Neural Machine Translation has made huge progress recently. However, it remains very fragile to unexpected or noisy inputs. If neural models have a very strong language modelling capacity, they may lose adequacy to source sentences during decoding due to this strong LM capability. An additional reason to this robustness issue is the exposure bias.
Recently, several Reinforcement Learning methods have been already proposed to tackle the exposure bias. More recently, adversarial methods have been shown to be useful to improve stability and robustness of MT models.
In this internship, we propose to explore and benchmark several adversarial methods for machine translation.
Supervisor: Stéphane Clinchant
- Robust Neural Machine Translation with Doubly Adversarial Inputs, Cheng et al, ACL19
- On evaluation of adversarial perturbations, Michel et al , NAACL-HLT 2019
- A Study of Reinforcement Learning for Neural Machine Translation, Wu et al , EMNLP’18
NAVER LABS is a world class team of self-motivated and highly engaged researchers, engineers and interface designers collaborating together to create next generation ambient intelligence technology and services that are rich with the organic understanding they have of users, their contexts and situations.
Since 2013 LABS has led NAVER’s innovation in technology through products such as the AI-based translation app ‘Papago’, the omni-tasking web browser ‘Whale’, the virtual AI assistant ‘WAVE’, in-vehicle information entertainment system ‘AWAY’ and M1, the 3D indoor mapping robot.
The team in Europe is multidisciplinary and extremely multicultural specializing in artificial intelligence, machine learning, computer vision, natural language processing, UX and ethnography. We collaborate with many partners in the European scientific community on R&D projects.
NAVER LABS Europe is located in the south east of France in Grenoble. The notoriety of Grenoble comes from its exceptional natural environment and scientific ecosystem with 21,000 jobs in public and private research. It is home to 1 of the 4 French national institutes in AI called MIAI (Multidisciplinary Innovation in Ai) It has a large student community (over 62,000 students) and is a lively and cosmopolitan place, offering a host of leisure opportunities. Grenoble is close to both the Swiss and Italian borders and is the ideal place for skiing, hiking, climbing, hang gliding and all types of mountain sports.