Comunicación

Over the past years, there have been various efforts at incorporating
monolingual syntax into statistical machine translation. The main difficulty
often seems that translation equivalence does not necessarily conform to
monolingual syntactic constituency structure, as assumed by traditional
compositional semantics. In this talk I will reflect on the origins of
syntactic structure to show that the information represented in a syntactic
tree can be represented in alternative ways. By representing this information
adequately, syntactic information can be brought into statistical MT models
effectively. In will give a bird's-eye view of three recent models developed
together with my co-workers and touch upon some relevant aspects. After
that I will discuss the reverse problem of how to represent word alignments
hierarchically in order to capture the relations of translation equivalence
adequately. By doing so, our aim is to use this new hierarchical representation
to pave the way for learning compositional statistical synchronous grammars
from word-aligned parallel corpora. This is ongoing work with the fundamental
results currently in submission.