The DBRD (pronounced dee-bird) dataset contains over 110k book reviews along with associated binary sentiment polarity labels. It is greatly influenced by the Large Movie Review Dataset and intended as a benchmark for sentiment classification in Dutch. The scripts that were used to scrape the reviews from Hebban can be found in the DBRD GitHub repository.
The dataset is ~79MB compressed and can be downloaded from here:
A language model trained with FastAI on Dutch Wikipedia can be downloaded from here:
Dutch language model trained on Wikipedia
The dataset includes three folders with data: test
(test split), train
(train split) and unsup
(remaining reviews).
Each review is assigned a unique identifier and can be deduced from the filename, as well as the rating: [ID]_[RATING].txt
. This is different from the Large Movie Review Dataset, where each file in a directory has a unique ID, but IDs are reused between folders.
The urls.txt
file contains on line L
the URL of the book review on Hebban for the book review with that ID, i.e., the URL of the book review in 48091_5.txt
can be found on line 48091 of urls.txt
. It cannot be guaranteed that these pages still exist.
.
├── README.md // the file you're reading
├── test // balanced 10% test split
│ ├── neg
│ └── pos:
├── train: // balanced 90% train split
│ ├── neg
│ └── pos
└── unsup // unbalanced positive and neutral
└── urls.txt // urls to reviews on Hebban
#all: 118516 (= #supervised + #unsupervised)
#supervised: 22252 (= #training + #testing)
#unsupervised: 96264
#training: 20028
#testing: 2224
Distribution of labels positive/negative/neutral
in rounded percentages.
training: 50/50/ 0
test: 50/50/ 0
unsup: 72/ 0/28
Train and test sets are balanced and contain no neutral reviews (for which rating==3
).
Since scraping Hebban induces a load on their servers, it's best to download the prepared dataset instead. This also makes sure your results can be compared to those of others. The scripts and instructions should be used mostly as a starting point for building a scraper for another website.
I'm making using of Selenium for automating user actions such as clicks. This library requires a browser driver that provides the rendering backend. I've made use of ChromeDriver.
If you're on macOS and you have Homebrew installed, you can install ChromeDriver by running:
brew install chromedriver
You can download ChromeDriver from the official download page.
The scripts are written for Python 3. To install the Python dependencies, run:
pip3 install -r ./requirements.txt
Two scripts are provided that can be run in sequence. You can also run run.sh
to run all scripts with defaults.
The first step is to gather all review URLs from Hebban. Run gather_urls.py
to fetch them and save them to a text file.
Usage: gather_urls.py [OPTIONS] OUTFILE
This script gathers review urls from Hebban and writes them to OUTFILE.
Options:
--offset INTEGER Review offset.
--step INTEGER Number of review urls to fetch per request.
--help Show this message and exit.
The second step is to scrape the URLs for review data. Run scrape_reviews.py
to iterate over the review URLs and save the scraped data to a JSON file.
Usage: scrape_reviews.py [OPTIONS] INFILE OUTFILE
Iterate over review urls in INFILE text file, scrape review data and
output to OUTFILE.
Options:
--encoding TEXT Output file encoding.
--indent INTEGER Output JSON file with scraped data.
--help Show this message and exit.
The third and final step is to prepare the dataset using the scraped reviews. By default, we limit the number of reviews to 110k, filter out some reviews and prepare train and test sets of 0.9 and 0.1 the total amount, respectively.
Usage: post_process.py [OPTIONS] INFILE OUTDIR
Options:
--encoding TEXT Input file encoding
--keep-incorrect-date TEXT Whether to keep reviews with invalid dates.
--sort TEXT Whether to sort reviews by date.
--maximum INTEGER Maximum number of reviews in output
--valid-size-fraction FLOAT Fraction of total to set aside as validation.
--shuffle TEXT Shuffle data before saving.
--help Show this message and exit.
v3: Changed name of the dataset from 110kDBRD to DBRD. The dataset itself remains unchanged.
v2: Removed advertisements from reviews and increased dataset size to 118,516.
v1: Initial release
Please use the following citation when making use of this dataset in your work.
@article{DBLP:journals/corr/abs-1910-00896,
author = {Benjamin van der Burgh and
Suzan Verberne},
title = {The merits of Universal Language Model Fine-tuning for Small Datasets
- a case with Dutch book reviews},
journal = {CoRR},
volume = {abs/1910.00896},
year = {2019},
url = {http://arxiv.org/abs/1910.00896},
archivePrefix = {arXiv},
eprint = {1910.00896},
timestamp = {Fri, 04 Oct 2019 12:28:06 +0200},
biburl = {https://dblp.org/rec/journals/corr/abs-1910-00896.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}
This dataset was created for testing out the ULMFiT (by Jeremy Howard and Sebastian Ruder) deep learning algorithm for text classification. It is implemented in the FastAI Python library that has taught me a lot. I'd also like to thank Timo Block for making his 10kGNAD dataset publicly available and giving me a starting point for this dataset. The dataset structure based on the Large Movie Review Dataset by Andrew L. Maas et al. Thanks to Andreas van Cranenburg for pointing out a problem with the dataset.
And of course I'd like to thank all the reviewers on Hebban for having taken the time to write all these reviews. You've made both book enthousiast and NLP researchers very happy :)
All code in this repository is licensed under a MIT License.
The dataset is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.