summarus

Abstractive and extractive summarization models, mostly for Russian language. Building on top of AllenNLP

You can also checkout the MBART-based Russian summarization model on Huggingface: mbart_ru_sum_gazeta

Based on the following papers:

• SummaRuNNer: A Recurrent Neural Network based Sequence Model for Extractive Summarization of Documents
• Get To The Point: Summarization with Pointer-Generator Networks
• Self-Attentive Model for Headline Generation
• Text Summarization with Pretrained Encoders
• Multilingual Denoising Pre-training for Neural Machine Translation

Contacts

• Telegram: @YallenGusev

Prerequisites

pip install -r requirements.txt

Commands

train.sh

Script for training a model based on AllenNLP 'train' command.

predict.sh

Script for model evaluation. The test dataset should have the same format as the train dataset.

summarus.util.train_subword_model

Script for subword model training.

Headline generation

• First paper: Importance of Copying Mechanism for News Headline Generation
• Slides: Importance of Copying Mechanism for News Headline Generation
• Second paper: Advances of Transformer-Based Models for News Headline Generation

Dataset splits:

• RIA original dataset: https://github.com/RossiyaSegodnya/ria_news_dataset
• RIA train/val/test: https://www.dropbox.com/s/rermx1r8lx9u7nl/ria.tar.gz
• RIA dataset preprocessed for mBART: https://www.dropbox.com/s/iq2ih8sztygvz0m/ria_data_mbart_512_200.tar.gz
• Lenta original dataset: https://github.com/yutkin/Lenta.Ru-News-Dataset
• Lenta train/val/test: https://www.dropbox.com/s/v9i2nh12a4deuqj/lenta.tar.gz
• Lenta dataset preprocessed for mBART: https://www.dropbox.com/s/4oo8jazmw3izqvr/lenta_mbart_data_512_200.tar.gz
• Telegram train dataset with split: https://www.dropbox.com/s/ykqk49a8avlmnaf/ru_all_split.tar.gz
• Telegram test dataset with multiple references: https://github.com/dialogue-evaluation/Russian-News-Clustering-and-Headline-Generation/blob/main/data/headline_generation/headline_generation_answers.jsonl

Models:

• ria_copynet_10kk
• ria_pgn_24kk
• ria_mbart
• rubert_telegram_headlines

Prediction script:

./predict.sh -t <path_to_test_dataset> -m ria_pgn_24kk.tar.gz -p subwords_summary -L ru 

Results

Train dataset: RIA, test dataset: RIA

Train dataset: RIA, eval dataset: Lenta

Summarization - CNN/DailyMail

Dataset splits:

• CNN/DailyMail jsonl dataset: https://www.dropbox.com/s/35ezpg78rtukkgh/cnn_dm_jsonl.tar.gz

Models:

• cnndm_pgn_25kk

Prediction script:

./predict.sh -t <path_to_test_dataset> -m cnndm_pgn_25kk.tar.gz -p words_summary -L en -R

Results:

Summarization - Gazeta, russian news dataset

• Paper: Dataset for Automatic Summarization of Russian News
• Gazeta dataset: https://github.com/IlyaGusev/gazeta
• Usage examples:

Models:

• gazeta_pgn_7kk
• gazeta_pgn_7kk_cov.tar.gz
• gazeta_pgn_25kk
• gazeta_pgn_words_13kk.tar.gz
• gazeta_summarunner_3kk

Prediction scripts:

./predict.sh -t <path_to_test_dataset> -m gazeta_pgn_7kk.tar.gz -p subwords_summary -L ru -T
./predict.sh -t <path_to_test_dataset> -m gazeta_summarunner_3kk.tar.gz -p subwords_summary_sentences -L ru -T

External models:

• gazeta_mbart (fairseq)
• gazeta_mbart (transformers)
• gazeta_mbart_lowercase (fairseq)

Results:

Demo

python demo/server.py --include-package summarus --model-dir <model_dir> --host <host> --port <port>

Citations

Headline generation (PGN):

@article{Gusev2019headlines,
    author={Gusev, I.O.},
    title={Importance of copying mechanism for news headline generation},
    journal={Komp'juternaja Lingvistika i Intellektual'nye Tehnologii},
    year={2019},
    volume={2019-May},
    number={18},
    pages={229--236}
}

Headline generation (transformers):

@InProceedings{Bukhtiyarov2020headlines,
    author={Bukhtiyarov, Alexey and Gusev, Ilya},
    title="Advances of Transformer-Based Models for News Headline Generation",
    booktitle="Artificial Intelligence and Natural Language",
    year="2020",
    publisher="Springer International Publishing",
    address="Cham",
    pages={54--61},
    isbn="978-3-030-59082-6",
    doi={10.1007/978-3-030-59082-6_4}
}

Summarization:

@InProceedings{Gusev2020gazeta,
    author="Gusev, Ilya",
    title="Dataset for Automatic Summarization of Russian News",
    booktitle="Artificial Intelligence and Natural Language",
    year="2020",
    publisher="Springer International Publishing",
    address="Cham",
    pages="{122--134}",
    isbn="978-3-030-59082-6",
    doi={10.1007/978-3-030-59082-6_9}
}