What is the correct answer for detect() on a multilingual string?

[fergald]

If a string is 100 chars of English followed by 900 characters of French, what is the ideal result? Is it the following?

[
  {language: "en", confidence:.1},
  {language: "fr", confidence:.9},
]

I haven't been able to come up with a better idea than that each language in the result should tell you what fraction of the string is in that language.

This gets more complicated when the language of segments of the string are themselves ambiguous. E.g. for an English article talking about words that are shared between Chinese and Japanese, what is the correct answer? Assuming the text is 80% English with 10% of it being Chinese/Japanese. What is the ideal result? Is it

  {language: "en", confidence:.8},
  {language: "ja", confidence:.0.1},
  {language: "zh", confidence:.0.1},
]

even though 20% of the text is Japanese and 20% is Chinese? I can't think of a better "correct" answer but maybe there is one.

Also from an implementation perspective, the above "correct" answer is relatively easy. Models may have a fixed maximum input size and the above can be calculated by breaking the string into chunks and averaging over the results for each chunk.

Questions

• Should we even be trying to spec level of detail?
• If so, should we spec the above?

[domenic]

Some interesting precedents:

• https://learn.microsoft.com/en-us/azure/ai-services/language-service/language-detection/how-to/call-api#mixed-language-content (returns the top language with <1 confidence)
• https://github.com/pemistahl/lingua-py?tab=readme-ov-file#116-detection-of-multiple-languages-in-mixed-language-texts (experimental mode returns multiple spans, with a different detected language for each)

Many other APIs don't seem to explicitly document their strategy.

[fergald]

I think annotated spans is the luxury version but will be harder to implement and overkill for many uses. Models that take a short text and assign weights to various languages already exist and by breaking a longer string into chunks and averaging over the results you naturally get something that measures how much of each language there is in a given text.