Hallucinations in Large Multilingual Translation Models

Abstract Hallucinated translations can severely undermine and raise safety issues when machine translation systems are deployed in the wild. Previous research on the topic focused on small bilingual models trained on high-resource languages, leaving a gap in our understanding of hallucinations in multilingual models across diverse translation scenarios. In this work, we fill this gap by conducting a comprehensive analysis—over 100 language pairs across various resource levels and going beyond English-centric directions—on both the M2M neural machine translation (NMT) models and GPT large language models (LLMs). Among several insights, we highlight that models struggle with hallucinations primarily in low-resource directions and when translating out of English, where, critically, they may reveal toxic patterns that can be traced back to the training data. We also find that LLMs produce qualitatively different hallucinations to those of NMT models. Finally, we show that hallucinations are hard to reverse by merely scaling models trained with the same data. However, employing more diverse models, trained on different data or with different procedures, as fallback systems can improve translation quality and virtually eliminate certain pathologies.


Introduction
Recent advancements in large-scale multilingual machine translation have brought us closer to realizing a universal translation system, capable of handling numerous languages and translation directions (Aharoni et al., 2019;Arivazhagan et al., 2019;Fan et al., 2020;Zhang et al., 2020;Wenzek et al., 2021;Goyal et al., 2022;NLLB Team et al., 2022).Concurrently, large language models (LLMs) have shown remarkable generalization to new tasks, including translation, where they are becoming increasingly stronger (Brown et al., 2020;Chowdhery et al., 2022;Hendy et al., 2023).Compared to traditional bilingual models, these systems can offer significant performance improvements and greatly simplify engineering efforts, as a single model can be used for all language pairs (Arivazhagan et al., 2019).As a result, they are an increasingly attractive choice for real-world applications.However, when deployed in the wild, these models may generate hallucinations: highly pathological translations detached from the source that can severely damage user trust and pose safety concerns (Perez et al., 2022).
The problem of hallucinations has long been recognized by researchers (Ji et al., 2022), and recent studies have contributed towards better understanding, detection and mitigation of these pathological translations.However, these studies have been conducted on small bilingual models (<100M parameters) trained on a single Englishcentric high-resource language pair (Raunak et al., 2021;Guerreiro et al., 2023b,a;Dale et al., 2023;Xu et al., 2023).This leaves a knowledge gap regarding the prevalence and properties of hallucinations in large-scale translation models across different translation directions, domains and data conditions.
In this work, we aim to fill this gap by investigating hallucinations on two different classes of models.The first and main class in our analysis is the de facto standard approach of massively multilingual supervised models: We use the M2M-100 family of multilingual NMT models (Fan et al., 2020), which includes the largest opensource multilingual NMT model with 12B parameters.The second class is the novel and promising approach of leveraging generative LLMs for translation.Contrary to conventional NMT models, these models are trained on massive amounts of 1500 monolingual data in many languages, with a strong bias towards English, and do not require parallel data.In our analysis, we use ChatGPT and GPT-4, as thay have been shown to achieve high translation quality over a wide range of language pairs (Hendy et al., 2023;Peng et al., 2023).
We organize our study by analyzing the two prevalent types of hallucinations in NMT considered in the literature: hallucinations under perturbation and natural hallucinations (Lee et al., 2018;Raunak et al., 2021;Guerreiro et al., 2023b).Firstly, we study hallucinations under perturbation and evaluate whether these translation systems are robust to source-side artificial perturbations.While previous studies have found that these perturbations (e.g., spelling errors and capitalization mistakes) can reliably induce hallucinations (Lee et al., 2018;Raunak et al., 2021), it is not clear whether those conclusions hold for large multilingual models.Secondly, we comprehensively investigate natural hallucinations, and evaluate their properties in the outputs of the massively multilingual M2M models on a vast range of conditions, spanning from English-centric to non-English-centric language pairs, translation directions with little supervision, and specialized medical domain data where hallucinations can have devastating impact.Finally, we study a hybrid setup where other models can be requested as fallback systems when an original system hallucinates, with the aim of mitigating hallucinations and improving translation quality on-the-fly.
We provide several key insights on properties of hallucinations, including: • models predominantly struggle with hallucinations in low-resource language pairs and translating out of English; critically, these hallucinations may contain toxic patterns that can be traced back to the training data; • smaller distilled models can, surprisingly, hallucinate less than large-scale models; we hypothesize that this is due to modeling choices that discourage hallucinations (e.g., leveraging less potent shallow decoders that rely more on the encoder representations); • LLMs produce hallucinations that are qualitatively different from those of conventional translation models, mostly consisting of offtarget translations, overgeneration, and even failed attempts to translate; • hallucinations are sticky and hard to reverse with models that share the same training data, whereas employing more diverse fallback systems can substantially improve overall translation quality and eliminate pathologies such as oscillatory hallucinations.
We release all our code and make available over a million translations in more than 100 translation directions to spur future research.1 2 Background

Large Multilingual Language Models
Massively multilingual neural machine translation has emerged as a powerful paradigm for building machine translation systems that can handle numerous languages (Akhbardeh et al., 2021;Wenzek et al., 2021;NLLB Team et al., 2022;Bapna et al., 2022;Chowdhery et al., 2022).These systems translate directly in multiple translation directions without relying on a pivot language.
The dominant strategy for achieving these systems is training large multilingual models on vast amounts of parallel data often obtained through a combination of data mining and data augmentation strategies, such as backtranslation (Sennrich et al., 2016;Edunov et al., 2018).Compared to traditional bilingual models, these systems bring significant improvements, particularly for lowresource and non-English-centric language pairs, as these benefit the most from multilingual transfer (Arivazhagan et al., 2019;Fan et al., 2020).
As an alternative, a novel strategy is to leverage LLMs.These systems are pretrained on massive nonparallel corpora and can be prompted to solve arbitrary tasks (Radford et al., 2019;Brown et al., 2020).In fact, this approach has led to impressive results across a wide variety of NLP tasks (Chowdhery et al., 2022;Zhang et al., 2022).Translation is no exception: LLMs can produce translations that are competitive with those of supervised translation models (Vilar et al., 2023;Peng et al., 2023;Hendy et al., 2023;Bawden and Yvon, 2023).

Hallucinations in Machine Translation
Hallucinations lie at the extreme end of translation pathologies and present a critical challenge in machine translation, as they can compromise the safety and reliability of real-world applications.
Importantly, hallucinations in machine translation are unlike hallucinations in other natural language generation tasks (e.g., abstractive summarization) (Ji et al., 2022).Whereas, for these other tasks, models often produce hallucinated outputs (Falke et al., 2019;Cao et al., 2022;Manakul et al., 2023), hallucinations in machine translation, possibly due to the more closed-ended nature of the task, are substantially rarer and hard to observe in clean, unperturbed data.This has led several previous studies to examine their properties by creating artificial scenarios where hallucinations are more likely to occur (e.g., introducing perturbations in the source text (Lee et al., 2018) or noise in the training data (Raunak et al., 2021)).To distinguish these two scenarios, hallucinations in machine translation are categorized into two types (Raunak et al., 2021): hallucinations under perturbation and natural hallucinations.
Hallucinations under Perturbation.A model generates a hallucination under perturbation when it produces a significantly lower quality translation for a slightly perturbed input compared to the original input (Lee et al., 2018).Hallucinations under perturbation explicitly reveal the lack of robustness of translation systems to perturbations in the source by finding translations that undergo significant negative shifts in quality due to these changes.
Natural Hallucinations.These translations occur naturally, without any perturbation.In this work, we follow the taxonomy introduced in Raunak et al. (2021) and extended in Guerreiro et al. (2023b).Under this taxonomy, hallucinations are translations that contain content that is detached from the source, and are further categorized as fluent detached hallucinations or oscillatory hallucinations.The former refers to translations that bear minimal or no relation at all to the source, while the latter refers to inadequate translations that contain erroneous repetitions of words and phrases.
While recent research has contributed towards understanding, detection, and mitigation of hallucinations (Guerreiro et al., 2023a,b;Dale et al., 2023;Xu et al., 2023), the scope of these studies has been restricted to small models (<100M parameters) trained on a single, high-resource language pair.In this work, we expand upon this prior research by studying hallucinations in large multilingual models across various translation directions, domains, and data conditions, thereby addressing an important gap in the literature.
3 Experimental Suite

Models
For supervised multilingual NMT models, we use the transformer-based (Vaswani et al., 2017) M2M-100 family of models (Fan et al., 2020): M2M (S) with 418M parameters, M2M (M) with 1.2B parameters, and M2M (L) with 12B parameters.These models were trained on a manyto-many parallel dataset of 7.5B sentences, and support 100 languages and thousands of language pairs (LPs).We also evaluate SMaLL100 (Mohammadshahi et al., 2022), a model with 330M parameters obtained via distillation of M2M (L).SMaLL100 was trained on a smaller training set obtained via uniform sampling across all language pairs to reduce the bias towards high-resource languages: only 100k parallel sentences from the M2M training data were used for each LP, for a total of 456M parallel sentences.For decoding, we run beam search with a beam size of 4.
As for the alternative strategy using LLMs, we use two GPT models: ChatGPT (gpt-3.5turbo)and GPT-4.2These models are upgraded versions of GPT-3.5-a175B GPT (Radford and Narasimhan, 2018;Radford et al., 2019;Brown et al., 2020) LLM-that has been fine-tuned with human feedback in the style of InstructGPT (Ouyang et al., 2022).In particular, ChatGPT has been shown to achieve impressive results for multiple multilingual tasks, including translation (Kocmi and Federmann, 2023;Fu et al., 2023;Hendy et al., 2023).Crucially, GPT models, unlike the majority of existing LLMs, exhibit extensive and strong capabilities across multiple language pairs, a prerequisite for carrying out the research in this work.To generate translations, we use the zero-shot prompt template used in Hendy et al. (2023) and the default API parameters.

Datasets
We chose datasets familiar to researchers and practitioners, ensuring no train/test overlap for the M2M models.To this end, we selected premier translation benchmarks: FLORES-101 (Goyal et al., 2022), TICO (Anastasopoulos et al., 2020), and WMT.FLORES-101 is a multi-parallel dataset that consists of Wikipedia text in 101 languages; we join the dev and devtest sets.TICO is a specialized medical-domain multilingual benchmark with COVID-19 related data, such as medical papers and news articles; we join the dev and test sets.Additionally, we use (i) WMT benchmarks from the M2M paper evaluation suite, as they were removed from the training data of M2M models;3 and (ii) WMT 2022 benchmarks (Kocmi et al., 2022), as they were created after the cutoff date of training data of all tested models.

Evaluation Metrics
Our main lexical metric is spBLEU (Goyal et al., 2022), as it has been widely employed in works on massively multilingual translation (Fan et al., 2020;Wenzek et al., 2021;NLLB Team et al., 2022). 4Moreover, we also adopt neural referencebased and reference-free COMET variants: COMET-22 and CometKiwi (Rei et al., 2022a,b).Lastly, we use the cross-lingual encoder LaBSE (Feng et al., 2022) to obtain sentence similarity scores, as these have been successfully used in prior research on detection of hallucinations (Guerreiro et al., 2023a;Dale et al., 2023).

Hallucinations under Perturbation
We start our analysis by focusing on artificially created hallucinations.We first provide an in-depth overview of our evaluation setting, focusing on the construction of the perturbed data and detection approach.Then, we present our results and analyze the properties of these hallucinations.

Evaluation Setting
Translation Directions.We use the FLORES dataset for these experiments, and focus specifically on translation out of English. 5We select all M2M bridge languages, as well as additional low-resource languages that were underrepresented among bridge languages. 6Overall, we generate translations for 31 different LP. 7 Additionally, we report results for the WMT 2022 benchmarks in Appendix A.3 to ensure evaluation on data that was released after the cutoff date of the training data of GPT models.8 Perturbations.We employ the same minimal perturbations used in Xu et al. (2023) to construct perturbed source sequences: misspelling errors, insertion of frequent tokens in the beginning of the sequence, and capitalization errors.Previous work has shown that these perturbations can trigger severe output errors (Lee et al., 2018;Karpukhin et al., 2019;Berard et al., 2019;Raunak et al., 2021).Additionally, we also experiment with the cascade approach in speech translation (Bentivogli et al., 2021), providing a real-world scenario where translation models have to deal with noisy inputs.In this approach, the audio (for consistency, we use the FLEURS dataset [Conneau et al., 2022], which consists of audio recordings of the sentences in the FLORES dataset) is first transcribed by an automatic speech recognition (ASR) system (we use a Whisper model [Radford et al., 2022] 9 ) and then translated by a translation system.Detection.Our detection approach is inspired by previous work on hallucinations under perturbation (Lee et al., 2018;Raunak et al., 2021;Ferrando et al., 2022;Xu et al., 2023).The algorithm is a simple 2-rule process: we fix (i) a minimum threshold quality score for the original translations, and (ii) an extremely low maximum quality score for the perturbed translations.A model generates a hallucination under perturbation when both translations meet the thresholds.
Table 1: Fraction of languages for which models produces at least one hallucination under perturbation, and average hallucination rate (and median, in subscript) among candidate translations across all languages at each resource level.† GPT models may have been exposed to the test samples.Crucially, rule (i) ensures that low-quality translations for unperturbed sources are not considered as candidates for hallucinations under perturbation.
We adapt rule (i) to ensure consistency across multiple models and LPs.We first obtain source sentences for which all models produce translations that meet a minimum quality threshold (spBLEU > 9), sort them by average quality across models, and select the top 20% as candidates.Finally, we apply rule (ii) and set the threshold to spBLEU < 3. 10 We selected these thresholds based on those used in previous works (Raunak et al., 2021;Ferrando et al., 2022;Xu et al., 2023).In Appendix A.1, we validate our detection method with human annotation on over 200 translations for 10 different language pairs.11

Results
Overall, we find that perturbations have the potential to trigger hallucinations, even in larger models.In what follows, we present several key insights.
Average hallucination rates for NMT models generally decrease with increasing resource levels.Table 1 shows that all NMT models exhibit lower hallucination rates as resource levels increase.This is expected and suggests that these models are better equipped to handle source-side perturbations for language pairs with more parallel data during training.In fact, hallucinations under perturbation for high-resource languages are almost non-existent.However, Figure 1 reveals variability across languages, and even within the models in the same family that share the same training data.For instance, when translating to Asturian (ast), M2M (L) and the distilled SMaLL100 show significantly higher hallucination rates than M2M (S).This suggests that hallucinations emerge in non-trivial ways unrelated to the training data.SMaLL100 exhibits lower hallucination rates than its teacher model M2M (L).Recall that SMaLL100 was trained using uniform sampling across all language pairs to reduce bias towards high-resource languages.The results in Table 1 may reflect a positive outcome from this approach: despite being much smaller than M2M (L), SMaLL100 hallucinates less and for fewer directions on low-and mid-resource language pairs.Hallucinations under perturbation are not correlated with the quality of original translations.The approach for detection of hallucinations under perturbation raises an interesting question: Are the original source sentences for which models produce higher quality translations less likely to lead to hallucinations when perturbed?Our analysis found a weak Pearson correlation12 between hallucinations under perturbation and spBLEU scores for the original unperturbed sources across all models.This indicates that the perturbations that we introduce in the source can cause models to undergo significant shifts in translation quality.
LLMs exhibit different hallucination patterns from conventional NMT models.Contrary to NMT models, the GPT models generate more hallucinations for mid-resource languages than for low-resource languages (Table 1).When compared to all other models, the results show that ChatGPT produces more hallucinations for midresource languages, while GPT-4 exhibits fewer hallucinations in low-resource directions and across fewer languages.Additionally, hallucinations from LLMs are qualitatively different from those of other models: They often consist of offtarget translations, overgeneration, or even failed attempts to translate (e.g., ''This is an English sentence, so there is no way to translate it to Vietnamese'').This further demonstrates that translation errors, even critical ones, obtained with LLMs are different from those produced by NMT models (Vilar et al., 2023;Hendy et al., 2023).
Interestingly, we also found that almost all hallucinations can be reversed with additional sampling from the model.This aligns with findings in Guerreiro et al. (2023b) and Manakul et al. (2023): As with traditional NMT models, LLM hallucinations may not necessarily imply model defect or inability to generate adequate translations, but could just stem from ''bad luck'' during generation.

Natural Hallucinations
We now turn to investigating natural hallucinations. 13We first provide an overview of our evaluation setting, focusing on the scenarios and detection approach.Then, we present a thorough analysis exploring various properties of hallucinations.

Evaluation Setting
Evaluation Scenarios.Analyzing multilingual translation models opens up several research scenarios that have not been explored in previous studies conducted on bilingual models.We will investigate three different multilingual scenarios, comprising over 100 translation directions.
We begin with an English-centric scenario, pairing 32 languages with English for a total of 64 translation directions. 14Then, we study a non-English-centric scenario inspired by Fan et al. (2020), exploring 25 language pairs corresponding to real-world use cases of translation not involving English. 15Finally, we examine hallucinations in medical data, where they can have severely compromising effects.We pair 9 languages 16 with English for a total of 18 directions.We report results for the first two setups using the FLORES dataset.For the final setup, we use the TICO dataset.
ALTI+ assesses the relative contributions of source and target prefixes to model predictions.As hallucinations are translations detached from 13 From now on, we use the terms natural hallucinations (both detached and oscillatory hallucinations) and hallucinations interchangeably.
14 ar ast az bn cs cy de el es fa fi fr he hi hr hu id ja ko lt nl oc pl ps pt ru sv sw ta tr vi zh.
15 hi-bn it-fr de-hu it-de cs-sk nl-fr fr-sw ro-ru ro-uk de-hr hr-sr be-ru hr-hu hr-cs el-tr hr-sk nl-de af-zu ro-hu hi-mr ro-tr uk-ru ro-hy ar-fr ro-de.
16 ar fr hi id ps pt ru sw zh.
Table 2: Fraction of LPs on the English-centric setup for which models produce at least one hallucination, and average hallucination rate (and median, in subscript) across all LPs at each resource level.
the source sequence, ALTI+ can be leveraged to detect them by identifying sentences with minimal source contribution.Notably, it faithfully reflects model behavior and explicitly signals model detachment from the source in any translation direction (Ferrando et al., 2022).This method has been successfully employed to detect hallucinated toxicity in a multilingual context in NLLB Team et al. ( 2022), and validated on human-annotated hallucinations in Dale et al. (2023), where it was shown that ALTI+ scores easily separate detached hallucinations from other translations.17TNG, on the other hand, is a simple, lightweight black-box heuristic targeting oscillatory hallucinations.It compares the count of the top repeated translation n-gram to the count of the top repeated source n-gram, ensuring a minimum difference of t.This method has been validated on human-annotated hallucinations and found to identify oscillatory hallucinations with perfect precision (Guerreiro et al., 2023b).Following previous work, we use n = 4 and t = 2 (Raunak et al., 2021;Guerreiro et al., 2023b) and exclude translations that meet the minimum quality threshold from Section 4.1. 18emark on Model Selection.We use ALTI+, a model-based detector, for reliable detection of detached hallucinations.Since we lack access to internal features from the GPT models, we exclude them from our model selection to ensure consistency in our analysis.Importantly, using alternative detectors could lead to misleading results and create discrepancies in our evaluation setup.

English-Centric Translation
We start by studying hallucinations on Englishcentric language pairs.We reveal key insights on how properties of hallucinations change across resource levels, models and translation directions.
Hallucinations in low-resource language pairs are not only more frequent, but also distinct.Table 2 shows that hallucinations occur frequently for low-resource directions, with all M2M models exhibiting average hallucination rates over 10%.Moreover, all models generate hallucinations for almost all low-resource language pairs.Regarding the type of hallucinations, Figure 2 shows that in low-resource directions, in contrast to mid-and high-resource ones, oscillatory hallucinations are less prevalent, while detached hallucinations occur more frequently.These findings suggest that, although massive multilingual models have significantly improved translation quality for low-resource languages, there is considerable room for improvement, and also highlight potential safety issues arising from translations in these directions.SMaLL100 consistently relies more on the source than other models.Despite being the smallest model, SMaLL100 shows remarkable hallucination rates across low-and mid-resource directions, hallucinating significantly less than its larger counterparts in low-resource ones.We hypothesize that these improved rates may be attributed not only to the uniform sampling of language  pairs during training, but also to its architecture.While SMaLL100 shares a 12-layer encoder with the other models to obtain source representations, it diverges by employing a shallow 3-layer decoder-instead of a 12-layer decoder-and placing the target language code on the encoder side.This design may encourage greater reliance on the more complex encoder representations, reducing detachment from the source.In fact, distinct patterns in ALTI+ scores support this hypothesis: SMaLL100 has higher scores and similar patterns across all resource levels (see Figure 3).In contrast, the M2M models tend to rely less on the source, especially in low-resource en-xx LPs.Importantly, however, SMaLL100's lower hallucination rates do not necessarily imply superior translation quality compared to the M2M models: We found a strong correlation between M2M models' COMET-22 scores and their respective hallucination rates for low-resource LPs, whereas, contrastingly, the correlation is weak for SMaLL100. 19This suggests that, despite relying more on the source, SMaLL100's translations may not necessarily be of higher quality than those of the M2M models.
Scaling up models within the same family reduces hallucination rates.Table 2 shows that increasing the size of the M2M models results in consistent reductions in hallucination rates.Relative improvements are more pronounced for mid-and high-resource language pairs, with M2M  (L) exhibiting fewer hallucinations and hallucinating for fewer languages than all other models.
Hallucinations are more frequent when translating out of English.Table 3 demonstrates that models are consistently more prone to hallucinate when translating out of English.In fact, models tend to detach more from the source text in these directions.This is evidenced by ALTI+ source contributions (see Figure 3) being lower across all en-xx language pairs compared to translating into English, which aligns with observations in Ferrando et al. (2022).Interestingly, we also discovered that the translation direction can influence the properties of hallucinations: (i) over 90% of off-target hallucinations occur when translating out of English, and (ii) nearly all hallucinations into English for mid-and high-resource language pairs are oscillatory (see Figure 2).
Toxic hallucinations can be traced back to the training data.To assess the prevalence of toxic text in detected hallucinations, we utilized the toxicity wordlists provided by NLLB Team et al.
(2022).We found that toxic text predominantly appears in translations out of English and almost exclusively for low-resource directions.For instance, over 1 in 8 hallucinations in Tamil contain toxic text.Interestingly, these hallucinations exhibit high lexical overlap among them and are repeated across models for multiple unique source sentences.Moreover, they are not necessarily reduced by scaling up the model size.These observations suggest that these hallucinations are likely to be traced back to the training data, aligning with observations in Raunak et al. ( 2021) and Guerreiro et al. (2023b).In fact, upon inspecting Table 4: Fraction of LPs on the non-English-centric setup for which models produce at least one hallucination, and average hallucination rate (and median, in subscript) across all LPs at each resource level.
the corpora that were used to create the training data, we found reference translations that exactly match the toxic hallucinations.Additionally, we found that these hallucinations can propagate through model distillation, as evidenced by SMaLL100 generating copies of its teacher model's toxic hallucinations.This highlights the necessity of rigorously filtering training data to ensure safe use of these models.

Beyond English-Centric Translation
We shift our focus to translation directions that do not involve English, typically corresponding to directions with less supervision during training.
Trends are largely similar to English-centric directions.Table 4 reveals trends that largely mirror those observed in the English-centric setup: (i) hallucinations are more frequent in lowresource directions; 20 (ii) SMaLL100 significantly outperforms the M2M models in low-resource language pairs; and (iii) scaling up to M2M (L) consistently yields substantial improvements over the smaller M2M models in low-and mid-resource directions.Additionally, the trends related to hallucination types are also similar: Detached hallucinations are more prevalent in low-resource directions, while oscillatory hallucinations overwhelmingly dominate in mid-and high-resource ones.
Less supervised language pairs exhibit extremely high hallucination rates.As expected, models struggle more with hallucinations for directions with less or even no supervision during training, such as ro-hy and af-zu.For instance, M2M (M) hallucinates for nearly half of the translations in these directions. 20We considered the resource level of the language pair to be the smallest resource level between the two languages.Positive values indicate higher rates for TICO.

Translation on Specialized Domains
We now study hallucinations in medical data using the TICO dataset.We compare hallucination rates with the FLORES dataset in 18 directions.
Hallucinations are not exacerbated under medical domain data.Table 5 reveals that hallucination rates for the TICO data are not consistently higher than those observed for the FLORES data.This finding diverges from previous work that investigated hallucinations in specialized domain data (Wang and Sennrich, 2020;Müller et al., 2020).We hypothesize that, unlike the smaller models typically trained on limited datasets from a single domain used in those works, the concept of ''domain shift'' is not as pronounced for M2M models.These models are not only much larger but, crucially, they are trained on a massive dataset containing over 7B sentences from various domains.This vast training set potentially mitigates the impact of domain shift and, consequently, reduces its influence on hallucinations.

Mitigation of Hallucinations through Fallback Systems
We now explore the potential of mitigating hallucinations and improving overall translation quality by employing a simple hybrid setup that can take advantage of multiple systems with possible complementary strengths.Put simply, we leverage an alternative system as a fallback when the original model produces hallucinations. 21Our analysis is focused on the broader English-centric evaluation setup studied in Section 5.2.

Employing Models of the Same Family as Fallback Systems
We begin by analyzing the performance of same-family models when employed as fallback systems for one another (e.g., using SMaLL100, M2M (M), and M2M (L) as fallbacks for M2M (S)). 22 Detached hallucinations are particularly sticky across M2M models.Figure 4 reveals that when employing M2M models as fallback systems, reversal rates-percentage of hallucinations from the original system that are reversed by the fallback system-are consistently higher for oscillatory hallucinations than for detached hallucinations.These findings not only align with those in Guerreiro et al. (2023b), where oscillatory hallucinations were found to be less related to model defects, but also further highlight the close connection between detached hallucinations and the training data.This connection can help explain their stickiness: Since the M2M models share the same training data, reversing these hallucinations using other M2M variants as fallbacks is more challenging.Interestingly, we also observe that M2M (L) particularly struggles to reverse the detached hallucinations generated by its 21 For consistency, we use the same detection approach described in Section 5.1. 22For simplicity, we consider the distilled SMaLL100 as a model from the M2M family.distilled counterpart SMaLL100, suggesting that model defects can persist and be shared during distillation.
Scaling up within the model family is not an effective strategy for mitigating hallucinations.In line with our findings in Section 5.2, Figure 4 shows that reversal rates using SMaLL100 as a fallback system are higher for detached hallucinations than for oscillatory hallucinations.Although SMaLL100 is a distilled M2M model, its training data, training procedure, and architecture differ from those of the M2M models.This distinction may make it a more complementary fallback system to other M2M models than simply scaling up within the same model family.This suggests that merely increasing the scale of models within the same family is not an effective strategy for mitigating hallucinations, and exploring alternative models with different training procedure and data may yield further improvements.We will analyze this alternative strategy in the next section.

Employing Diverse Fallback Systems
Building on the findings from the previous section, we will investigate how models outside the M2M family can be employed to mitigate hallucinations and improve translation quality.We will test this approach with two different models: (i) we will prompt the GPT models as detailed in Section 3,23 and (ii) we will use a high-quality 3.3B parameter model from the NLLB family of NMT models (NLLB) (NLLB Team et al., 2022).
Diverse fallback systems can significantly improve translation quality.Figure 5a shows that diverse fallback systems can significantly enhance translation quality of originally hallucinated translations compared to same-family models.This improvement is most pronounced for low-resource directions, where both the LLMs and NLLB consistently boost translation quality.Moreover, NLLB generally outperforms the GPT models for low-and mid-resource language pairs, aligning with previous work that found that these models lag behind supervised models in these directions (Hendy et al., 2023).Nonetheless, even for these language pairs, GPT models exhibit superior performance to the dedicated M2M translation systems, further underscoring the limitations of relying on same-family models as fallbacks.
Oscillatory hallucinations are practically nonexistent when employing diverse fallbacks.Figure 5b demonstrates another benefit of using diverse fallback systems: oscillatory hallucinations are almost completely eliminated.Consistent with our findings in Section 4, we observe that LLMs produce very few, if any, oscillations, slightly improving the rates obtained with NLLB.This provides further evidence that hallucinations obtained with LLMs exhibit different properties and surface forms.Investigating these differences is a relevant research direction for future work.

Conclusion
We have comprehensively investigated hallucinations in massively multilingual translation models, exploring a wide range of translation scenarios that remained overlooked in previous work.
Our analysis provided several key insights on the prevalence and properties of hallucinations across models of different scales and architectures, translation directions, and data conditions, including: the prevalence of hallucinations in lowresource languages and when translating out of English; the emergence of toxicity in hallucinations, which can be directly traced back to the training data; how model distillation may bring reduced hallucination rates compared to larger models; how scaling up within the same model family generally decreases the rate of hallucinations; and how LLMs produce qualitatively different hallucinations compared to conventional NMT models.Finally, we also examined how fallback systems can be employed to mitigate hallucinations.We found that hallucinations can be sticky and difficult to reverse when using models of the same family.However, diverse fallbacks with different training procedure and data can significantly improve translation quality and virtually eliminate pathologies such as oscillatory hallucinations.
Table 7: Counts and rate, in percentage, of extractable samples from the FLORES dataset.
Table 8: Average hallucination rate, in percentage, for each language pair on the WMT 2022 benchmarks.the most, one additional sentence for some target languages.
Notably, even if minor train-test overlap exists, their influence on BLEU scoring differences on translations obtained with LLMs has been found to be small, or almost negligible, when translating out of English (Vilar et al., 2023;Chowdhery et al., 2022).Alongside the results in Table 7, we are confident that our analysis with the GPT models in Section 4 is valuable and valid, irrespective of any hypothetical, albeit small, traintest overlap.Nevertheless, in the next section, we experiment on benchmarks created after the training data cutoff of these LLMs. 28Crucially, the trends in the results closely follow those reported in Section 4.

A.3 Hallucinations under Perturbation on WMT 2022 Benchmarks
We now investigate hallucinations under perturbation for WMT 2022 benchmarks (Kocmi et al., 2022) created after the cutoff data of the training data of GPT models.Additionally, to include 28 This chronology suggests that the train-test overlap risk is reduced.However, as the training data is not available, we cannot entirely rule out overlap with the test data.more languages to our analysis, we also experiment with WMT 2021 benchmarks (Akhbardeh et al., 2021) (for Hausa, ha, and Icelandic, is), as they are less likely to overlap with the training data of GPT models.We apply the same perturbations and detection method detailed in Section 4. 1. 29   Trends with GPT models largely mirror those reported on the FLORES dataset.Table 8 reveals trends that align closely with those from Section 4: (i) SMaLL100 consistently exhibits lower hallucination rates compared to other NMT models, including its teacher model, M2M (L); (ii) GPT-4 shows impressive performance across the board, with no instances of hallucinations across all language pairs; and (iii) ChatGPT may hallucinate more than NMT models for some language pairs, generating hallucinations that are qualitatively different from those produced by these models, often consisting of failed attempts to translate (e.g., ''Sorry, but my language module cannot translate English text into Russian.'').

Figure 1 :
Figure 1: Hallucination rates among candidate translations for each model in the languages considered.Patternfilled cells indicate at least one hallucination for a given model-language pair.

Figure 2 :
Figure 2: Heatmap of the rate of hallucinations detected with TNG (oscillatory hallucinations) among all hallucinations.Patterned cells indicate at least one natural hallucination for a given model-LP pair.

Figure 3 :
Figure 3: Distribution of SMaLL100 and M2M (L) ALTI+ scores for en-xx and xx-en directions.

Figure 4 :
Figure 4: Reversal rates for oscillatory (Osc.) and detached (Det.)hallucinations when same-family systems (dots) are used as fallback to an original model (labeled above subplots).

Figure 5 :
Figure 5: We analyze overall translation quality improvements on the original model hallucinated translations (represented with dashed lines) across different resource levels via COMET-22 scores in (a), and overall prevalence of oscillatory hallucinations among the fallback translations in (b).

Table 3 :
Average hallucination rates (%) across all LPs at each direction (into or out of English).

Table 5 :
Delta average hallucination rate at each resource level for FLORES and TICO medical data.