sage-mt5-large

Summary

The model corrects spelling errors and typos in both Russian and English languages by bringing all the words in the text to the norm of the language. Corrector had been trained based on the model mT5-large architecture. An extensive dataset with “artificial” errors was taken as a training corpus: the corpus was assembled on the basis of the Russian-language Wikipedia and transcripts of Russian-language videos, then typos and spelling errors were automatically introduced into it using the library SAGE.

Public references

SAGE library announcement, DataFest 2023
Paper about synthetic error generation methods, Dialogue 2023
SAGE EACL 2024 paper

Examples

Input	Output
Перведи мне текст на аглиском: "Screw you kuys, I am goin hme (c).	Переведи мне текст на английском: "Screw you guys, I am going home" (c).
И не чсно прохожим в этот день непогожйи почему я веселый такйо	И мне ясно прохожим в этот день непогожий, почему я веселый такой
If you bought something goregous, you well be very happy.	If you bought something gorgeous, you will be very happy.

Metrics

Quality

Below are automatic metrics for determining the correctness of the spell checkers. We compare our solution with both open automatic spell checkers and the ChatGPT family of models on all six available datasets:

RUSpellRU: texts collected from (LiveJournal), with manually corrected typos and errors;
MultidomainGold: examples from 7 text sources, including the open web, news, social media, reviews, subtitles, policy documents and literary works;
MedSpellChecker: texts with errors from medical anamnesis;
GitHubTypoCorpusRu: spelling errors and typos in commits from GitHub;
BEA60K: English spelling errors collected from several domains;
JFLEG: 1601 sentences in English, which contain about 2 thousand spelling errors;

RUSpellRU, MultidomainGold, MedSpellChecker, GitHubTypoCorpusRu are datasets for the Russian spellchecking and BEA60K and JFLEG are those for the English language.

RUSpellRU

Model	Precision	Recall	F1
sage-mt5-large	55.7	68.5	61.4
sage-mt5-large (ft.)	88.4	71.6	79.1
sage-ai-service	93.5	82.4	87.6
gpt-3.5-turbo	39.6	62.3	48.5
gpt-4	69.5	81.0	74.8

MultidomainGold

Model	Precision	Recall	F1
sage-mt5-large	35.4	57.9	43.9
sage-mt5-large (ft.)	65.3	62.7	63.9
sage-ai-service	70.9	68.8	69.9
gpt-3.5-turbo	17.8	56.1	27.0
gpt-4	31.1	78.1	44.5

MedSpellChecker

Model	Precision	Recall	F1
sage-mt5-large	35.1	70.8	47.0
sage-mt5-large (ft.)	77.7	77.5	77.6
sage-ai-service	73.4	76.2	74.9
gpt-3.5-turbo	15.1	53.6	23.5
gpt-4	48.9	88.7	63.1

GitHubTypoCorpusRu

Model	Precision	Recall	F1
sage-mt5-large	47.4	53.8	50.4
sage-mt5-large (ft.)	69.5	46.0	55.3
sage-ai-service	76.1	51.2	61.2
gpt-3.5-turbo	23.7	43.9	30.8
gpt-4	34.7	60.5	44.1

BEA60K

Model	Precision	Recall	F1
sage-mt5-large	64.7	83.8	73.0
gpt-3.5-turbo	66.9	84.1	74.5
gpt-4	68.6	85.2	76.0
Bert (https://github.com/neuspell/neuspell)	65.8	79.6	72.0
SC-LSTM (https://github.com/neuspell/neuspell)	62.2	80.3	72.0

JFLEG

Model	Precision	Recall	F1
sage-mt5-large	74.9	88.4	81.1
gpt-3.5-turbo	77.8	88.6	82.9
gpt-4	77.9	88.3	82.8
Bert (https://github.com/neuspell/neuspell)	78.5	85.4	81.8
SC-LSTM (https://github.com/neuspell/neuspell)	80.6	86.1	83.2

How to use

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained("ai-forever/sage-mt5-large")
model = AutoModelForSeq2SeqLM.from_pretrained("ai-forever/sage-mt5-large", device_map='cuda')

sentence = "Перведи мне текст на аглиском: \"Screw you kuys, I am goin hme (c)."
inputs = tokenizer(sentence, max_length=None, padding="longest", truncation=False, return_tensors="pt")
outputs = model.generate(**inputs.to(model.device), max_length = inputs["input_ids"].size(1) * 1.5)
print(tokenizer.batch_decode(outputs, skip_special_tokens=True))

# ["Переведи мне текст на английском: "Screw you guys, I am going home" (c)."]

Limitations

For the Russian language the model is intended to be fine-tuned for better performance.

Resources

SAGE library, GitHub
sage-fredt5-large, HuggingFace
sage-fredt5-distilled-95m, HuggingFace
sage-m2m100-1.2B, HuggingFace
sage-mt5-large, HuggingFace

License

Model mT5-large, on the basis of which our solution is made, and its source code are supplied under the Apache-2.0 license. Our solution comes with MIT license.