TY  - GEN
Y1  - 2019///
TI  - Document Meta-Information as Weak Supervision for Machine Translation
N2  - Data-driven machine translation has advanced considerably since the first pioneering work
in the 1990s with recent systems claiming human parity on sentence translation for highresource tasks. However, performance degrades for low-resource domains with no available
sentence-parallel training data. Machine translation systems also rarely incorporate the
document context beyond the sentence level, ignoring knowledge which is essential for
some situations. In this thesis, we aim to address the two issues mentioned above by
examining ways to incorporate document-level meta-information into data-driven machine
translation. Examples of document meta-information include document authorship and
categorization information, as well as cross-lingual correspondences between documents,
such as hyperlinks or citations between documents. As this meta-information is much more
coarse-grained than reference translations, it constitutes a source of weak supervision for
machine translation. We present four cumulatively conducted case studies where we devise
and evaluate methods to exploit these sources of weak supervision both in low-resource
scenarios where no task-appropriate supervision from parallel data exists, and in a full
supervision scenario where weak supervision from document meta-information is used to
supplement supervision from sentence-level reference translations. All case studies show
improved translation quality when incorporating document meta-information.
ID  - heidok26780
AV  - public
A1  - Jehl, Laura Elisabeth
KW  - Natural Language Processing
KW  -  Neural Machine Translation
KW  -  Statistical Machine Translation
KW  -  Domain Adaptation
UR  - https://archiv.ub.uni-heidelberg.de/volltextserver/26780/
ER  -