Traditionalhand-crafted linguisticallyinformed features have often been used for distinguishing between translated and original non-translated texts. By contrast, to date, neural architectures without manual feature engineering have been less explored for this task. In this work, we (i) compare the traditional feature-engineering-based approach to the feature-learning-based one and (ii) analyse the neural architectures in order to investigate how well the hand-crafted features explain the variance in the neural models' predictions. We use pre-trained neural word embeddings, as well as several end-to-end neural architectures in both monolingual and multilingual settings and compare them to feature-engineering-based SVM classifiers. We show that (i) neural architectures outperform other approaches by more than 20 accuracy points, with the BERT-based model performing the best in both the monolingual and multilingual settings; (ii) while many individual hand-crafted translationese features correlate with neural model predictions, feature importance analysis shows that the most important features for neural and classical architectures differ; and (iii) our multilingual experiments provide empirical evidence for translationese universals across languages.

This paper presents the Automatic Postediting (APE) systems submitted by the DFKI-MLT group to the WMT'18 APE shared task. Three monolingual neural sequenceto-sequence APE systems were trained using target-language data only: one using an attentional recurrent neural network architecture and two using the attention-only (transformer) architecture. The training data was composed of machine translated (MT) output used as source to the APE model aligned with their manually post-edited version or reference translation as target. We made use of the provided training sets only and trained APE models applicable to phrase-based and neural MT outputs. Results show better performances reached by the attention-only model over the recurrent one, significant improvement over the baseline when post-editing phrase-based MT output but degradation when applied to neural MT output.

In learning-based functionality stealing, the attacker is trying to build a local model based on the victim's outputs. The attacker has to make choices regarding the local model's architecture, optimization method and, specifically for NLP models, subword vocabulary, such as BPE. On the machine translation task, we explore (1) whether the choice of the vocabulary plays a role in model stealing scenarios and (2) if it is possible to extract the victim's vocabulary. We find that the vocabulary itself does not have a large effect on the local model's performance. Given gray-box model access, it is possible to collect the victim's vocabulary by collecting the outputs (detokenized subwords on the output). The results of the minimum effect of vocabulary choice are important more broadly for black-box knowledge distillation.