sage-fredt5-large

banner

Summary

The model corrects spelling and punctuation errors and typos by bringing all the words in the text to the norm of the Russian language. Corrector had been trained based on the model FRED-T5-large. An extensive dataset with โ€œartificialโ€ 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

Examples

Input Output
ะ˜ ะฝะต ั‡ัะฝะพ ะฟั€ะพั…ะพะถะธะผ ะฒ ัั‚ะพั‚ ะดะตะฝัŒ ะฝะตะฟะพะณะพะถะนะธ ะฟะพั‡ะตะผัƒ ั ะฒะตัะตะปั‹ะน ั‚ะฐะบะนะพ ะ˜ ะฝะต ััะฝะพ ะฟั€ะพั…ะพะถะธะผ ะฒ ัั‚ะพั‚ ะดะตะฝัŒ ะฝะตะฟะพะณะพะถะธะน, ะฟะพั‡ะตะผัƒ ั ะฒะตัะตะปั‹ะน ั‚ะฐะบะพะน.
ะšะฐะถะดะน ะดะตะฝัŒ ะฒะพั‚ั‚ะฐะบ ะดะตะปะพะน, ะธ ัะฟะตะฝะฐ ะฑะฐะปะตั‚ัŒ ะฝะธะฑัƒะดะตั‚. ะ ะฒะพั‚ะฐะบ ะบะฐะถะดั‹ะน ะดะตะฝัŒ ะฝะธะดะตะปะฐะน ะšะฐะถะดั‹ะน ะดะตะฝัŒ ะฒะพั‚ ั‚ะฐะบ ะดะตะปะฐะน ะธ ัะฟะธะฝะฐ ะฑะพะปะตั‚ัŒ ะฝะต ะฑัƒะดะตั‚. ะ ะฒะพั‚ ั‚ะฐะบ ะบะฐะถะดั‹ะน ะดะตะฝัŒ ะฝะต ะดะตะปะฐะน.
ะžัะฝะพะฒะฐั ั†ะตะปัŒ ะผะตั€ะพะฟั€ะธัั‚ะธั ะฟั€ะฐะบั‚ะธั‡ะตัะบะฐั ะพั‚ั€ะฐะฑะพั‚ะบะฐ ะฝะฐะฒั‹ะบะพะฒ ะฟะพ ะพะบะฐะทะฐะฝะธัŽ ะฟะพะผะพั‰ะธ ะณั€ะฐะถะดะฐะฝะฐะผ, ะฟะพะฟะฐะฒัˆะธะผ ะฒ ะ”ะขะŸ ะฐ ั‚ะฐะบะถะต ะฟะพะฒั‹ัˆะตะฝะธะต ะธ ัะพะฒะตั€ัˆะตะฝัั‚ะฒะพะฒะฐะฝะธะต ัƒั€ะพะฒะฝั ะฟั€ะพั„ะตััะธะพะฝะฐะปัŒะฝะพะน ะฟะพะดะณะพั‚ะพะฒะบะธ ัะพั‚ั€ัƒะดะฝะธะบะพะฒ ะœะงะก ะฟั€ะธ ะฟั€ะพะฒะตะดะตะฝะธะธ ะฐะฒะฐั€ะธะนะฝะพ-ัะฟะฐัะฐั‚ะตะปัŒะฝั‹ั… ั€ะฐะฑะพั‚ ะฟะพ ะปะธะบะฒะธะดะฐั†ะธะธ ะฟะพัะปะตะดัั‚ะฒะธะน ะดะพั€ะพะถะฝะพ-ั‚ั€ะฐะฝัะฟะพั€ั‚ะฝั‹ั… ะฟั€ะพะธัˆะตัั‚ะฒะธะน ัะพะบั€ะฐั‰ะตะฝะธะต ะฒั€ะตะผะตะฝะฝั‹ั… ะฟะพะบะฐะทะฐั‚ะตะปะตะน ั€ะตะฐะณะธั€ะพะฒะฐะฝะธั. ะžัะฝะพะฒะฝะฐั ั†ะตะปัŒ ะผะตั€ะพะฟั€ะธัั‚ะธั โ€” ะฟั€ะฐะบั‚ะธั‡ะตัะบะฐั ะพั‚ั€ะฐะฑะพั‚ะบะฐ ะฝะฐะฒั‹ะบะพะฒ ะฟะพ ะพะบะฐะทะฐะฝะธัŽ ะฟะพะผะพั‰ะธ ะณั€ะฐะถะดะฐะฝะฐะผ, ะฟะพะฟะฐะฒัˆะธะผ ะฒ ะ”ะขะŸ, ะฐ ั‚ะฐะบะถะต ะฟะพะฒั‹ัˆะตะฝะธะต ะธ ัะพะฒะตั€ัˆะตะฝัั‚ะฒะพะฒะฐะฝะธะต ัƒั€ะพะฒะฝั ะฟั€ะพั„ะตััะธะพะฝะฐะปัŒะฝะพะน ะฟะพะดะณะพั‚ะพะฒะบะธ ัะพั‚ั€ัƒะดะฝะธะบะพะฒ ะœะงะก ะฟั€ะธ ะฟั€ะพะฒะตะดะตะฝะธะธ ะฐะฒะฐั€ะธะนะฝะพ-ัะฟะฐัะฐั‚ะตะปัŒะฝั‹ั… ั€ะฐะฑะพั‚ ะฟะพ ะปะธะบะฒะธะดะฐั†ะธะธ ะฟะพัะปะตะดัั‚ะฒะธะน ะดะพั€ะพะถะฝะพ-ั‚ั€ะฐะฝัะฟะพั€ั‚ะฝั‹ั… ะฟั€ะพะธััˆะตัั‚ะฒะธะน, ัะพะบั€ะฐั‰ะตะฝะธะต ะฒั€ะตะผะตะฝะฝั‹ั… ะฟะพะบะฐะทะฐั‚ะตะปะตะน ั€ะตะฐะณะธั€ะพะฒะฐะฝะธั

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 four 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;

RUSpellRU

Model Pr. (spell) Rec. (spell) F1 (spell) Pr. (punc) Rec. (punc) F1 (punc) Pr. (case) Rec. (case) F1 (case)
sage-fredt5-large 57.3 68.0 62.2 86.7 46.1 60.2 92.1 67.8 78.1
sage-fredt5-large (ft) 88.4 80.9 84.5 88.2 85.3 86.8 95.5 94.0 94.7
sage-ai-service 90.3 86.3 88.2 90.3 86.6 88.4 95.2 95.9 95.6
gpt-3.5-turbo 33.6 58.5 42.7 85.9 64.6 73.7 84.9 73.9 79.0
gpt-4 54.9 76.7 64.0 84.0 82.3 83.2 91.5 90.2 90.9

MultidomainGold

Model Pr. (spell) Rec. (spell) F1 (spell) Pr. (punc) Rec. (punc) F1 (punc) Pr. (case) Rec. (case) F1 (case)
sage-fredt5-large 43.4 49.7 46.3 21.8 21.3 21.6 58.8 23.9 34.0
sage-fredt5-large (ft) 80.3 75.1 77.6 69.0 66.5 67.7 78.6 80.0 79.3
sage-ai-service 81.6 77.7 79.6 70.2 67.5 68.8 80.5 80.5 80.5
gpt-3.5-turbo 18.8 48.1 27.1 42.0 31.8 36.2 47.1 51.3 49.1
gpt-4 25.4 68.0 37.0 57.8 54.3 56.0 54.0 67.5 60.0

MedSpellChecker

Model Pr. (spell) Rec. (spell) F1 (spell) Pr. (punc) Rec. (punc) F1 (punc) Pr. (case) Rec. (case) F1 (case)
sage-fredt5-large 35.2 54.5 42.8 19.2 13.2 15.7 48.7 36.8 41.9
sage-fredt5-large (ft) 72.5 72.2 72.3 74.6 66.4 70.3 79.3 85.1 82.1
sage-ai-service 71.3 73.5 72.4 75.1 69.2 72.0 80.9 72.8 76.6
gpt-3.5-turbo 14.7 45.9 22.3 69.9 52.3 59.8 26.4 41.8 32.3
gpt-4 37.8 72.3 49.6 81.4 64.3 71.9 73.0 62.1 67.1

GitHubTypoCorpusRu

Model Pr. (spell) Rec. (spell) F1 (spell) Pr. (punc) Rec. (punc) F1 (punc) Pr. (case) Rec. (case) F1 (case)
sage-fredt5-large 46.0 46.6 46.3 22.7 18.3 20.2 12.0 13.2 12.6
sage-fredt5-large (ft) 67.5 53.2 59.5 48.5 38.0 42.6 37.3 50.0 42.7
sage-ai-service 70.8 56.3 62.7 48.9 35.8 41.4 32.9 45.3 38.1
gpt-3.5-turbo 23.7 38.7 29.4 37.6 23.3 28.7 19.6 35.9 25.3
gpt-4 27.0 52.8 35.7 45.9 32.6 38.2 25.7 36.8 30.2

How to use

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained("ai-forever/sage-fredt5-large")
model = AutoModelForSeq2SeqLM.from_pretrained("ai-forever/sage-fredt5-large", device_map='cuda')

sentence = "ะ˜ ะฝะต ั‡ัะฝะพ ะฟั€ะพั…ะพะถะธะผ ะฒ ัั‚ะพั‚ ะดะตะฝัŒ ะฝะตะฟะพะณะพะถะนะธ ะฟะพั‡ะตะผัƒ ั ะฒะตัะตะปั‹ะน ั‚ะฐะบะนะพ"
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))

# ["ะ˜ ะฝะต ััะฝะพ ะฟั€ะพั…ะพะถะธะผ ะฒ ัั‚ะพั‚ ะดะตะฝัŒ ะฝะตะฟะพะณะพะถะธะน, ะฟะพั‡ะตะผัƒ ั ะฒะตัั‘ะปั‹ะน ั‚ะฐะบะพะน?"]

Limitations

  • The model is intended to be fine-tuned on sets with natural errors for better performance. The realized model is a pre-train and pre-train task is different from the usual spell checking in terms of density of the noise in a corpus and its origin;
  • Complex formatting may cause some trouble in output generation.

Resources

License

Model FRED-T5-large, on the basis of which our solution is made, and its source code are supplied under the MIT license. Our solution comes with MIT license also.

Specifications

  • File size: 3.3 Gb;
  • Framework: pytorch
  • Version: v1.0
  • Developer: SberDevices, AGI NLP

Contacts

nikita.martynov.98@list.ru

Downloads last month
1,057
Safetensors
Model size
0.8B params
Tensor type
F32
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Collection including ai-forever/sage-fredt5-large

Evaluation results