Under-representation of languages in training data

Imagine a student who is most comfortable writing in Telugu and asks a chatbot a detailed question about a school subject. The answer comes back weaker than it would have been in English. Some sentences feel off, and a few facts are hard to trust. The student did nothing wrong, and the question was not harder than an English one. The gap comes from how the model was built and what it was taught on.

This article follows that gap through three places where it shows up: the data a model learns from, the translation shortcuts people use to work around it, and the way text is split into tokens, which affects both speed and price. Along the way you will learn how to compute a simple ratio that measures how under-represented a language is in web data, and how to read that ratio without drawing the wrong conclusion.

A model learns patterns from the text it is trained on. If a language appears only rarely in that text, the model has fewer examples to learn from, and it tends to perform worse in that language. The source points to a concrete case: Telugu, Marathi, and Punjabi are the three languages where GPT-4 performed worst on the MMLU benchmarks, and these same three languages are among the most under-represented in Common Crawl.

The mechanism is about how much material the model sees, not about whether the language is worth learning. The source describes this as a major reason for worse performance, which is why representation in the training data is the first place to look when a model is weak in a language.

Low representation is a major factor, but it is not the only one. The source notes that a language's structure and the culture it carries can make it harder for a model to learn, even when there is a reasonable amount of text. Coverage tells you how much material exists. Structure and culture affect how difficult that material is to learn from.

Since LLMs are generally good at translation, a tempting workaround is to translate every non-English query into English, let the model answer in English, and then translate that answer back into the original language. The source treats this as a reasonable but imperfect option, for two separate reasons. The diagram below traces the path a query takes through this pipeline and marks where meaning can be lost.

The first reason is a dependency. The translation step only works well if the model understands the original language sufficiently. For an under-represented language, the translation itself can be the weak point, so the workaround assumes a capability that may be missing.

The second reason is information loss. Some languages mark the relationship between speakers directly in their pronouns. Vietnamese is the source's example. When such a sentence is translated into English, the various pronouns all become I and you, so the information about how the two speakers relate to each other disappears. The English model never sees it, and the translated-back reply cannot restore what was dropped.

Quality problems are not the only way language affects a model. The source reports that model behavior can also differ by language in ways that matter for safety. NewsGuard tested ChatGPT-3.5 in April 2023 by asking it to write misinformation articles about China. The same kind of request was made in English, in simplified Chinese, and in traditional Chinese. The model declined to produce false claims for six of seven English prompts, but it produced false claims all seven times in simplified Chinese and all seven times in traditional Chinese.

The source states that the cause of this difference is unclear. What it does establish is that the same model can behave differently depending on the language of the request, so a safety check run only in English does not cover the other languages.

A model reads and writes text as tokens, and the amount of work it does depends on how many tokens are involved. According to the source, a model's inference latency and cost are both proportional to the number of tokens in the input and the response. If a language needs more tokens to express the same meaning, every request in that language takes longer and costs more, even when the model's quality is the same.

The source gives a benchmark. Yennie Jun tested GPT-4 on MASSIVE, a dataset of one million short texts translated across 52 languages. For the same meaning, Burmese and Hindi needed far more tokens than English or Spanish. The median token length was 7 in English, 32 in Hindi, and 72 in Burmese.

The same division works for Hindi: $\frac{32}{7} \approx 4.6$, so a Hindi message needs about 4.6 times the tokens of the English one. The ratio is the tool for comparing any two languages, and it is the number to use when estimating a bill or response time from a token count.

The source measures under-representation by comparing two percentages for each language: the percent of the world population that speaks it, and the percent of Common Crawl written in it. Dividing the first by the second gives a single number. The question is not simply whether a language appears in the data, but whether its presence is in line with how many people speak it.

If the two shares matched perfectly, the ratio would be 1. The further the ratio rises above 1, the more the language is under-represented relative to its speakers. A ratio below 1 indicates the opposite: a language that is over-represented in the crawl compared with its speaker share. Because both numbers are percentages, the percent sign cancels and the result is a plain number.

The source connects the representation gap to the behavior of general-purpose models such as GPT-4. English holds a very large share of Common Crawl, so models that draw on this text tend to work much better in English than in other languages. The concept map below links the pieces: the crawl composition, the training data, the model, and the per-language scores.

The clearest example in the source is the MMLU benchmark, a suite of 14,000 multiple-choice problems spanning 57 subjects. GPT-4 performed much better on MMLU in English than on the same kind of questions in under-represented languages such as Telugu. The questions in the other languages were produced by translating the English versions with Azure AI Translator.

A second line of evidence comes from math. Yennie Jun gave GPT-4 six problems from Project Euler in an informal test. GPT-4 solved problems posed in English more than three times as often as problems posed in Armenian or Farsi, and it failed every one posed in Burmese and Amharic. Because the test used only six problems, read it as an illustration of a pattern rather than a precise measurement. Its value is that the same direction of disparity appears in a task that requires multi-step reasoning, not just recall of facts.

One response to these gaps is to train models that focus on a particular non-English language. The source says many such models exist. Chinese is the most active language after English, with models such as ChatGLM, YAYI, and Llama-Chinese. Other examples include CroissantLLM for French, PhoGPT for Vietnamese, and Jais for Arabic, and the source notes there are many more.