The Fifth Workshop on Syntax, Semantics and Structure in Statistical Translation (SSST-5) seeks to build on the foundations established in the first four SSST workshops, which brought together a large number of researchers working on diverse aspects of structure and representation in relation to statistical machine translation. Its program each year has comprised high-quality papers discussing current work spanning topics including: new grammatical models of translation; new learning methods for syntax-based models; formal properties of synchronous/transduction grammars (hereafter S/TGs); discriminative training of models incorporating linguistic features; using S/TGs for semantics and generation; and syntax- and semantics-based evaluation of machine translation.
The need for structural mappings between languages is widely recognized in the fields of statistical machine translation and spoken language translation, and there is a growing consensus that these mappings are appropriately represented using a family of formalisms that includes synchronous/transduction grammars and their tree-transducer equivalents. To date, flat-structured models, such as the word-based IBM models of the early 1990s or the more recent phrase-based models, remain widely used. But tree-structured mappings arguably offer a much greater potential for learning valid generalizations about relationships between languages.
Within this area of research there is a rich diversity of approaches. There is active research ranging from formal properties of S/TGs to large-scale end-to-end systems. There are approaches that make heavy use of linguistic theory, and approaches that use little or none. There is theoretical work characterizing the expressiveness and complexity of particular formalisms, as well as empirical work assessing their modeling accuracy and descriptive adequacy across various language pairs. There is work being done to invent better translation models, and work to design better algorithms. Recent years have seen significant progress on all these fronts. In particular, systems based on these formalisms are now top contenders in MT evaluations.
At the same time, SMT has seen a movement toward semantics over the past five years, which has been reflected at recent SSST workshops. The issues of deep syntax and shallow semantics are closely linked. Semantic SMT research now includes semantic role labeling (SRL) for MT evaluation, SRL for SMT, and WSD for SMT.
In order to emphasize structure and representation at semantic and not only syntactic levels, “Semantics” has been explicitly added to the name of this year’s Workshop (the acronym remains SSST), and is a special workshop theme. Special sessions will be devoted to the Semantics theme.