The MLLP-UPV German-English Machine Translation System for WMT18

Abstract

This paper describes the statistical machine translation system built by the MLLP research group of Universitat Politècnica de València for the German→English news translation shared task of the EMNLP 2018 Third Conference on Machine Translation (WMT18). We used an ensemble of Transformer architecture-based neural machine translation systems. To train our system under "constrained" conditions, we filtered the provided parallel data with a scoring technique using character-based language models, and we added parallel data based on synthetic source sentences generated from the provided monolingual corpora.

VivesDebate: A New Annotated Multilingual Corpus of Argumentation in a Debate Tournament

“…During this competition, 16 different teams from universities belonging to the autonomous regions of Valencia, Catalonia and the Balearic islands debated in the Catalan language on the topic "Should surrogacy be legalised?" In addition to the original language of the annotated data, automatic translations to Spanish and English languages using the MLLP machine translation toolkit [30,31] have also been included. The results and the evaluation of the debates were directly retrieved from the organisation, but were post-processed by us in order to focus on the argumentative aspects of the debates and to preserve the anonymity of the jury and the participant teams.…”

The MLLP-UPV Supervised Machine Translation Systems for WMT19 News Translation Task

“…This is highlighted by the fact that a majority of participating systems in the WMT18 News Translation Task apply filtering techniques to ParaCrawl. Additionally, the experiments carried out for our 2018 submission (Iranzo-Sánchez et al, 2018) show that using a noisy corpus such as ParaCrawl without filtering can result in a worse performance compared with a baseline system that simply excludes the noisy corpus from the training data.…”

“…• LM-based filtering (Iranzo-Sánchez et al, 2018): This approach uses language models for estimating the quality of a sentence pair, under the assumption that a low-perplexity sentence is more likely to be an adequate sentence for training. Using in-domain data, we train one language model for each language, and then use them to score the corresponding side of the sentence pair, giving us perplexity scores (s, t).…”

“…• German → English: We have used 20M sentences from our WMT18 submission (Iranzo-Sánchez et al, 2018), and an additional 24M sentences generated using a system with the same configuration as WMT18, but trained with 3 GPUs instead of 1. The monolingual sentences were randomly sampled from News Crawl 2017.…”

The Universitat d’Alacant Submissions to the English-to-Kazakh News Translation Task at WMT 2019

“…The same filtering was applied to the monolingual Kazakh Common Crawl corpus. In addition, inspired by Iranzo-Sánchez et al (2018), we ranked its sentences by perplexity computed by a character-based 7-gram language model and discarded the half of the corpus with the highest perplexity. The language model was trained 2 on the high-quality Kazakh monolingual News Commentary corpus.…”