Machine Translation (MT) for language pairs with long distance dependencies and word reordering, such as German-English, is prone to producing output that is lexically or syntactically incoherent. Statistical MT (SMT) models used explicit or latent syntax to improve reordering, however failed at capturing other long distance dependencies. This thesis explores how explicit sentence-level syntactic information can improve translation for such complex linguistic phenomena. In particular, we work at the level of the syntactic-semantic interface with representations conveying the predicate-argument structures. These are essential to preserving semantics in translation and SMT systems have long struggled to model them. String-to-tree SMT systems use explicit target syntax to handle long-distance reordering, but make strong independence assumptions which lead to inconsistent lexical choices. To address this, we propose a Selectional Preferences feature which models the semantic affinities between target predicates and their argument fillers using the target dependency relations available in the decoder. We found that our feature is not effective in a string-to-tree system for German-English and that often the conditioning context is wrong because of mistranslated verbs. To improve verb translation, we proposed a Neural Verb Lexicon Model (NVLM) incorporating sentence-level syntactic context from the source which carries relevant semantic information for verb disambiguation. When used as an extra feature for re-ranking the output of a German-English string-to-tree system, the NVLM improved verb translation precision by up to 2.7% and recall by up to 7.4%. While the NVLM improved some aspects of translation, other syntactic and lexical inconsistencies are not being addressed by a linear combination of independent models. In contrast to SMT, neural machine translation (NMT) avoids strong independence assumptions thus generating more fluent translations and capturing some long-distance dependencies. Still, incorporating additional linguistic information can improve translation quality. We proposed a method for tightly coupling target words and syntax in the NMT decoder. To represent syntax explicitly, we used CCG supertags, which encode subcategorization information, capturing long distance dependencies and attachments. Our method improved translation quality on several difficult linguistic constructs, including prepositional phrases which are the most frequent type of predicate arguments. These improvements over a strong baseline NMT system were consistent across two language pairs: 0.9 BLEU for German-English and 1.2 BLEU for Romanian-English.

La thèse, effectuée dans le cadre d'une bourse CIFRE, et prolongeant un des aspects du projet ANR Traouiero, aborde d'abord la production, l'extension et l'amélioration de corpus multilingues par traduction automatique (TA) et post-édition contributive (PE). Des améliorations fonctionnelles et techniques ont été apportées aux logiciels SECTra et iMAG, et on a progressé vers une définition générique de la structure d'un corpus multilingue, multi-annoté et multimédia, pouvant contenir des documents classiques aussi bien que des pseudo-documents et des méta-segments. Cette partie a été validée par la création de bons corpus bilingues français-chinois, l'un d'eux résultant de la toute première application à la traduction littéraire.Une seconde partie, initialement motivée par un besoin industriel, a consisté à construire des systèmes de TA de type Moses, spécialisés à des sous-langages, en français↔chinois, et à étudier la façon de les améliorer dans le cadre d'un usage en continu avec possibilité de PE. Dans le cadre d'un projet interne sur le site du LIG et d'un projet (TABE-FC) en coopération avec l'université de Xiamen, on a pu démontrer l'intérêt de l'apprentissage incrémental en TA statistique, sous certaines conditions, grâce à une expérience qui s'est étalée sur toute la thèse.La troisième partie est consacrée à des contributions et mises à disposition de supports informatiques et de ressources. Les principales se placent dans le cadre du projet COST MUMIA de l'EU et résultent de l'exploitation de la collection CLEF-2011 de 1,5 M de brevets partiellement multilingues. De grosses mémoires de traductions en ont été extraites (17,5 M segments), 3 systèmes de TA en ont été tirés, et un site Web de support à la RI multilingue sur les brevets a été construit. On décrit aussi la réalisation en cours de JianDan-eval, une plate-forme de construction, déploiement et évaluation de systèmes de TA. / The thesis, conducted as part of a CIFRE grant, and extending one of the aspects of the ANR project Traouiero, first addresses the production, extension and improvement of multilingual corpora by machine translation (MT) and contributory post-editing (PE). Functional and technical improvements have been made to the SECTra and iMAG software produced in previous PhD theses (P.C. Huynh, H.T. Nguyen), and progress has ben made toward a generic definition of the structure of a multilingual, annotated and multi-media corpus that may contain usual documents as well as pseudo-documents (such as Web pages) and meta-segments. This part has been validated by the creation of good French-Chinese bilingual corpora, one of them resulting from the first application to literary translation (a Jules Verne novel).A second part, initially motivated by an industrial need, has consisted in building MT systems of Moses type, specialized to sub-languages, for french↔chinese, and to study how to improve them in the context of a continuous use with the possibility of PE. As part of an internal project on the LIG website and of a project (TABE-FC) in cooperation with Xiamen University, it has been possible to demonstrate the value of incremental learning in statistical MT, under certain conditions, through an experiment that spread over the whole thesis.The third part of the thesis is devoted to contributing and making available computer tools and resources. The main ones are related to the COST project MUMIA of the EU and result from the exploitation of the CLEF-2011 collection of 1.5 million partially multilingual patents. Large translation memories have been extracted from it (17.5 million segments), 3 MT systems have been produced (de-fr, en-fr, fr-de), and a website of support for multilingual IR on patents has been constructed. One also describes the on-going implementation of JianDan-eval, a platform for building, deploying and evaluating MT systems.