Note: This blog post was originally written in Japanese for our Japanese website. We use our machine translation platforms to translate and make automatic corrections, and then partially edit to fit the content in English.
As we discussed in Part 3 of this series, LLMs, which power today's generative AI, and neural machine translation (NMT), a more traditional form of machine translation, differ in their translation capabilities. To summarize briefly, LLMs have the edge in fluency (how natural the translated text sounds), while NMT has the edge in accuracy (how faithfully it reflects the source) and translation speed. But that's not the only difference between the two.
In this post, we'll look at how the training data used for LLMs and NMT differs, and the translation quality issues that result.
Differences in training data and the mechanism of pivot translation
NMT training data consists of pairs of source and target texts: the source text written in one language (for example, Japanese) and the target text in another (for example, English). As a result, each translation model can typically handle only one language direction (for example, Japanese to English).
Many commercial machine translation services offer models that translate from English into other languages and from other languages into English. For example, when translating from Japanese to Chinese, the system first translates from Japanese to English, then from English to Chinese. This is called "pivot translation."
(Note that our machine translation solution, Kawamura NMT, uses translation models for both Japanese-to-other-language and other-language-to-Japanese pairs, so it performs direct translation rather than pivot translation.)
LLM training data, on the other hand, consists of text in a wide range of languages. As a result, a single language model can handle many language pairs, with no need for pivot translation. On that basis, LLMs would appear better suited.
Differences in LLM translation quality caused by language bias
However, a 2023 study we referenced in Part 3 found that while LLM translations are on par with NMT when translating into English, quality drops when translating out of English or between two non-English languages. In other words, LLM performance varies significantly by language direction.
This comes down to a language bias in LLM training data. Most widely used LLMs are trained on a data set weighted heavily toward English. Because these models excel at generating English sentences, they perform well on translations into English, but struggle when the target language isn't English.
LLMs are constantly increasing the volume of data they train on, though. If they keep learning from more non-English text, will translation quality improve for language pairs that don't involve English?
Declining quality of training data and the limitations faced by LLMs
LLMs rely on publicly available data, much of it scraped from the web, for the bulk of their training. But the share of web content that is user-generated or AI-generated (including AI-translated) keeps growing. No matter how much of this lower-quality data an LLM trains on, it can't be expected to improve performance as a result.
In fact, it's well established that training an AI model on its own generated text causes the model to break down. Experiments show that generative AI eventually starts producing nothing but nonsensical output, a phenomenon sometimes called "Habsburg AI," after the Habsburg dynasty, whose repeated intermarriage led to its decline.
The problem of training data shortages and a practical solution using NMT
Training LLMs requires high-quality data, but some studies estimate that publicly available high-quality data could run out as early as 2026*, an issue often called the "2026 Problem." For lower-resource languages, machine-translated text already makes up a large share of what's publicly available, so the decline in translation quality driven by lower-quality training data is expected to become more severe over time.
That said, NMT faces the same shortage of high-quality training data. One solution is to build an organization-specific translation model trained on that organization's own proprietary data, and NMT has an advantage over LLMs here: it's easier to retrain with additional data.
With XMAT, our AI translation platform, you can create and use translation models tailored to your organization on a self-service basis. We can also help with related needs, such as building higher-quality bilingual data or ensuring your bilingual data stays within your organization.
Reference
-
How Good Are GPT Models at Machine Translation? A Comprehensive Evaluation
-
AI Models Collapse When Trained on Recursively Generated Data
-
Will we run out of data? Limits of LLM scaling based on human-generated data
-
A Shocking Amount of the Web is Machine Translated: Insights from Multi-Way Parallelism
Kawamura International's services
As organizations face increasing pressure on translation quality and data security, having the right infrastructure matters. With XMAT, you get high-quality, secure machine translation with no per-character fees. Beyond text and document translation, you can also use the translation editor and the AI-powered (LLM) revision feature.
At Kawamura International, we do more than provide machine translation. We offer a range of solutions to help streamline your translation processes. If you're facing translation challenges, we'd love to hear from you.
* This post was originally published on our Japanese website in January 2026.
