The unreasonable effectiveness of Neural Models in Language Decoding

Tony O'Dowd


Abstract
This tutorial will provide an in-depth look at the experiments, jointly carried out by KantanMT and eBay during 2018, to determine which Neural Model delivers the best translation performance for eBay Customer Service content. It will lay out the timeline, process and mechanisms used to customise Neural MT models and how these were used in conjunction with Human Based evaluations to determine which approach to Neural MT provided the best translation outcomes.The tutorial will cover the following topics and methods:- Structural differences in Neural Networks and how they assist the language decoding process – RNN, CNN and TNN will be covered in detailed.- Customisation of Neural MT using the KantanMT Platform- Using MQM Framework for the evaluation and comparison of Translation Outputs and comparison to Human Translation- Collation and analysis of experimental findings in reaching our decision to standardise on Transformer type networks.Participants of the tutorial will get a clear understanding of Neural Model types and the differences, it will also cover how to customise these models and then how to set up a controlled experiment to determine translation performance.
Anthology ID:
W19-7601
Volume:
Proceedings of Machine Translation Summit XVII: Tutorial Abstracts
Month:
August
Year:
2019
Address:
Dublin, Ireland
Editor:
Laura Rossi
Venue:
MTSummit
SIG:
Publisher:
European Association for Machine Translation
Note:
Pages:
Language:
URL:
https://aclanthology.org/W19-7601
DOI:
Bibkey:
Cite (ACL):
Tony O'Dowd. 2019. The unreasonable effectiveness of Neural Models in Language Decoding. In Proceedings of Machine Translation Summit XVII: Tutorial Abstracts, Dublin, Ireland. European Association for Machine Translation.
Cite (Informal):
The unreasonable effectiveness of Neural Models in Language Decoding (O’Dowd, MTSummit 2019)
Copy Citation: