We aim to shed light on the strengths and weaknesses of the newly introduced neural machine translation paradigm. To that end, we conduct a multifaceted evaluation in which we compare outputs produced by state-of-the-art neural machine translation and phrase-based machine translation systems for 9 language directions across a number of dimensions. Specifically, we measure the similarity of the outputs, their fluency and amount of reordering, the effect of sentence length and performance across different error categories. We find out that translations produced by neural machine translation systems are considerably different, more fluent and more accurate in terms of word order compared to those produced by phrase-based systems. Neural machine translation systems are also more accurate at producing inflected forms, but they perform poorly when translating very long sentences.

This paper presents a quantitative fine-grained manual evaluation approach to comparing the performance of different machine translation (MT) systems. We build upon the well-established Multidimensional Quality Metrics (MQM) error taxonomy and implement a novel method that assesses whether the differences in performance for MQM error types between different MT systems are statistically significant. We conduct a case study for English-toCroatian, a language direction that involves translating into a morphologically rich language, for which we compare three MT systems belonging to different paradigms: pure phrase-based, factored phrase-based and neural. First, we design an MQM-compliant error taxonomy tailored to the relevant linguistic phenomena of Slavic languages, which made the annotation process feasible and accurate. Errors in MT outputs were then annotated by two annotators following this taxonomy. Subsequently, we carried out a statistical analysis which showed that the best-performing system (neural) reduces the errors produced by the worst system (pure phrase-based) by more than half (54%). Moreover, we conducted an additional analysis of agreement errors in which we distinguished between short (phrase-level) and long distance (sentence-level) errors. We discovered that phrase-based MT approaches are of limited use for long distance agreement phenomena, for which neural MT was found to be especially effective.

We compare three approaches to statistical machine translation (pure phrase-based, factored phrase-based and neural) by performing a fine-grained manual evaluation via error annotation of the systems' outputs. The error types in our annotation are compliant with the multidimensional quality metrics (MQM), and the annotation is performed by two annotators. Inter-annotator agreement is high for such a task, and results show that the best performing system (neural) reduces the errors produced by the worst system (phrase-based) by 54%.

Statistical and rule-based methods are complementary approaches to machine translation (MT) that have different strengths and weaknesses. This complementarity has, over the last few years, resulted in the consolidation of a growing interest in hybrid systems that combine both data-driven and linguistic approaches. In this paper we address the situation in which the amount of bilingual resources that is available for a particular language pair is not sufficiently large to train a competitive statistical MT system, but the cost and slow development cycles of rule-based MT systems cannot be afforded either. In this context, we formalise a new method that uses scarce parallel corpora to automatically infer a set of shallowtransfer rules to be integrated into a rule-based MT system, thus avoiding the need for human experts to handcraft these rules.Our work is based on the alignment template approach to phrase-based statistical MT, but the definition of the alignment template is extended to encompass different generalisation levels. It is also greatly inspired by the work of Sánchez-Martínez and Forcada published in 2009 (Journal of Artificial Intelligence Research 34) in which alignment templates were also considered for shallow-transfer rule inference. However, our approach overcomes many relevant limitations of that work, principally those related to the inability to find the correct generalisation level for the alignment templates, and to select the subset of alignment templates that ensures an adequate segmentation of the input sentences by the rules eventually obtained. Unlike previous approaches in literature, our formalism does not require linguistic knowledge about the languages involved in the translation. Moreover, it is the first time that conflicts between rules are resolved by choosing the most appropriate ones according to a global minimisation function rather than proceeding in a pairwise greedy fashion.Experiments conducted using five different language pairs with the free/open-source rule-based MT platform Apertium show that translation quality significantly improves when compared to the method proposed by Sánchez-Martínez and Forcada (2009), and is close to that obtained using handcrafted rules. For some language pairs, our approach is even able to outperform them. Moreover, the resulting number of rules is considerably smaller, which eases human revision and maintenance.

In the context of neural machine translation, data augmentation (DA) techniques may be used for generating additional training samples when the available parallel data are scarce. Many DA approaches aim at expanding the support of the empirical data distribution by generating new sentence pairs that contain infrequent words, thus making it closer to the true data distribution of parallel sentences. In this paper, we propose to follow a completely different approach and present a multi-task DA approach in which we generate new sentence pairs with transformations, such as reversing the order of the target sentence, which produce unfluent target sentences. During training, these augmented sentences are used as auxiliary tasks in a multi-task framework with the aim of providing new contexts where the target prefix is not informative enough to predict the next word. This strengthens the encoder and forces the decoder to pay more attention to the source representations of the encoder. Experiments carried out on six lowresource translation tasks show consistent improvements over the baseline and over DA methods aiming at extending the support of the empirical data distribution. The systems trained with our approach rely more on the source tokens, are more robust against domain shift and suffer less hallucinations.