EN-DE-Bidirectional-Europarl-UdS corpus

<h2><strong>Bidirectional document- and sentence-aligned corpus of Europarl proceedings (EN-DE, DE-EN).</strong></h2> <p>The corpus is described in the paper:</p> <p><strong>Mitigating Translationese with GPT-4: Strategies and Performance</strong></p> <p><em>Maria Kunilovskaya, Koel Dutta Chowdhury, Heike Przybyl, Cristina Espa&ntilde;a-Bonet, Josef van Genabith</em></p> <p><em>in: </em>Proceedings of the 25th Annual conference of the European Association for Machine Translation (EAMT-2024) 24-27 June, Sheffield (UK) Association for Computational Linguistics</p> <p>This version of Europarl-UdS corpus is built from the Europarl proceedings (published up to 10 July 2018) collected using Jose Martinez's pipeline (https://github.com/chozelinek/europarl) in November 2023.</p> <p><br>We include several versions and subsets of this parallel corpus:</p> <ol> <li>The initial input (0_align/xml_translationese/) to the parallel corpus-building pipeline <ul> <li><a href="../api/records/11127626/draft/files/xml_translationese.zip/content" target="_blank" rel="noopener noreferrer">xml_translationese.zip</a></li> </ul> </li> <li>Raw-text-based sentence-aligned documents for both translation directions (columns=['sdoc_id', 'sseg_id', 'sseg', 'tseg', 'hunalign_qua']): <ul> <li><a href="../api/records/11127626/draft/files/deen_wide2018_cap0_score0.3.tsv.gz/content" target="_blank" rel="noopener noreferrer">deen_wide2018_cap0_score0.3.tsv.gz</a></li> <li><a href="../api/records/11127626/draft/files/deen_wide2018_cap0_score0.3.tsv.gz/content" target="_blank" rel="noopener noreferrer">ende_wide2018_cap0_score0.5.tsv.gz</a></li> <li>meta.zip contains four files with XML tags with metadata to each document in the corpus.&nbsp;</li> </ul> </li> <li>The same documents annotated with Stanza (with the conllu-style vertical format for each segment and document wrapped in XML tags containing metadata) <ul> <li>ORG_WR_DE_EN.conllu.xml.gz (original German)</li> <li>ORG_WR_EN_DE.conllu.xml.gz (original English)</li> <li>TR_DE_EN.conllu.xml.gz (translated English)</li> <li>TR_EN_DE.conllu.xml.gz (translated German)</li> </ul> </li> <li>The sentence-level subset of the corpus with extracted morphosyntactic (or lexicogrammatical) features as described in the paper (documents longer than 450 tokens, 1500 documents per translation direction) <ul> <li>seg-450-1500.feats.tsv.gz</li> </ul> </li> <li>A multi-parallel subset of 200 most_translated documents re-written by GPT-4 under various prompting conditions as described in the paper <ul> <li>ratio2.5_de_7aligned_2056segs.tsv</li> <li>ratio2.5_en_7aligned_2109segs.tsv</li> </ul> </li> </ol> <p>Note that the annotation retains the alignment. The processing code is available here https://github.com/SFB1102/b7-b6-prompting-eamt2024/1_parse_extract_feats</p> <p>&nbsp;</p>