Correct sentence recognition in YouTube importing (Korean)

Hi there. I’m not sure about the exact process how LingQ handles YouTube importing of subtitles, especially for Korean. But I do know that unless subtitles contain punctuation, LingQ fails to accurately identify the ends of sentences. This means that sentence view always breaks long sentences up into sometimes-strange mini-units, and makes the sentence translation function useless.

Punctuation is not very common in Korean in general, and when people make even very accurate YT subtitles they often leave out the periods at the end of sentences, so this creates a consistent problem in Korean language YT imports.

Here is what a YT video with very good subtitles looks like when imported.

You can see that there are no periods anywhere, but any basic algorithm familiar with the language should be able to determine immediately that the “ㅂ/습니다” formal sentence ending demarcates sentences very clearly. (There is also the rhythm of speech, with natural caesura at the ends of sentences, which I think LingQ already detects quite well when matching audio to subtitles?)

Despite this, in sentence view each long sentence is broken up into weird units like this:

This means that the translation function is often misleading and incorrect, since subject particles, whole subject/verb clauses, and even single words may be broken up between units.

For those of us reading advanced material, I wonder if there’s some way that better sentence recognition could be improved when importing Korean? (and perhaps other languages, but I’m not sure whether this is a problem for other languages too.) It would certainly make sentence view useful again.

Thanks!

1 Like

AFAIK LingQ just imports the subtitles from YouTube / Netflix unchanged. I’m not aware of any sentence recognition technology. What I see in sentence mode corresponds to one line in the subtitle. Each line comes with a start and end timestamp this allows LingQ to synchronize the audio to the text, this is used in sentence mode and the “karaoke mode”.
The timestamps are most likely the reason LingQ opts to preserves the formatting.
Discarding the subtitle segments and reassembling the text according to linguistic criteria would also discard the timestamps.
Of course it is not impossible to align audio and text using forced alignment algorithms but for them to be accurate you might need to use a machine learning approach with is computationally expensive and probably still imperfect.
I could be missing something, but I don’t see a simple solution. If timestamps were of no concern it would certainly open up possibilities.
A possible alternative might be to decouple the translator from sentence mode and create translations per lesson. Providing context to the translator might also improve the quality, especially when LingQ switches to ChatGPT in the future.

1 Like