The fully working model is on huggingface, and a nice chat is on HF Space as well.
A pet project as the (only) other solution does not seem to be maintainable by either owner or others. Some others do exist and are, like, bad.
E.g. for input
я купил iphone 10X за 14990 руб без 3-x часов полдень и т.д.
output of
- russian_stt_text_normalization itself is
я купил ифон десять кс за четыре девять девять ноль рублей без третьи часов полдень и т.д.
- text-normalization-ru-terrible is
я купил айфон сто икс за тысячу четыреста девяносто рубле без третьих часов пол
,- text-normalization-ru-new is
я купил ифон десять икс за четырнадцать тысяч девять
.
I went along with! Took these steps:
- Get a dataset.
Done with notebooks to find and to itn texts, then to construct dataset.
- Download any vast (informal?) russian raw text corpus. Could be
- Find occurances w/ regexp patterns like
r"двадцат\S+"
, - Make sure there is nothing but cyrillic.
- Make inverse text normalization (that task is more straightforward and many good solutions do exist).
- Used
NeMo Text Processinganother python package with some additions.
- Used
- Polish things roughly like balance (as
два
seems to be far more common thanдвумястами
), get rid of ITN mistakes etc.
- Train an MVP.
Done with notebooks to train and to distributed train a model.
- Get a relatively big LLM as we are going to prune it after (and to onnx it as well so that the resulting performance is compatible with the solution I've mentioned).
-
Seems to be ai-forever/FRED-T5-1.7B as it is encoder-decoder, trainable on single RTX3060 12GB and good enough to get an MVP.
Turned out that 12GB is enough to inference it only so I've trained ai-forever/FRED-T5-large.
I've managed to run FRED-T5-1.7B train on two 12GB GPUs using
tensor_parallel
package but model did not perform notably better. Also, the point is to have a small and fast model to infer it on CPU.
-
- Train, like, any barely working model.
- Several attempts are required as it is not clear which prompt is better. May be
<SC1>Было у отца [3]<extra_id_0> сына и [2-3]<extra_id_1> пиджака с блёстками.
Turned out the pattern below works well so I've made no experiments here.
- Several attempts are required as it is not clear which prompt is better. May be
- Test and analyze.
Regret deeply.
- Get a relatively big LLM as we are going to prune it after (and to onnx it as well so that the resulting performance is compatible with the solution I've mentioned).
- To obtain a dataset of a better quality, we want to ask really big smart ass LLM to (not inverse!) normalize texts during the training.
- Unfortunately, LLM experiments failed. I took instruct models (Mistral-7B-Instruct-v0.2, ruGPT-3.5 13B LoRA, GigaSaiga, Saiga2 7B) and plain generation ones (ruGPT-3.5-13B-GPTQ and Vikhr-7b-0.1), but there were always too many mistakes which can not be catch automatically. Well, they were, so I decided to...
- Take the Kaggle Text Normalization Challenge dataset! So I had latin normalization as well.
Done with a notebook to process kaggle data.
- Train everything again at last. Put on hf.
There are but a few packages from namely
- NVidia's NeMo,
- Oknolaz,
- SergeyShk and its forks from
- averkij and
- flockentanz.
NeMo works well but tends to miss many cases I won't have missed (see the comparison table below). I used it as the first attempt but did my research then.
Oknolaz needs to be fed with extracted numbers only and does many mistakes in that case even so bad choice for us.
SergeyShk does either
replace_groups
—тысяча сто
to1100
butсто двести триста
to400
orreplace
—сто двести триста
to100 200 300
butтысяча сто
to1000 100
.
It is obvious that addition should be done on decreasing values only so there are some forks to fix it (the overall code is a mess so that I didn't want to do it myself anyway).
averkij and flockentanz work fine both but have some bugs so I took the second one and fixed them. Also I cover cases like с половиной
and одна целая две десятых
.
Original | 🟡 NeMo TP | 🔴 Oknolaz replace |
🔴 SergeyShk replace_groups |
🔴 SergeyShk replace |
🔴 averkij replace |
🔴 flockentanz replace_groups_sa |
🟢 flockentanz fixed |
---|---|---|---|---|---|---|---|
сто двести триста да хоть тысячу раз |
🟢100 200 300 да хоть 1000 раз |
🔴600000 |
🔴400 да хоть 1000 раз |
🟢100 200 300 да хоть 1000 раз |
🔴10200 300 да хоть 1000 раз |
🟢100 200 300 да хоть 1000 раз |
🟢100 200 300 да хоть 1000 раз |
тысяча сто |
🟢1100 |
🟢1100 |
🟢1100 |
🔴1000 100 |
🟢1100 |
🟢1100 |
🟢1100 |
я видел сто-двести штук |
🟡я видел сто-двести штук |
🔴300 |
🟢я видел 100-200 штук |
🟢я видел 100-200 штук |
🟢я видел 100-200 штук |
🟢я видел 100-200 штук |
🟢я видел 100-200 штук |
восемь девятьсот двадцать два пять пять пять тридцать пять тридцать пять, лучше позвонить, чем занимать |
🟡восемь 922 пять пять пять 35 35 , лучше позвонить, чем занимать |
🔴8 |
🔴115, лучше позвонить, чем занимать |
🔴8 900 20 2 5 5 5 30 5 30 5, лучше позвонить, чем занимать |
🟢8 922 5 5 5 35 35, лучше позвонить, чем занимать |
🟢8 922 5 5 5 35 35, лучше позвонить, чем занимать |
🟢8 922 5 5 5 35 35, лучше позвонить, чем занимать |
три с половиной человека |
🟡три с половиной человека |
🔴3 |
🟡3 с половиной человека |
🟡3 с половиной человека |
🟢3.5 человека |
🟡3 с половиной человека |
🟢3.5 человека |
миллион сто тысяч сто зайцев |
🟢1100100 зайцев |
❌list index out of range |
🔴1000100100 зайцев |
🔴1000000 100000 100 зайцев |
1100100 зайцев |
🔴1000100100 зайцев |
🟢1100100 зайцев |
одни двойки и ни одной пятёрки |
🟡одни двойки и ни одной пятёрки |
🟡No valid number words found! ... |
🟡1 двойки и ни 1 пятёрки |
🟡1 двойки и ни 1 пятёрки |
🟡1 двойки и ни 1 пятёрки |
🟡1 двойки и ни 1 пятёрки |
🟡1 двойки и ни 1 пятёрки |
без одной минуты два |
🟢 01:59 |
🔴2 |
🟢без 1 минуты 2 |
🟢без 1 минуты 2 |
🟢без 1 минуты 2 |
🟢без 1 минуты 2 |
🟢без 1 минуты 2 |
вторая дача пять соток |
🟡вторая дача пять соток |
🔴5 |
🟢2 дача 5 соток |
🟢2 дача 5 соток |
🟢2 дача 5 соток |
🟢2 дача 5 соток |
🟢2 дача 5 соток |
двести пятьдесят с половиной тысяч отборных солдат Ирака |
🟡250 с половиной 1000 отборных солдат Ирака |
🔴250000 |
🟡250 с половиной 1000 отборных солдат Ирака |
🔴200 50 с половиной 1000 отборных солдат Ирака |
🔴2050000.5 отборных солдат Ирака |
🟡250 с половиной 1000 отборных солдат Ирака |
🟢250500 отборных солдат Ирака |
ноль целых ноль десятых минус две целых шесть сотых |
🟢0,0 -2,06 |
🟡Redundant number word! ... |
🔴0 целых 0.0 минус 2 целых 0.06 |
🔴0 целых 0.0 минус 2 целых 0.06 |
🔴0 целых 0.0 минус 2 целых 0.06 |
🔴0 целых 0.0 минус 2 целых 0.06 |
🟢0 минус 2.06 |