🎯 sage-mt5-large

The model is a part of SAGE v1.1.0 release

../../_images/sage_banner.jpg

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 library.

Table of contents

Public references

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

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)."]

API

class sage.spelling_correction.t5_correctors.T5ModelForSpellingCorruption(model_name_or_path: str | PathLike)

Bases: Corrector

T5-based models.

batch_correct(sentences: List[str], batch_size: int, prefix: str | None = '', **generation_params) → List[List[Any]]

Corrects multiple sentences.

Parameters:

  • sentences (list of str) – input sentences to correct;

  • batch_size (int) – size of subsample of input sentences;

  • prefix (str) – some models need some sort of a prompting;

  • generation_params (dict) – parameters passed to generate method of a HuggingFace model;

Returns:

corresponding corrections

Return type:

list of list of str

correct(sentence: str, prefix: str | None = '', **generation_params) → List[str]

Corrects a single input sentence.

Parameters:

  • sentence (str) – a source sentence;

  • prefix (str) – some models need some sort of a prompting;

  • generation_params (dict) – parameters passed to generate method of a HuggingFace model;

Returns:

corresponding corrected sentence

Return type:

list of str

evaluate(dataset_name_or_path: str | PathLike | None, metrics: List, batch_size: int, prefix: str = '', dataset_split: str = 'test', **generation_params) → Dict[str, float]

Evaluate the particular model on the spellcheck datasets.

Parameters:

  • dataset_name_or_path (str) – a path to a locally situated dataset or a name of a dataset on HuggingFace;

  • metrics (list of str) – set of metrics to be used to report performance;

  • batch_size (int) – size of subsample of input sentences;

  • prefix (str) – some models need some sort of a prompting;

  • dataset_split (str) – train / test / dev part to be evaluated on;

  • generation_params (dict) – parameters passed to generate method of a HuggingFace model;

Returns:

mapping between metric’s name and its corresponding value

Return type:

dict[str, float]

classmethod from_pretrained(model_name_or_path: str | PathLike)

Initialize the T5-type corrector from a pre-trained checkpoint. The latter can be either locally situated checkpoint or a name of a model on HuggingFace.

Parameters:

model_name_or_path (str or os.PathLike) – the aforementioned name or path to checkpoint;

Returns:

corrector initialized from pre-trained weights

Return type:

object of T5ModelForSpellingCorruption

Limitations

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

Resources

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.

Specifications

  • File size: 5 Gb;

  • Framework: pytorch

  • Format: AI Service

  • Version: v1.0

  • Developer: SberDevices, AGI NLP

Contacts

nikita.martynov.98@list.ru