Datasets:
license:
- cc-by-sa-4.0
language:
- de
multilinguality:
- monolingual
size_categories:
- 10M<n<100M
task_categories:
- sentence-similarity
This dataset card is still a draft version. The dataset has not been uploaded yet.
This is a record of German language paraphrases. These are text pairs that have the same meaning but are expressed in different words. The source of the paraphrases are different parallel German / English text corpora. The English texts were machine translated back into German to obtain the paraphrases.
This dataset can be used for example to train semantic text embeddings. To do this, for example, SentenceTransformers and the MultipleNegativesRankingLoss can be used.
Maintainers
This dataset is open sourced by Philip May
and maintained by the One Conversation
team of Deutsche Telekom AG.
Our pre-processing
Apart from the back translation, we have added more columns (for details see below). We have carried out the following pre-processing and filtering:
- We dropped text pairs where one text was longer than 499 characters.
- In the GlobalVoices v2018q4 texts we have removed the
" · Global Voices"
suffix.
Your post-processing
You probably don't want to use the dataset as it is, but filter it further. This is what the additional columns of the dataset are for. For us it has proven useful to delete the following pairs of sentences:
min_char_len
less than 15jaccard_similarity
greater than 0.3de_token_count
greater than 30en_de_token_count
greater than 30cos_sim
less than 0.85
Columns description
uuid
: a uuid calculated with Pythonuuid.uuid4()
de
: the original German texts from the corpusen_de
: the German texts translated back from Englishcorpus
: the name of the corpusmin_char_len
: the number of characters of the shortest textjaccard_similarity
: the Jaccard similarity coefficient of both sentences - see below for more detailsde_token_count
: number of tokens of thede
text, tokenized with deepset/gbert-largeen_de_token_count
: number of tokens of thede
text, tokenized with deepset/gbert-largecos_sim
: the cosine similarity of both sentences measured with sentence-transformers/paraphrase-multilingual-mpnet-base-v2
Anomalies in the texts
It is noticeable that the OpenSubtitles texts have weird dash prefixes. This looks like this:
- Hast du was draufgetan?
To remove them you could apply this function:
import re
def clean_text(text):
text = re.sub("^[-\s]*", "", text)
text = re.sub("[-\s]*$", "", text)
return text
df["de"] = df["de"].apply(clean_text)
df["en_de"] = df["en_de"].apply(clean_text)
Parallel text corpora used
Corpus name & link | Number of paraphrases |
---|---|
OpenSubtitles | 18,764,810 |
WikiMatrix v1 | 1,569,231 |
Tatoeba v2022-03-03 | 313,105 |
TED2020 v1 | 289,374 |
News-Commentary v16 | 285,722 |
GlobalVoices v2018q4 | 70,547 |
sum | . 21,292,789 |
Back translation
We have made the back translation from English to German with the help of Fairseq.
We used the transformer.wmt19.en-de
model for this purpose:
en2de = torch.hub.load(
"pytorch/fairseq",
"transformer.wmt19.en-de",
checkpoint_file="model1.pt:model2.pt:model3.pt:model4.pt",
tokenizer="moses",
bpe="fastbpe",
)
How the Jaccard similarity was calculated
To calculate the Jaccard similarity coefficient
we are using the SoMaJo tokenizer
to split the texts into tokens.
We then lower()
the tokens so that upper and lower case letters no longer make a difference. Below you can find a code snippet with the details:
from somajo import SoMaJo
LANGUAGE = "de_CMC"
somajo_tokenizer = SoMaJo(LANGUAGE)
def get_token_set(text, somajo_tokenizer):
sentences = somajo_tokenizer.tokenize_text([text])
tokens = [t.text.lower() for sentence in sentences for t in sentence]
token_set = set(tokens)
return token_set
def jaccard_similarity(text1, text2, somajo_tokenizer):
token_set1 = get_token_set(text1, somajo_tokenizer=somajo_tokenizer)
token_set2 = get_token_set(text2, somajo_tokenizer=somajo_tokenizer)
intersection = token_set1.intersection(token_set2)
union = token_set1.union(token_set2)
jaccard_similarity = float(len(intersection)) / len(union)
return jaccard_similarity
Load this dataset with Pandas
If you want to download the csv file and then load it with Pandas you can do it like this:
df = pd.read_csv("train.csv")
Citations & Acknowledgements
OpenSubtitles
- citation: P. Lison and J. Tiedemann, 2016, OpenSubtitles2016: Extracting Large Parallel Corpora from Movie and TV Subtitles. In Proceedings of the 10th International Conference on Language Resources and Evaluation (LREC 2016)
- also see http://www.opensubtitles.org/
- license: no special license has been provided at OPUS for this dataset
WikiMatrix v1
- citation: Holger Schwenk, Vishrav Chaudhary, Shuo Sun, Hongyu Gong and Paco Guzman, WikiMatrix: Mining 135M Parallel Sentences in 1620 Language Pairs from Wikipedia, arXiv, July 11 2019
- license: CC-BY-SA 4.0
Tatoeba v2022-03-03
- citation: J. Tiedemann, 2012, Parallel Data, Tools and Interfaces in OPUS. In Proceedings of the 8th International Conference on Language Resources and Evaluation (LREC 2012)
- license: CC BY 2.0 FR
- copyright: https://tatoeba.org/eng/terms_of_use
TED2020 v1
- citation: Reimers, Nils and Gurevych, Iryna, Making Monolingual Sentence Embeddings Multilingual using Knowledge Distillation, In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, November 2020
- acknowledgements to OPUS for this service
- license: please respect the TED Talks Usage Policy
News-Commentary v16
- citation: J. Tiedemann, 2012, Parallel Data, Tools and Interfaces in OPUS. In Proceedings of the 8th International Conference on Language Resources and Evaluation (LREC 2012)
- license: no special license has been provided at OPUS for this dataset
GlobalVoices v2018q4
- citation: J. Tiedemann, 2012, Parallel Data, Tools and Interfaces in OPUS. In Proceedings of the 8th International Conference on Language Resources and Evaluation (LREC 2012)
- license: no special license has been provided at OPUS for this dataset
Licensing
Copyright (c) 2022 Philip May, Deutsche Telekom AG
This work is licensed under CC-BY-SA 4.0.