/russian_paraphrasers

Russian paraphrasers. Generate paraphrases with mt5, gpt2, etc.

Primary LanguagePythonMIT LicenseMIT

Russian Paraphrasers

The library for Russian paraphrase generation. Paraphrase generation is an increasingly popular task in NLP that can be used in many areas:

  • style transfer:
    • translation from rude to polite
    • translation from professional to simple language
  • data augmentation: increasing the number of examples for training ML-models
  • increasing the stability of ML-models: training models on a wide variety of examples, in different styles, with different sentiment, but the same meaning / intent of the user

Install

pip install --upgrade pip
pip install -r requirements.txt
pip install russian_paraphrasers

in this case you need to add git+https://github.com/Maluuba/nlg-eval.git@master

Warning important in requirements.txt (versions!):

sentence-transformers==0.4.0
transformers>=4.0.1
git+https://github.com/Maluuba/nlg-eval.git@master

Or you can install last version from git:

pip install git+https://github.com/RussianNLP/russian_paraphrasers@master

Usage

Open In Colab

  1. First, import one of the models and set general parameters for your paraphraser:
from russian_paraphrasers import GPTParaphraser

paraphraser = GPTParaphraser(model_name="gpt2", range_cand=False, make_eval=False)
from russian_paraphrasers import Mt5Paraphraser

paraphraser = Mt5Paraphraser(model_name="mt5-base", range_cand=False, make_eval=False)

You can choose 1) to filter candidates or not 2) to add some evaluation of best candidates or all n samples.

Arguments:

  • model_name: mt5-small, mt5-base, mt5-large, gpt2
  • range_cand: True/False
  • make_eval: True/False
  1. Pass sentence (obligatory) and parameters for generating to generate function and see the results.
sentence = "Мама мыла раму."
results = paraphraser.generate(
    sentence, n=10, temperature=1, 
    top_k=10, top_p=0.9, 
    max_length=100, repetition_penalty=1.5,
    threshold=0.7
)

You can set the threshold parameter to range candidates, it is calculated as similarity score between original vector and the candidate vector.

Results for one sentence look like this:

{'average_metrics': {'Bleu_1': 0.06666666665333353,
                     'Bleu_2': 2.3570227263379004e-09,
                     'Bleu_3': 8.514692649183842e-12,
                     'Bleu_4': 5.665278056606597e-13,
                     'ROUGE_L': 0.07558859975216851},
 'best_candidats': ['В чём цель существования человека?',
                    'Для чего нужна жизнь?',
                    'Что такое жизнь в смысле смысла ее существования, и зачем '
                    'она нужна человеку.'],
 'predictions': ['В чём счастье людей, проживающих в мире сегодня',
                 'В чём счастье человека?)',
                 'Для чего нужна жизнь и какова цель ее существования?',
                 'Что означает фраза в том чтобы жить жизнью?',
                 'В чём ценность человеческой Жизни?',
                 'В чём счастье людей в мире? и т. д.',
                 'Зачем нужна жизнь и что в ней главное докуменция дл',
                 'В чём цель существования человека?',
                 'Что такое жизнь в смысле смысла ее существования, и зачем '
                 'она нужна человеку.',
                 'Для чего нужна жизнь?']
}

Models

All models were fine-tuned on the same dataset (see below) and uploaded to hugging_face. Available models:

To be continued... =)

Dataset

All models were finetuned on the dataset based on two parts:

  1. part of the ParaPhraser data, about 200k filtered examples
  2. filtered questions to chatbots and filtered subtitles from here

The dataset is in the folder dataset.

The article is published and was presenteed in BSNLP. Read it here.