Document-level machine translation. Ensuring translational consistency of non-local phenomena

Resumo

In this thesis, we study the automatic translation of documents by taking into account cross-sentence phenomena. This document-level information is typically ignored by most of the standard state-of-the-art Machine Translation (MT) systems, which focus on translating texts processing each of their sentences in isolation. Translating each sentence without looking at its surrounding context can lead to certain types of translation errors, such as inconsistent translations for the same word or for elements in a coreference chain. We introduce methods to attend to document-level phenomena in order to avoid those errors, and thus, reach translations that properly convey the original meaning. Our research starts by identifying the translation errors related to such document-level phenomena that commonly appear in the output of state-of-the-art Statistical Machine Translation (SMT) systems. For two of those errors, namely inconsistent word translations as well as gender and number disagreements among words, we design simple and yet effective post-processing techniques to tackle and correct them. Since these techniques are applied a posteriori, they can access the whole source and target documents, and hence, they are able to perform a global analysis and improve the coherence and consistency of the translation. Nevertheless, since following such a two-pass decoding strategy is not optimal in terms of efficiency, we also focus on introducing the context-awareness during the decoding process itself. To this end, we enhance a document-oriented SMT system with distributional semantic information in the form of bilingual and monolingual word embeddings. In particular, these embeddings are used as Semantic Space Language Models (SSLMs) and as a novel feature function. The goal of the former is to promote word translations that are semantically close to their preceding context, whereas the latter promotes the lexical choice that is closest to its surrounding context, for those words that have varying translations throughout the document. In both cases, the context extends beyond sentence boundaries. Recently, the MT community has transitioned to the neural paradigm. The finalstep of our research proposes an extension of the decoding process for a Neural Machine Translation (NMT) framework, independent of the model architecture, by shallow fusing the information from a neural translation model and the context semantics enclosed in the previously studied SSLMs. The aim of this modification is to introduce the benefits of context information also into the decoding process of NMT systems, as well as to obtain an additional validation for the techniques we explored. The automatic evaluation of our approaches does not reflect significant variations. This is expected since most automatic metrics are neither context- nor semantic-aware and because the phenomena we tackle are rare, leading to few modifications with respect to the baseline translations. On the other hand, manual evaluations demonstrate the positive impact of our approaches since human evaluators tend to prefer the translations produced by our document-aware systems. Therefore, the changes introduced by our enhanced systems are important since they are related to how humans perceive translation quality for long texts.