CzeGPT-2 headline generator
CzeGPT-2_headline_generator is a Czech summarizer built upon the CzeGPT-2 model. The model has the same architectural dimensions as the GPT-2 small (12 layers, 12 heads, 1024 tokens on input/output, and embedding vectors with 768 dimensions) resulting in 124M trainable parameters. It was fine-tuned and evaluated on the SumeCzech summarization dataset containing about 1M Czech news articles.
Tokenizer
Along, we also provide a Czech trained tokenizer (vocab and merges) with vocab size of 50257 that was used during the pre-training phase and fine-tuning. It is the byte-level BPE tokenizer as used in the original GPT-2 paper.
Training results
The model was evaluated on the test and ood-test partitions of the SumeCzech dataset and compared to the best summarizers yet evaluated on this benchmark (the results taken from here). The headline generator is trained to decide itself when to stop (generate an <|endoftext|> token). If you want a variable summary length, refer to our summary generator
We manage to exceed current state-of-the art on all standard metrics.
Test set
| Model | ROUGERAW-1 | ROUGERAW-2 | ROUGERAW-L |
|---|---|---|---|
| CzeGPT-2 | 17.3/17.0/16.7 | 4.4/4.3/4.2 | 15.5/15.2/14.9 |
| First | 7.4/13.5/8.9 | 1.1/2.2/1.3 | 6.5/11.7/7.7 |
| TextRank | 6.0/16.5/8.3 | 0.8/2.3/1.1 | 5.0/13.8/6.9 |
| Tensor2Tensor | 8.8/7.0/7.5 | 0.8/0.6/0.7 | 8.1/6.5/7.0 |
| NE Density | 6.6/10.7/7.3 | 0.8/1.4/0.9 | 5.9/9.4/6.4 |
| Seq2Seq | 16.1/14.1/14.6 | 2.5/2.1/2.2 | 14.6/12.8/13.2 |
| Seq2SeqNER | 16.2/14.1/14.7 | 2.5/2.1/2.2 | 14.7/12.8/13.3 |
OOD test set
| Model | ROUGERAW-1 | ROUGERAW-2 | ROUGERAW-L |
|---|---|---|---|
| CzeGPT-2 | 17.9/17.6/17.2 | 5.9/5.7/5.5 | 16.4/16.2/15.8 |
| First | 6.7/13.6/8.3 | 1.3/2.8/1.6 | 5.9/12.0/7.4 |
| TextRank | 5.8/16.9/8.1 | 1.1/3.4/1.5 | 5.0/14.5/6.9 |
| Tensor2Tensor | 6.3/5.1/5.5 | 0.5/0.4/0.4 | 5.9/4.8/5.1 |
| NE Density | 6.3/11.4/7.1 | 1.3/2.3/1.4 | 5.7/10.2/6.3 |
| Seq2Seq | 13.1/11.8/12.0 | 2.0/1.7/1.8 | 12.1/11.0/11.2 |
| Seq2SeqNER | 16.2/14.1/14.7 | 2.5/2.1/2.2 | 14.7/12.8/13.3 |
The numbers in the tables denote precision/recall/F1-score
Error Analysis
As we think the current standard ROUGERAW metric is not suitable enough for the summarization task (even though it is the best we have at the time), we performed also a manual error analysis of the generated summaries using human annotators. You can find more about the methodology and results in our paper referenced at the bottom of this card.
Running the predictions
The repository includes a simple Jupyter Notebook that can help with first steps when using the model.
Summary generator
See also our model fine-tuned for summary generation task.
How to cite
@unpublished{hajek_horak2022,
author = "Adam Hájek and Aleš Horák",
title = "CzeGPT-2 – New Model for Czech Summarization Task",
note = "preprint available at \url{https://openreview.net/forum?id=H43eQtxZefq}",
month = "3",
year = "2022",
}
- Downloads last month
- 19