Numbers missing from English translations and a few other issues

#3
by mpasila - opened

I noticed that a lot of the numbers were missing from the English translations which still show up in the Japanese ones.
Another issue is even though most of the "Ch Chapter" stuff was removed, there are still the names of the chapters left on some of the English translations.
And some translations have unnecessary text at the end like "It is also very straightforward writing, which meshes with me well, so I finished 3 chapters within 8 hours."
And there seems to sometimes be notes added by the translator like: (TN: remember he was carrying the black haired girl since last chapter)
edit: I forgot to also mention some weirdness with some words that have been written like this "finis.h.i.+ng" like I've noticed that several times with words having dots in between them for no reason.

Not sure if there's any easy way to remove these besides manually going through everything (for numbers you could just look up any Japanese sentences with them and spot the missing English ones).

I'm planning to do another sentence alignment run, but especially with the translation notes, every translator formats them differently (square brackets, "T/N" instead of TN, etc.). I'm going to have to do a lot of work on regex filtering.

As for the added stuff at the end, the only reason they'd be present is if they were in the Japanese text, in which case it'd be a normal parallel text, or if the translator added it without line breaks, which would require something like splitting long sentences at the end and beginning of texts, or some equally extravagant method.

The extra punctuation in words I can filter out easily via regex.

I plan cleaning the dataset to be a progressive project, so I'll continue to update it as issues come in.

What do you mean by 'missing numbers'? As in, missing chapter numbers, or missing numbers in general?

Most of the current issues seem to be down to scraper artifacts and my regex filters only being designed for a small test dataset that isn't representative of the big one. I'll need to fix them up.

Owner

Forgot to mention, I've been training a tinyllama model on the dataset, and while it's only a quarter of the way through an epoch, it's pretty promising. Once it's reached a satisfactory level (might take a few days) I'll upload it to my huggingface, so keep an eye out and see what you think.

What do you mean by 'missing numbers'? As in, missing chapter numbers, or missing numbers in general?

In general like here's a few examples:
"ある意味十三親王邸の顔みたいなメイドだ。
In a way, she’s like a head maid of the th Prince’s residence."
"少し前にバイロンからもらった、十万リィーンが入っている箱だ。
This was the box that Byron had given me a while ago, containing ,000 reens."

Owner

That looks like it might be caused by a regex rule I setup to filter out footnote citations like [1] embedded in the text. I'll look into that.

Thanks for finding them, let me know if you see anything else!

Owner

I might tighten up the sentence match threshold and dropout conditions along with some form of quality estimation (maybe a bleu or chrf match against a machine translation to detect them?) to make a smaller version of the dataset to post alongside the big one.

I also noticed this thing where it's missing something in the Japanese text but not in the English one:
【:稀人side】
[First Night: Rare People Side]

Owner

@mpasila I updated the dataset, and it should fix the issues you found.
I didn't address the random punctuation issue you found though, since it would be a pain to remove without collateral damage. It's sparse enough that it likely won't make any meaningful difference.

Nice I'll probably try training a new model with it.

Sign up or log in to comment