Alexandre Bérard - Neural Machine Translation Architectures and Applications

12:00
Friday
15
Jun
2018
Organized by: 
Laurent Besacier
Speaker: 
Alexandre Bérard
Teams: 

 

Composition du jury :

  • Philippe Langlais - Professeur à l'Université de Montréal
  • Béatrice Daille - Professeur à l'Université de Nantes
  • François Yvon - Professeur à l'Université Paris Sud
  • Pascale Sébillot - Professeur à l'INSA de Rennes
  • Marc Tommasi - Professeur à l'Université de Lille
    Directeurs de thèse :
  • Olivier Pietquin - Professeur à l'Université de Lille
  • Laurent Besacier - Professeur à l'Université Grenoble Alpes

 

This thesis is centered on two main objectives: adaptation of Neural Machine Translation techniques to new tasks and research replication.
Our efforts towards research replication have led to the production of two resources: MultiVec, a framework that facilitates the use of several techniques related to word embeddings (Word2vec, Bivec and Paragraph Vector); and a framework for Neural Machine Translation that implements several architectures and can be used for regular MT, Automatic Post-Editing, and Speech Recognition or Translation. These two resources are publicly available and now extensively used by the research community.
We extend our NMT framework to work on three related tasks: Machine Translation (MT), Automatic Speech Translation (AST) and Automatic Post-Editing (APE). For the machine translation task, we replicate pioneer neural-based work, and do a case study on TED talks where we advance the state-of-the-art.
Automatic speech translation consists in translating speech from one language to text in another language. In this thesis, we focus on the unexplored problem of end-to-end speech translation, which does not use an intermediate source language text transcription. We propose the first model for end-to-end AST and apply it on two benchmarks: translation of audiobooks and of basic travel expressions.
Our final task is automatic post-editing, which consists in automatically correcting the outputs of an MT system in a black-box scenario, by training on data that was produced by human post-editors. We replicate and extend published results on the WMT 2016 and 2017 tasks, and propose new neural architectures for low-resource automatic post-editing.