MBARTRuSumGazeta

Model description

This is a ported version of fairseq model.

For more details, please see, Dataset for Automatic Summarization of Russian News.

Intended uses & limitations

How to use

from transformers import MBartTokenizer, MBartForConditionalGeneration

article_text = "..."
model_name = "IlyaGusev/mbart_ru_sum_gazeta"
tokenizer = MBartTokenizer.from_pretrained(model_name)
model = MBartForConditionalGeneration.from_pretrained(model_name)

input_ids = tokenizer.prepare_seq2seq_batch(
    [article_text],
    src_lang="en_XX", # fairseq training artifact
    return_tensors="pt",
    padding="max_length",
    truncation=True,
    max_length=600
)["input_ids"]

output_ids = model.generate(
    input_ids=input_ids,
    max_length=162,
    no_repeat_ngram_size=3,
    num_beams=5,
    top_k=0
)[0]

summary = tokenizer.decode(output_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)
print(summary)

Limitations and bias

  • The model should work well with Gazeta.ru articles, but for any other agencies it can suffer from domain shift

Training data

Training procedure

Eval results

Model R-1-f R-2-f R-L-f METEOR BLEU
gazeta_mbart 32.3 14.3 27.9 25.5 48.9

Predicting all summaries:

import json
import torch
from transformers import MBartTokenizer, MBartForConditionalGeneration


def gen_batch(inputs, batch_size):
    batch_start = 0
    while batch_start < len(inputs):
        yield inputs[batch_start: batch_start + batch_size]
        batch_start += batch_size


def predict(
    model_name,
    test_file,
    predictions_file,
    targets_file,
    max_source_tokens_count=600,
    max_target_tokens_count=160,
    use_cuda=True,
    batch_size=4
):
    inputs = []
    targets = []
    with open(test_file, "r") as r:
        for line in r:
            record = json.loads(line)
            inputs.append(record["text"])
            targets.append(record["summary"])

    tokenizer = MBartTokenizer.from_pretrained(model_name)
    device = torch.device("cuda:0") if use_cuda else torch.device("cpu")
    model = MBartForConditionalGeneration.from_pretrained(model_name).to(device)
    predictions = []
    for batch in gen_batch(inputs, batch_size):
        input_ids = tokenizer.prepare_seq2seq_batch(
            batch,
            src_lang="en_XX",
            return_tensors="pt",
            padding="max_length",
            truncation=True,
            max_length=max_source_tokens_count
        )["input_ids"].to(device)
        output_ids = model.generate(
            input_ids=input_ids,
            max_length=max_target_tokens_count + 2,
            no_repeat_ngram_size=3,
            num_beams=5,
            top_k=0
        )
        summaries = tokenizer.batch_decode(output_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)
        for s in summaries:
            print(s)
        predictions.extend(summaries)
    with open(predictions_file, "w") as w:
        for p in predictions:
            w.write(p.strip() + "\n")
    with open(targets_file, "w") as w:
        for t in targets:
            w.write(t.strip() + "\n")

predict("IlyaGusev/mbart_ru_sum_gazeta", "gazeta_test.jsonl", "predictions.txt", "targets.txt")

Evaluation: https://github.com/IlyaGusev/summarus/blob/master/evaluate.py

Flags: --language ru --tokenize-after --lower

BibTeX entry and citation info

@InProceedings{10.1007/978-3-030-59082-6_9,
    author="Gusev, Ilya",
    editor="Filchenkov, Andrey and Kauttonen, Janne and Pivovarova, Lidia",
    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"
}
Downloads last month
675
Hosted inference API
Summarization