Transforming Global Content

Transforming Global Content

Projects on Machine Translation (MT) modelling, MT training data scarcity and human factors are pivotal to extending research in MT, human translation and their business impact.

New deep learning techniques are being augmented with linguistic knowledge to constrain the MT decoding space explosion due to increasing model complexity. Cloud-based models seed MT engines built on-the-fly using small amounts of data targeted to the translational requirements of the input document. We extend our previous research on domain adaptation to new ADAPT sectors and data types using grounding semantics, filtering out ‘noisy’ input, and where data is in short supply, supplement parallel training data with comparable corpora. We also extend our previous ethnographic studies of real users of MT output, which will uncover cognitive and social barriers to MT acceptability. Novel evaluation schemes are being developed which meet industry needs for flexible, configurable quality measures that reflect directly their core organisational goals.

Research team

Publications

Conference

Distributed dimensionality reduction of industrial data based on clustering

  • Posted: 31 May 2018
  • Author: , Yongyan Zhang, Guo Xie, Wenqing Wang, Xiaofan Wang, Fucai Qian, Xulong Du
  • Publication: ICIEA 2018 - 13th IEEE Conference on Industrial Electronics and Applications (ICIEA)
Journal Article

Translation Facilitates Comprehension of Health-Related Crisis Information: Kenya as an Example

  • Posted: 1 Jul 2017
  • Author: Sharon O’Brien
  • Publication: JoSTrans: The Journal of Specialised Translation
Transforming Global Content

Elastic-substitution decoding for Hierarchical SMT: efficiency, richer search and double labels

  • Posted: 1 Aug 2017
  • Author: Andy Way, Gideon Maillette de Buy Wenniger, Khalil Simaan
  • Publication: MT Summit XVI - 16th Machine Translation Summit
Conference

Building English-to-Serbian machine translation system for IMDb movie reviews

  • Posted: 2 Aug 2019
  • Author: Pintu Lohar, Maja Popović, Andy Way
  • Publication: BSNLP 2019 - 7th Workshop on. Balto-Slavic Natural Language Processing

Research Goals

We provide MT with increased intelligence, by developing engines incorporating syntax, semantics and discourse features, constrained MT models using deep learning techniques, cloud-based data models for use by (disposable) MT engines, and engines for sentiment analysis and translation.

We connect texts with the real world, and investigate different ways to leverage grounding semantics (in contrast to abstract semantics), including named entities and relations, multimodality, and discourse semantics to improve translation quality in various scenarios. We use the state-of-the-art neural MT framework to incorporate grounding semantics and rich linguistic features.

Through a human-factors oriented approach, we seek to understand what the blocking points are, in order to overcome them. We take a cognitive ergonomic approach to this, which investigates three types of factors: cognitive (i.e. best presentation), physical (i.e. reduced editing effort), and organisational (i.e. best organisation for the adoption of MT).

Twitter
6 pm
@Adaptcentre
Congratulations to @LindaDoyle on making history @tcddublin! Looking forward to you leading on #ImagineTrinity over… twitter.com/i/web/status/1…

Newsletter

Sign up to our newsletter for all the latest updates on ADAPT news, events and programmes.
Archived newsletters can be viewed here.