RobBERT

RobBERT is a Dutch state-of-the-art BERT-based language model based on RoBERTa.

Read more on our blog post or on the paper.

Getting started

RobBERT can easily be used in two different ways, namely either using Fairseq RoBERTa code or using HuggingFace Transformers

Using Huggingface Transformers

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.

Using Fairseq

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.

Performance Evaluation Results

All experiments are described in more detail in our paper.

Sentiment analysis

Predicting whether a review is positive or negative using the Dutch Book Reviews Dataset.

Die/Dat (coreference resolution)

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.

Finetuning on whole dataset

Finetuning on 10K examples

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.

Using zero-shot word masking task

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.

Part-of-Speech Tagging.

Using the Lassy UD dataset.

Interestingly, we found that when dealing with small data sets, RobBERT v2 significantly outperforms other models.

Named Entity Recognition

Using the CoNLL 2002 evaluation script.

Training procedure

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.

Limitations and bias

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.

Replicating the paper experiments

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 pipenv in your terminal) by running pipenv install.

Classification

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.

Sentiment analysis using the Dutch Book Review Dataset

• Download the Dutch book review dataset from https://github.com/benjaminvdb/110kDBRD, and save it to data/raw/110kDBRD
• Run src/preprocess_dbrd.py to 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.ipynb to finetune the model.

Predicting the Dutch pronouns die and dat

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._

Credits and citation

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.

Even though MIT doesn't require it, we would like to ask if you could nevertheless cite our paper if it helped you!

@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"