When early machine translation (MT) systems failed, they failed conspicuously. A broken translation was usually a jigsaw puzzle of ungrammatical phrases, untranslated source words, and jarring syntax. Monolingual readers could instantly tell that the system had stumbled.
Modern translation systems—powered by deep Neural Machine Translation (NMT) architectures and Large Language Models (LLMs)—fail in a fundamentally different, more dangerous way. They produce outputs that are grammatically flawless, stylistically polished, and completely wrong. This phenomenon is known as deceptive fluency or translation hallucination.
As Duygu Ataman et al. observe in their survey, Machine Translation in the Era of Large Language Models:A Survey of Historical and Emerging Problems, this deceptive fluency presents a profound challenge to automated translation trust. In high-stakes settings like medical, legal, or safety-critical applications, a beautifully written lie is far more damaging than an obvious, ungrammatical failure. Wilker Aziz pointed out in Predicting Machine Translation Adequacy that monolingual users easily identify grammatical errors but lack the means to detect when a translation has completely inverted the sentiment or fabricated facts.
To understand why modern systems hallucinate, we have to look at the internal tug-of-war that occurs during text generation.
The Tug-of-War: Fluency vs. Adequacy
Every modern autoregressive translation model generates text token by token. At each step, the model balances two competing forces:
- Adequacy (The Source Signal): The conditioning context provided by the source text encoder. This force demands that the next word accurately reflect the original meaning.
- Fluency (The Target Prior): The language modeling capability of the decoder. This force demands that the next word make grammatical and stylistic sense given the target tokens generated so far.
A hallucination occurs when the target language model prior completely overpowers the source signal. Instead of translating, the model begins to generate text autonomously, prioritizing target-language naturalness over source-language fidelity.
[Source Text] ---> ENCODER ---\
===> [Tug-of-War] ---> Next Token
[Tokens So Far] -> DECODER ---/
This imbalance is exacerbated by several architectural and data-driven factors:
Missing Coverage Tracking
Traditional phrase-based translation systems kept strict checklists of which source words had been processed. Modern attention-based architectures do not inherently possess this constraint. As Haitao Mi et al. explain in Coverage Embedding Models for Neural Machine Translation, without explicit mechanisms to monitor translation progress, models frequently suffer from over-translation (repeating content) or under-translation (omitting critical information entirely).
Data Memorization and Noise
Models are trained to memorize long-tail patterns and clean up noisy training data. In The Curious Case of Hallucinations in Neural Machine Translation, Vikas Raunak demonstrates that hallucinations under input noise are strongly linked to this memorization behavior. When confronted with minor perturbations or unfamiliar inputs, the model defaults to generating fluent, detached text sequences it saw during training rather than attempting an accurate mapping.
The Structural Signature of a Hallucination
Because hallucinated text looks perfectly normal on the surface, standard evaluation metrics like BLEU—which look at n-gram overlaps—fail to penalize them effectively. If a model generates a beautiful sentence that shares a few common words with the reference, it can still score highly despite a total semantic failure (Qiyu Wu et al. in Word Alignment as Preference for Machine Translation).
However, if we look inside the model during generation, hallucinations leave distinct structural signatures.
Attention Pathology
In a healthy translation step, the model distributes its attention across the source tokens dynamically, moving from the beginning to the end of the sentence as translation progresses.
Weijia Xu et al. investigated this internal behavior in Understanding and Detecting Hallucinations in Neural Machine Translation via Model Introspection. They discovered that when a model begins to hallucinate, its cross-attention mechanism "gets stuck." Instead of scanning the source sentence, the attention weights concentrate heavily on a tiny, often irrelevant subset of input tokens. The model effectively disconnects its reading mechanism while keeping its writing mechanism running at full speed.
Healthy Attention Matrix: Hallucinating Attention Matrix:
Target Target
^ ^
| . . .
* | .
* . .
| . .
* . | .
* . .
| .
* . . | .
* . .
| * . . . | .
* . .
+-------------> Source +-------------> Source
(Dynamic Scan) (Stuck on Single Token)
Downstream Confidence Drops
Another way to detect hallucinations without a reference text is to observe how the translation performs in a downstream task. In SimQA: Detecting Simultaneous MT Errors through Word-by-Word Question Answering, HyoJung Han et al. demonstrated that when an MT system fabricates facts to maintain flow, the confidence scores of a downstream question-answering model processing that output drop precipitously. The downstream model's uncertainty acts as a proxy indicator that the translation has diverged from the source facts.
Restoring Alignment
To mitigate deceptive fluency, researchers focus on forcing the model to anchor its target generation strictly to the source semantics.
Word Alignment Preferences
One prominent strategy is to explicitly reward the model for maintaining tight word-level correspondences. Qiyu Wu et al. developed a method called Word Alignment Preference (WAP), which leverages Direct Preference Optimization (DPO). By training the model to prefer translations that exhibit clean, verifiable source-to-target token alignments, they significantly reduce both hallucinations (extrapolating extra info) and omissions (dropping source info).
Mutual Information Normalization
Another approach modifies the training objective to prevent target fluency from dominating. Youngwon Lee 0003 et al. introduced Normalized Pointwise Mutual Information (NPMI) in Normalizing Mutual Information for Robust Adaptive Training for Translation. By incorporating a source-side language model to scale the token importance scores, NPMI ensures that a target token is only generated if its probability is strictly justified by the source text, rather than just being a highly probable next word in the target language.
Through reference-free Quality Estimation (QE) frameworks—such as those evaluated in the WMT shared tasks (Lucia Specia et al., Findings of the WMT 2021 Shared Task on Quality Estimation)—the translation community continues to refine methods that balance target-side elegance with strict semantic accountability, working to ensure that automated translations are not just fluent, but trustworthy.
Key Foundational Work
Robust Processing in Machine Translation (1984) — contributed 5 concept elements to this direction.