability rather than inability to translate.This observation applies to both LLM and traditional NMT hallucination findings discussed by the authors.

5 Natural Hallucinations

Natural hallucinations are more prevalent in low-resource and out-of-English directions, with detached hallucinations especially prominent there. The study also finds toxicity, model-scaling effects, and distinct behavior on medical data.

  • Prevalence and types: More than 10% average hallucination rates occur for all M2M models in low-resource directions, where detached hallucinations are more prevalent than oscillatory ones.The findings suggest weaker reliance on source context in low-resource translation.
  • Model differences: SMaLL100 hallucinates less than larger M2M models in low-resource settings and consistently relies more on source text, although this does not necessarily imply higher translation quality.The authors associate this pattern with training and architectural differences.
  • Model scaling: Scaling up within the M2M family consistently reduces hallucination rates, with larger relative improvements in mid- and high-resource directions.M2M (L) hallucinates for fewer languages than the smaller models.
  • Translation direction: Hallucinations are more frequent when translating out of English, accompanied by lower ALTI+ source contributions than when translating into English.Translation direction also changes hallucination properties.
  • Toxicity: Toxic text primarily appears in hallucinations generated out of English and almost exclusively affects low-resource language pairs.Over 1 in 8 hallucinations in Tamil contain toxic text.
  • Medical data: Hallucination rates on TICO medical data do not consistently exceed those on FLORES Wikipedia data.The authors hypothesize that domain shift is less pronounced for large M2M models than for smaller specialized-domain systems.

6 Mitigation of Hallucinations through Fallback Systems

Fallback systems can mitigate hallucinations, but models sharing training data and architecture are limited at reversing them. External models, especially NLLB and ChatGPT, improve quality and largely eliminate oscillatory hallucinations.

  • Same-family fallbacks: Same-family fallback models reverse oscillatory hallucinations more often than detached hallucinations, showing that detached hallucinations are especially sticky.Shared training data and architecture make reversal more difficult.
  • Same-family fallbacks: Scaling up within the same model family is not an effective hallucination-mitigation strategy.SMaLL100 can be more complementary because its training data, procedure, and architecture differ from the M2M models.
  • External fallbacks: External fallback systems, particularly NLLB, significantly improve COMET-22 translation quality over same-family fallbacks, especially for low-resource languages.NLLB generally outperforms ChatGPT for low- and mid-resource directions, while ChatGPT still surpasses dedicated M2M systems there as a fallback.
  • External fallbacks: Oscillatory hallucinations are almost entirely eliminated when ChatGPT is used as the fallback system.ChatGPT produces very few, if any, oscillations and slightly improves the rates obtained with NLLB.

7 Conclusion

The paper broadens hallucination analysis beyond prior bilingual, high-resource settings by examining massively multilingual models across varied directions, scales, and data conditions. It finds that external fallback models improve translation quality and virtually eliminate oscillatory hallucinations.

  • Conclusion: The study examines hallucinations across massively multilingual models, varied scales, translation directions, resource levels, and data conditions.It covers scenarios previously overlooked by work focused on bilingual high-resource models.
  • Conclusion: External models can significantly improve translation performance and virtually eliminate pathologies such as oscillatory hallucinations.The study also releases code and over a million translations and detection results.

Limitations

The study's conclusions are constrained by its model coverage, detection metrics, and limited access to ChatGPT's training and internal information.

  • Scope: The study mainly focuses on the M2M family, so it is unclear how its findings generalize to other multilingual model families such as NLLB.
  • Detection: The detection approaches inherit limitations from their underlying metrics, including the use of BLEU for perturbation hallucinations despite weaker WMT22 ranking than reference-based neural metrics.
  • ChatGPT evaluation: Because ChatGPT is behind API walls and sparsely documented, the study could not ensure it had not seen evaluation sets or evaluate source-text contributions for detached-hallucination detection.

Supplemental Material

The supplemental material documents the languages, benchmarks, perturbations, and dataset configurations used to evaluate multilingual translation hallucinations.

  • Datasets: FLORES contains 101 languages and 2009 parallel sentences per translation direction after joining its dev and devtest splits.The benchmark consists of sentences extracted from Wikipedia and translated from original English sentences.
  • Datasets: TICO is a medical-domain multilingual dataset with 2100 parallel sentences per translation direction after joining its dev and test sets.Its English open-source sources include scientific articles and government health announcements.
  • Datasets: WMT evaluation uses the English-centric benchmarks from the original M2M study plus WMT21 and WMT22 test sets released after model training.The benchmark specifications are summarized in Table 7.
  • Perturbations: The perturbation procedure randomly misspells words, inserts frequent tokens or punctuation, and title-cases words.Misspellings change characters with probability 0.01, while title-casing occurs with probability 0.1 and is guaranteed to affect at least one word.
  • Languages: The perturbation analysis pairs English with bridge and additional low-resource languages, excluding Tamil because too few suitable candidates remained after detection filtering.The analyzed language list includes 31 English-paired languages.

C.2 Supplementary Results

Supplementary results examine perturbation robustness, language coverage, translation quality, attribution scores, and hallucination rates across FLORES English-centric directions.

  • Perturbation results: The Pearson correlation between original translation quality and detected hallucinations under perturbation is very weak across all models.The corresponding model-based statistics are reported in Table 8.
  • Perturbation results: Perturbation hallucinations expose a lack of robustness because source perturbations can produce significant negative shifts in translation quality.They do not necessarily entail detachment from the source, unlike natural hallucinations.
  • Experimental setups: The English-centric setup evaluates language pairs formed by English and bridge or additional low-resource languages, while the non-English-centric setup covers 25 specified directions.The specialized-domain setup pairs English with nine TICO-supported languages.
  • Attribution and quality: For M2M models, ALTI+ shows smaller source contributions when translating out of English and for low-resource pairs.SMaLL100 has significantly higher and more resource-level-consistent ALTI+ scores than the M2M models.
  • Attribution and quality: SMaLL100’s lower low-resource hallucination rates do not necessarily imply better translation quality than other M2M models.M2M COMET-22 scores strongly correlate with low-resource hallucination rates, whereas the relationship differs for SMaLL100.

D.2.2 Evaluation with WMT.

WMT and supplementary multilingual evaluations report hallucination prevalence, oscillatory-hallucination prevalence, and translation quality across English-centric, non-English-centric, and specialized-domain directions.

  • WMT evaluation: The WMT analysis reports translation quality with spBLEU and COMET-22 and aggregates hallucination results by language pair and resource level.Table 12 includes the fraction of language pairs with at least one hallucination and average hallucination rates, with medians where available.
  • WMT evaluation: High-resource WMT language pairs are omitted because they served as validation data for setting model-based ALTI+ thresholds.This omission limits direct comparison with the included resource levels.
  • WMT evaluation: The reported hallucination-prevalence trends for WMT language pairs hold for the broader analysis presented in the main text.The WMT heatmap covers model-specific hallucination rates across translation directions.
  • Cross-setup analyses: Heatmaps separately summarize model-specific hallucination rates and the relative prevalence of oscillatory hallucinations detected only by TNG.These analyses cover WMT, non-English-centric, and TICO specialized-domain setups.
  • Cross-setup analyses: The non-English-centric and TICO analyses report translation quality with spBLEU and COMET-22 alongside language-pair hallucination heatmaps.The corresponding tables and figures provide setup-specific evaluation summaries.

E Fallback System Analysis

The fallback-system analysis compares COMET-22 quality on hallucinated translations across resource levels and examines external-model fallback behavior.

  • Fallback evaluation: The non-English-centric fallback results largely follow the trends analyzed in the main text.Figure 16 reports COMET-22 scores for original hallucinated translations by direction.
  • Model comparison: ChatGPT does not consistently outperform all other models in non-English-centric directions, especially at low and mid-resource levels.The paper links this pattern to ChatGPT’s non-parallel, heavily English-centric training corpus.
  • Fallback evaluation: Fallback analysis measures translation-quality improvements over original hallucinated translations with COMET-22 across resource levels.Original model hallucinated translations are represented with dashed lines.
Loading 2303.16104v1…