We present a simple yet powerful data augmentation method for boosting Neural Machine Translation (NMT) performance by leveraging information retrieved from a Translation Memory (TM). We propose and test two methods for augmenting NMT training data with fuzzy TM matches. Tests on the DGT-TM data set for two language pairs show consistent and substantial improvements over a range of baseline systems. The results suggest that this method is promising for any translation environment in which a sizeable TM is available and a certain amount of repetition across translations is to be expected, especially considering its ease of implementation.

We identify a number of aspects that can boost the performance of Neural Fuzzy Repair (NFR), an easy-to-implement method to integrate translation memory matches and neural machine translation (NMT). We explore various ways of maximising the added value of retrieved matches within the NFR paradigm for eight language combinations, using Transformer NMT systems. In particular, we test the impact of different fuzzy matching techniques, sub-word-level segmentation methods and alignment-based features on overall translation quality. Furthermore, we propose a fuzzy match combination technique that aims to maximise the coverage of source words. This is supplemented with an analysis of how translation quality is affected by input sentence length and fuzzy match score. The results show that applying a combination of the tested modifications leads to a significant increase in estimated translation quality over all baselines for all language combinations.

This paper studies the impact of machine translation (MT) on the translation workflow at the Directorate-General for Translation (DGT), focusing on two language pairs and two MT paradigms: English-into-French with statistical MT and English-into-Finnish with neural MT. We collected data from 20 professional translators at DGT while they carried out real translation tasks in normal working conditions. The participants enabled/disabled MT for half of the segments in each document. They filled in a survey at the end of the logging period. We measured the productivity gains (or losses) resulting from the use of MT and examined the relationship between technical effort and temporal effort. The results show that while the usage of MT leads to productivity gains on average, this is not the case for all translators. Moreover, the two technical effort indicators used in this study show weak correlations with post-editing time. The translators’ perception of their speed gains was more or less in line with the actual results. Reduction of typing effort is the most frequently mentioned reason why participants preferred working with MT, but also the psychological benefits of not having to start from scratch were often mentioned.

This paper describes the submission of the UGENT-LT3 SCATE system to the WMT15 Shared Task on Quality Estimation (QE), viz. English-Spanish word and sentence-level QE. We conceived QE as a supervised Machine Learning (ML) problem and designed additional features and combined these with the baseline feature set to estimate quality. The sentence-level QE system re-uses the word level predictions of the word-level QE system. We experimented with different learning methods and observe improvements over the baseline system for wordlevel QE with the use of the new features and by combining learning methods into ensembles. For sentence-level QE we show that using a single feature based on word-level predictions can perform better than the baseline system and using this in combination with additional features led to further improvements in performance.

We propose three linguistically motivated metrics to quantify syntactic equivalence between a source sentence and its translation. Syntactically Aware Cross (SACr) measures the degree of word group reordering by creating syntactically motivated groups of words that are aligned. Secondly, an intuitive approach is to compare the linguistic labels of the word-aligned source and target tokens. Finally, on a deeper linguistic level, Aligned Syntactic Tree Edit Distance (ASTrED) compares the dependency structure of both sentences. To be able to compare source and target dependency labels we make use of Universal Dependencies (UD). We provide an analysis of our metrics by comparing them with translation process data in mixed models. Even though our examples and analysis focus on English as the source language and Dutch as the target language, the proposed metrics can be applied to any language for which UD models are attainable. An open-source implementation is made available.