README.md

metadata

language: nl
thumbnail: https://github.com/iPieter/RobBERT/raw/master/res/robbert_logo.png
tags:
  - Dutch
  - RoBERTa
  - RobBERT
license: mit
datasets:
  - oscar
  - Shuffled Dutch section of the OSCAR corpus (https://oscar-corpus.com/)

RobBERT

Model description

RobBERT v2 is a Dutch state-of-the-art RoBERTa-based language model.

More detailled information can be found in the RobBERT paper.

How to use

from transformers import RobertaTokenizer, RobertaForSequenceClassification
tokenizer = RobertaTokenizer.from_pretrained("pdelobelle/robbert-v2-dutch-base")
model = RobertaForSequenceClassification.from_pretrained("pdelobelle/robbert-v2-dutch-base")

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.

Model	Accuracy [%]
ULMFiT	93.8
BERTje	93.0
RobBERT v2	95.1

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

Model	Accuracy [%]	F1 [%]
Baseline (LSTM)		75.03
mBERT	98.285	98.033
BERTje	98.268	98.014
RobBERT v2	99.232	99.121

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.

Model	Accuracy [%]	F1 [%]
mBERT	92.157	90.898
BERTje	93.096	91.279
RobBERT v2	97.816	97.514

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.

Model	Accuracy [%]
ZeroR	66.70
mBERT	90.21
BERTje	94.94
RobBERT v2	98.75

Part-of-Speech Tagging.

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.

Named Entity Recognition

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

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.

BibTeX entry and citation info

@misc{delobelle2020robbert,
    title={RobBERT: a Dutch RoBERTa-based Language Model},
    author={Pieter Delobelle and Thomas Winters and Bettina Berendt},
    year={2020},
    eprint={2001.06286},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}