Systran (France)

Non-autoregressive machine translation (NAT) has recently made great progress. However, most works to date have focused on standard translation tasks, even though some edit-based NAT models, such as the Levenshtein Transformer (LevT), seem well suited to translate with a Translation Memory (TM). This is the scenario considered here. We first analyze the vanilla LevT model and explain why it does not do well in this setting. We then propose a new variant, TM-LevT, and show how to effectively train this model. By modifying the data presentation and introducing an extra deletion operation, we obtain performance that are on par with an autoregressive approach, while reducing the decoding load. We also show that incorporating TMs during training dispenses to use knowledge distillation, a well-known trick used to mitigate the multimodality issue.

When building machine translation systems, one often needs to make the best out of heterogeneous sets of parallel data in training, and to robustly handle inputs from un-expected domains in testing. This multi-domain scenario has attracted a lot of recent work, that fall under the general umbrella of transfer learning. In this study, we revisit multi-domain machine translation, with the aim to formulate the motivations for developing such systems and the associated expectations with respect to performance. Our experiments with a large sample of multi-domain systems show that most of these expectations are hardly met and suggest that further work is needed to better analyze the current behaviour of multi-domain systems and to make them fully hold their promises.

Abstract The present study examines adults’ processing of the Turkish Aorist, an irregular pattern the acquisition of which is already shown to pose a huge challenge to children up until the age of 8. We investigate how adults produce the Aorist when presented with non-existent monosyllabic roots and to what extent the outcome reflects the application of a rule or the effect of analogy on stored exemplars. Experimental results, associated with a thorough type/token frequency analysis of the aorist in the BOUN corpus, suggest that a rule-based generalization on the basis of the type count of the -Ar affix is at work for over 85% of productions. Similarity-based generalizations, however, though quite few in number appear to be heavily conditioned by the matching last consonant, the matching first and last consonants, overlapping trigram and quadrigram sequences, but not rhyme in Turkish.

Domain adaptation is an old and vexing problem for machine translation systems.The most common and successful approach to supervised adaptation is to fine-tune a baseline system with in-domain parallel data.Standard fine-tuning however modifies all the network parameters, which makes this approach computationally costly and prone to overfitting.A recent, lightweight approach, instead augments a baseline model with supplementary (small) adapter layers, keeping the rest of the model unchanged.This has the additional merit to leave the baseline model intact and adaptable to multiple domains.In this paper, we conduct a thorough analysis of the adapter model in the context of a multidomain machine translation task.We contrast multiple implementations of this idea using two language pairs.Our main conclusions are that residual adapters provide a fast and cheap method for supervised multi-domain adaptation; our two variants prove as effective as the original adapter model and open perspective to also make adapted models more robust to label domain errors.

This work is inspired by a typical machine translation industry scenario in which translators make use of in-domain data for facilitating translation of similar or repeating sentences. We introduce a generic framework applied at inference in which a subset of segment pairs are first extracted from training data according to their similarity to the input sentences. These segments are then used to dynamically update the parameters of a generic NMT network, thus performing a lexical micro-adaptation. Our approach demonstrates strong adaptation performance to new and existing datasets including pseudo in-domain data. We evaluate our approach on a heterogeneous English-French training dataset showing accuracy gains on all evaluated domains when compared to strong adaptation baselines.

This paper presents an approach combining lexico-semantic resources and distributed representations of words applied to the evaluation in machine translation (MT). This study is made through the enrichment of a well-known MT evaluation metric: METEOR. This metric enables an approximate match (synonymy or morphological similarity) between an automatic and a reference translation. Our experiments are made in the framework of the Metrics task of WMT 2014. We show that distributed representations are a good alternative to lexico-semantic resources for MT evaluation and they can even bring interesting additional information. The augmented versions of METEOR, using vector representations, are made available on our Github page.

This paper describes an initial version of a general purpose French/English statistical machine translation system.The main features of this system are the open-source Moses decoder, the integration of a bilingual dictionary and a continuous space target language model.We analyze the performance of this system on the test data of the WMT'08 evaluation.

This paper describes SYSTRAN submissions for the shared task of the third Workshop on Statistical Machine Translation at ACL. Our main contribution consists in a French-English statistical model trained without the use of any human-translated parallel corpus. In substitution, we translated a monolingual corpus with SYSTRAN rule-based translation engine to produce the parallel corpus. The results are provided herein, along with a measure of error analysis.

Latency is an explicit and possibly insidious issue associated with distributed interactive simulations (DISs), whether they are legacy simulations or simulations based on the DIS protocol or on the HLA (High Level Architecture) and RTI (Run-Time Infrastructure) standards. The implementation details of individual simulations, whether located entirely in one room or geographically distributed across a continent, will directly impact the latencies between various elements of the simulation. At SYSTRAN Corp., we have developed a Windows NT-based system, called SNAPpy (Simulation Network Analysis Project), that seeks to provide tools to both measure and reduce the latency associated with DISs. SNAPpy includes two major sections: an NIU (network interface unit) that is multi-threaded and built on an SMP (symmetric multiprocessor) and Windows NT platform, and the SNAP latency measurement section. This paper reviews the architecture of SNAPpy, going into specific details about its make-up and performance. Finally, SNAPpy's potential as a translator of legacy simulations into the HLA/RTI domain is described, and we introduce the SNAP-Lite Latency Measurement System, a new SYSTRAN product.