RobBERT is the state-of-the-art Dutch BERT model. It is a large pre-trained general Dutch language model that can be fine-tuned on a given dataset to perform any text classification, regression or token-tagging task. As such, it has been successfully used by many researchers and practitioners for achieving state-of-the-art performance for a wide range of Dutch natural language processing tasks, including:
- Emotion detection
- Sentiment analysis (book reviews, news articles*)
- Coreference resolution
- Named entity recognition (CoNLL, job titles*, SoNaR)
- Part-of-speech tagging (Small UD Lassy, CGN)
- Zero-shot word prediction
- Humor detection
- Cyberbullying detection
- Correcting dt-spelling mistakes*
and also achieved outstanding, near-sota results for:
* Note that several evaluations use RobBERT-v1, and that the second and improved RobBERT-v2 outperforms this first model on everything we tested
(Also note that this list is not exhaustive. If you used RobBERT for your application, we are happy to know about it! Send us a mail, or add it yourself to this list by sending a pull request with the edit!)
To use the RobBERT model using HuggingFace transformers, use the name pdelobelle/robbert-v2-dutch-base.
More in-depth information about RobBERT can be found in our blog post and in our paper.
- How To Use
- Technical Details From The Paper
- Pre-Training Procedure Details
- Investigating Limitations and Bias
- How to Replicate Our Paper Experiments
- Name Origin of RobBERT
- Credits and citation
RobBERT uses the RoBERTa architecture and pre-training but with a Dutch tokenizer and training data. RoBERTa is the robustly optimized English BERT model, making it even more powerful than the original BERT model. Given this same architecture, RobBERT can easily be finetuned and inferenced using code to finetune RoBERTa models and most code used for BERT models, e.g. as provided by HuggingFace Transformers library.
RobBERT can easily be used in two different ways, namely either using Fairseq RoBERTa code or using HuggingFace Transformers
By default, RobBERT has the masked language model head used in training. This can be used as a zero-shot way to fill masks in sentences. It can be tested out for free on RobBERT's Hosted infererence API of Huggingface. You can also create a new prediction head for your own task by using any of HuggingFace's RoBERTa-runners, their fine-tuning notebooks by changing the model name to pdelobelle/robbert-v2-dutch-base, or use the original fairseq RoBERTa training regimes.
You can easily download RobBERT v2 using 🤗 Transformers. Use the following code to download the base model and finetune it yourself, or use one of our finetuned models (documented on our project site).
from transformers import RobertaTokenizer, RobertaForSequenceClassification
tokenizer = RobertaTokenizer.from_pretrained("pdelobelle/robbert-v2-dutch-base")
model = RobertaForSequenceClassification.from_pretrained("pdelobelle/robbert-v2-dutch-base")Starting with transformers v2.4.0 (or installing from source), you can use AutoTokenizer and AutoModel.
You can then use most of HuggingFace's BERT-based notebooks for finetuning RobBERT on your type of Dutch language dataset.
Alternatively, you can also use RobBERT using the RoBERTa architecture code.
You can download RobBERT v2's Fairseq model here: (RobBERT-base, 1.5 GB).
Using RobBERT's model.pt, this method allows you to use all other functionalities of RoBERTa.
All experiments are described in more detail in our paper, with the code in our GitHub repository.
Predicting whether a review is positive or negative using the Dutch Book Reviews Dataset.
| Model | Accuracy [%] |
|---|---|
| ULMFiT | 93.8 |
| BERTje | 93.0 |
| RobBERT v2 | 95.1 |
We measured how well the models are able to do coreference resolution by predicting whether "die" or "dat" should be filled into a sentence. For this, we used the EuroParl corpus.
| Model | Accuracy [%] | F1 [%] |
|---|---|---|
| Baseline (LSTM) | 75.03 | |
| mBERT | 98.285 | 98.033 |
| BERTje | 98.268 | 98.014 |
| RobBERT v2 | 99.232 | 99.121 |
We also measured the performance using only 10K training examples. This experiment clearly illustrates that RobBERT outperforms other models when there is little data available.
| Model | Accuracy [%] | F1 [%] |
|---|---|---|
| mBERT | 92.157 | 90.898 |
| BERTje | 93.096 | 91.279 |
| RobBERT v2 | 97.816 | 97.514 |
Since BERT models are pre-trained using the word masking task, we can use this to predict whether "die" or "dat" is more likely. This experiment shows that RobBERT has internalised more information about Dutch than other models.
| Model | Accuracy [%] |
|---|---|
| ZeroR | 66.70 |
| mBERT | 90.21 |
| BERTje | 94.94 |
| RobBERT v2 | 98.75 |
Using the Lassy UD dataset.
| Model | Accuracy [%] |
|---|---|
| Frog | 91.7 |
| mBERT | 96.5 |
| BERTje | 96.3 |
| RobBERT v2 | 96.4 |
Interestingly, we found that when dealing with small data sets, RobBERT v2 significantly outperforms other models.
Using the CoNLL 2002 evaluation script.
| Model | Accuracy [%] |
|---|---|
| Frog | 57.31 |
| mBERT | 90.94 |
| BERT-NL | 89.7 |
| BERTje | 88.3 |
| RobBERT v2 | 89.08 |
We pre-trained RobBERT using the RoBERTa training regime. We pre-trained our model on the Dutch section of the OSCAR corpus, a large multilingual corpus which was obtained by language classification in the Common Crawl corpus. This Dutch corpus is 39GB large, with 6.6 billion words spread over 126 million lines of text, where each line could contain multiple sentences, thus using more data than concurrently developed Dutch BERT models.
RobBERT shares its architecture with RoBERTa's base model, which itself is a replication and improvement over BERT. Like BERT, it's architecture consists of 12 self-attention layers with 12 heads with 117M trainable parameters. One difference with the original BERT model is due to the different pre-training task specified by RoBERTa, using only the MLM task and not the NSP task. During pre-training, it thus only predicts which words are masked in certain positions of given sentences. The training process uses the Adam optimizer with polynomial decay of the learning rate l_r=10^-6 and a ramp-up period of 1000 iterations, with hyperparameters beta_1=0.9 and RoBERTa's default beta_2=0.98. Additionally, a weight decay of 0.1 and a small dropout of 0.1 helps prevent the model from overfitting.
RobBERT was trained on a computing cluster with 4 Nvidia P100 GPUs per node, where the number of nodes was dynamically adjusted while keeping a fixed batch size of 8192 sentences. At most 20 nodes were used (i.e. 80 GPUs), and the median was 5 nodes. By using gradient accumulation, the batch size could be set independently of the number of GPUs available, in order to maximally utilize the cluster. Using the Fairseq library, the model trained for two epochs, which equals over 16k batches in total, which took about three days on the computing cluster. In between training jobs on the computing cluster, 2 Nvidia 1080 Ti's also covered some parameter updates for RobBERT v2.
In the RobBERT paper, we also investigated potential sources of bias in RobBERT.
We found that the zeroshot model estimates the probability of hij (he) to be higher than zij (she) for most occupations in bleached template sentences, regardless of their actual job gender ratio in reality.
By augmenting the DBRB Dutch Book sentiment analysis dataset with the stated gender of the author of the review, we found that highly positive reviews written by women were generally more accurately detected by RobBERT as being positive than those written by men.
You can replicate the experiments done in our paper by following the following steps. You can install the required dependencies either the requirements.txt or pipenv:
- Installing the dependencies from the requirements.txt file using
pip install -r requirements.txt - OR install using Pipenv (install by running
pip install pipenvin your terminal) by runningpipenv install.
In this section we describe how to use the scripts we provide to fine-tune models, which should be general enough to reuse for other desired textual classification tasks.
- Download the Dutch book review dataset from https://github.com/benjaminvdb/DBRD, and save it to
data/raw/DBRD - Run
src/preprocess_dbrd.pyto prepare the dataset. - To not be blind during training, we recommend to keep aside a small evaluation set from the training set. For this run
src/split_dbrd_training.sh. - Follow the notebook
notebooks/finetune_dbrd.ipynbto finetune the model.
We fine-tune our model on the Dutch Europarl corpus. You can download it first with:
cd data\raw\europarl\
wget -N 'http://www.statmt.org/europarl/v7/nl-en.tgz'
tar zxvf nl-en.tgz
As a sanity check, now you should have the following files in your data/raw/europarl folder:
europarl-v7.nl-en.en
europarl-v7.nl-en.nl
nl-en.tgz
Then you can run the preprocessing with the following script, which fill first process the Europarl corpus to remove sentences without any die or dat.
Afterwards, it will flip the pronoun and join both sentences together with a <sep> token.
python src/preprocess_diedat.py
. src/preprocess_diedat.sh
note: You can monitor the progress of the first preprocessing step with watch -n 2 wc -l data/europarl-v7.nl-en.nl.sentences. This will take a while, but it's certainly not needed to use all inputs. This is after all why you want to use a pre-trained language model. You can terminate the python script at any time and the second step will only use those._
Most BERT-like models have the word BERT in their name (e.g. RoBERTa, ALBERT, CamemBERT, and many, many others). As such, we queried our newly trained model using its masked language model to name itself <mask>bert using all kinds of prompts, and it consistently called itself RobBERT. We thought it was really quite fitting, given that RobBERT is a very Dutch name (and thus clearly a Dutch language model), and additionally has a high similarity to its root architecture, namely RoBERTa.
Since "rob" is a Dutch words to denote a seal, we decided to draw a seal and dress it up like Bert from Sesame Street for the RobBERT logo.
This project is created by Pieter Delobelle, Thomas Winters and Bettina Berendt.
We are grateful to Liesbeth Allein, for her work on die-dat disambiguation, Huggingface for their transformer package, Facebook for their Fairseq package and all other people whose work we could use.
We release our models and this code under MIT.
If you would like to cite our paper or model, you can use the following BibTeX code:
@inproceedings{delobelle2020robbert,
title = "{R}ob{BERT}: a {D}utch {R}o{BERT}a-based {L}anguage {M}odel",
author = "Delobelle, Pieter and
Winters, Thomas and
Berendt, Bettina",
booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2020",
month = nov,
year = "2020",
address = "Online",
publisher = "Association for Computational Linguistics",
url = "https://www.aclweb.org/anthology/2020.findings-emnlp.292",
doi = "10.18653/v1/2020.findings-emnlp.292",
pages = "3255--3265"
}