File size: 9,998 Bytes
007953f 28bc819 8e21d2d 28bc819 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 |
---
tags:
- cs
- fill-mask
---
# CZERT
This repository keeps trained Czert-B-base-cased-long-zero-shot model for the paper [Czert β Czech BERT-like Model for Language Representation
](https://arxiv.org/abs/2103.13031)
For more information, see the paper
This is long version of Czert-B-base-cased created without any finetunning on long documents. Positional embedings were created by simply repeating the positional embeddings of the original Czert-B model. For tokenization, please use BertTokenizer. Cannot be used with AutoTokenizer.
## Available Models
You can download **MLM & NSP only** pretrained models
~~[CZERT-A-v1](https://air.kiv.zcu.cz/public/CZERT-A-czert-albert-base-uncased.zip)
[CZERT-B-v1](https://air.kiv.zcu.cz/public/CZERT-B-czert-bert-base-cased.zip)~~
After some additional experiments, we found out that the tokenizers config was exported wrongly. In Czert-B-v1, the tokenizer parameter "do_lower_case" was wrongly set to true. In Czert-A-v1 the parameter "strip_accents" was incorrectly set to true.
Both mistakes are repaired in v2.
[CZERT-A-v2](https://air.kiv.zcu.cz/public/CZERT-A-v2-czert-albert-base-uncased.zip)
[CZERT-B-v2](https://air.kiv.zcu.cz/public/CZERT-B-v2-czert-bert-base-cased.zip)
or choose from one of **Finetuned Models**
| | Models |
| - | - |
| Sentiment Classification<br> (Facebook or CSFD) | [CZERT-A-sentiment-FB](https://air.kiv.zcu.cz/public/CZERT-A_fb.zip) <br> [CZERT-B-sentiment-FB](https://air.kiv.zcu.cz/public/CZERT-B_fb.zip) <br> [CZERT-A-sentiment-CSFD](https://air.kiv.zcu.cz/public/CZERT-A_csfd.zip) <br> [CZERT-B-sentiment-CSFD](https://air.kiv.zcu.cz/public/CZERT-B_csfd.zip) | Semantic Text Similarity <br> (Czech News Agency) | [CZERT-A-sts-CNA](https://air.kiv.zcu.cz/public/CZERT-A-sts-CNA.zip) <br> [CZERT-B-sts-CNA](https://air.kiv.zcu.cz/public/CZERT-B-sts-CNA.zip)
| Named Entity Recognition | [CZERT-A-ner-CNEC](https://air.kiv.zcu.cz/public/CZERT-A-ner-CNEC-cased.zip) <br> [CZERT-B-ner-CNEC](https://air.kiv.zcu.cz/public/CZERT-B-ner-CNEC-cased.zip) <br>[PAV-ner-CNEC](https://air.kiv.zcu.cz/public/PAV-ner-CNEC-cased.zip) <br> [CZERT-A-ner-BSNLP](https://air.kiv.zcu.cz/public/CZERT-A-ner-BSNLP-cased.zip)<br>[CZERT-B-ner-BSNLP](https://air.kiv.zcu.cz/public/CZERT-B-ner-BSNLP-cased.zip) <br>[PAV-ner-BSNLP](https://air.kiv.zcu.cz/public/PAV-ner-BSNLP-cased.zip) |
| Morphological Tagging<br> | [CZERT-A-morphtag-126k](https://air.kiv.zcu.cz/public/CZERT-A-morphtag-126k-cased.zip)<br>[CZERT-B-morphtag-126k](https://air.kiv.zcu.cz/public/CZERT-B-morphtag-126k-cased.zip) |
| Semantic Role Labelling |[CZERT-A-srl](https://air.kiv.zcu.cz/public/CZERT-A-srl-cased.zip)<br> [CZERT-B-srl](https://air.kiv.zcu.cz/public/CZERT-B-srl-cased.zip) |
## How to Use CZERT?
### Sentence Level Tasks
We evaluate our model on two sentence level tasks:
* Sentiment Classification,
* Semantic Text Similarity.
<!-- tokenizer = BertTokenizerFast.from_pretrained(CZERT_MODEL_PATH, strip_accents=False)
model = TFAlbertForSequenceClassification.from_pretrained(CZERT_MODEL_PATH, num_labels=1)
or
self.tokenizer = BertTokenizerFast.from_pretrained(CZERT_MODEL_PATH, strip_accents=False)
self.model_encoder = AutoModelForSequenceClassification.from_pretrained(CZERT_MODEL_PATH, from_tf=True)
-->
### Document Level Tasks
We evaluate our model on one document level task
* Multi-label Document Classification.
### Token Level Tasks
We evaluate our model on three token level tasks:
* Named Entity Recognition,
* Morphological Tagging,
* Semantic Role Labelling.
## Downstream Tasks Fine-tuning Results
### Sentiment Classification
| | mBERT | SlavicBERT | ALBERT-r | Czert-A | Czert-B |
|:----:|:------------------------:|:------------------------:|:------------------------:|:-----------------------:|:--------------------------------:|
| FB | 71.72β
Β±β
0.91 | 73.87β
Β±β
0.50 | 59.50β
Β±β
0.47 | 72.47β
Β±β
0.72 | **76.55**β
Β±β
**0.14** |
| CSFD | 82.80β
Β±β
0.14 | 82.51β
Β±β
0.14 | 75.40β
Β±β
0.18 | 79.58β
Β±β
0.46 | **84.79**β
Β±β
**0.26** |
Average F1 results for the Sentiment Classification task. For more information, see [the paper](https://arxiv.org/abs/2103.13031).
### Semantic Text Similarity
| | **mBERT** | **Pavlov** | **Albert-random** | **Czert-A** | **Czert-B** |
|:-------------|:--------------:|:--------------:|:-----------------:|:--------------:|:----------------------:|
| STA-CNA | 83.335β
Β±β
0.063 | 83.593β
Β±β
0.050 | 43.184β
Β±β
0.125 | 82.942β
Β±β
0.106 | **84.345**β
Β±β
**0.028** |
| STS-SVOB-img | 79.367β
Β±β
0.486 | 79.900β
Β±β
0.810 | 15.739β
Β±β
2.992 | 79.444β
Β±β
0.338 | **83.744**β
Β±β
**0.395** |
| STS-SVOB-hl | 78.833β
Β±β
0.296 | 76.996β
Β±β
0.305 | 33.949β
Β±β
1.807 | 75.089β
Β±β
0.806 | **79.827β
Β±β
0.469** |
Comparison of Pearson correlation achieved using pre-trained CZERT-A, CZERT-B, mBERT, Pavlov and randomly initialised Albert on semantic text similarity. For more information see [the paper](https://arxiv.org/abs/2103.13031).
### Multi-label Document Classification
| | mBERT | SlavicBERT | ALBERT-r | Czert-A | Czert-B |
|:-----:|:------------:|:------------:|:------------:|:------------:|:-------------------:|
| AUROC | 97.62β
Β±β
0.08 | 97.80β
Β±β
0.06 | 94.35β
Β±β
0.13 | 97.49β
Β±β
0.07 | **98.00**β
Β±β
**0.04** |
| F1 | 83.04β
Β±β
0.16 | 84.08β
Β±β
0.14 | 72.44β
Β±β
0.22 | 82.27β
Β±β
0.17 | **85.06**β
Β±β
**0.11** |
Comparison of F1 and AUROC score achieved using pre-trained CZERT-A, CZERT-B, mBERT, Pavlov and randomly initialised Albert on multi-label document classification. For more information see [the paper](https://arxiv.org/abs/2103.13031).
### Morphological Tagging
| | mBERT | Pavlov | Albert-random | Czert-A | Czert-B |
|:-----------------------|:---------------|:---------------|:---------------|:---------------|:---------------|
| Universal Dependencies | 99.176β
Β±β
0.006 | 99.211β
Β±β
0.008 | 96.590β
Β±β
0.096 | 98.713β
Β±β
0.008 | **99.300β
Β±β
0.009** |
Comparison of F1 score achieved using pre-trained CZERT-A, CZERT-B, mBERT, Pavlov and randomly initialised Albert on morphological tagging task. For more information see [the paper](https://arxiv.org/abs/2103.13031).
### Semantic Role Labelling
<div id="tab:SRL">
| | mBERT | Pavlov | Albert-random | Czert-A | Czert-B | dep-based | gold-dep |
|:------:|:----------:|:----------:|:-------------:|:----------:|:----------:|:---------:|:--------:|
| span | 78.547 Β± 0.110 | 79.333 Β± 0.080 | 51.365 Β± 0.423 | 72.254 Β± 0.172 | **81.861 Β± 0.102** | \- | \- |
| syntax | 90.226 Β± 0.224 | 90.492 Β± 0.040 | 80.747 Β± 0.131 | 80.319 Β± 0.054 | **91.462 Β± 0.062** | 85.19 | 89.52 |
SRL results β dep columns are evaluate with labelled F1 from CoNLL 2009 evaluation script, other columns are evaluated with span F1 score same as it was used for NER evaluation. For more information see [the paper](https://arxiv.org/abs/2103.13031).
</div>
### Named Entity Recognition
| | mBERT | Pavlov | Albert-random | Czert-A | Czert-B |
|:-----------|:---------------|:---------------|:---------------|:---------------|:---------------|
| CNEC | **86.225β
Β±β
0.208** | **86.565β
Β±β
0.198** | 34.635β
Β±β
0.343 | 72.945β
Β±β
0.227 | 86.274β
Β±β
0.116 |
| BSNLP 2019 | 84.006β
Β±β
1.248 | **86.699β
Β±β
0.370** | 19.773β
Β±β
0.938 | 48.859β
Β±β
0.605 | **86.729 Β± 0.344** |
Comparison of f1 score achieved using pre-trained CZERT-A, CZERT-B, mBERT, Pavlov and randomly initialised Albert on named entity recognition task. For more information see [the paper](https://arxiv.org/abs/2103.13031).
## Licence
This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License. http://creativecommons.org/licenses/by-nc-sa/4.0/
## How should I cite CZERT?
For now, please cite [the Arxiv paper](https://arxiv.org/abs/2103.13031):
```
@article{sido2021czert,
title={Czert -- Czech BERT-like Model for Language Representation},
author={Jakub Sido and OndΕej PraΕΎΓ‘k and Pavel PΕibΓ‘Ε and Jan PaΕ‘ek and Michal SejΓ‘k and Miloslav KonopΓk},
year={2021},
eprint={2103.13031},
archivePrefix={arXiv},
primaryClass={cs.CL},
journal={arXiv preprint arXiv:2103.13031},
}
```
|