Syntax-to-Morphology Mapping in Factored SMT

Morphology is many morphologically complex languages encodes
information that is expressed with syntax involving a series of
function words in a language like English. In such cases, factored
phrase-based SMT is not directly applicable as there is no relation
between the morphological structures of words on both sides. In this
talk we present a scheme to employ factored phrase-based SMT when the
morphological structures of the source and target languages are widely
disparate, and experiment with it in SMT between English and Turkish.
Our approach relies on syntactic analysis on the morphologically poor
side (English) and then encodes a wide variety of local and non-local
syntactic structures as \emph{complex structural tags} which appear as
additional factors in the training data. On the
morphologically-complex side (Turkish), we only perform morphological
analysis and disambiguation but treat the complete complex
morphological tag as a factor, instead of separating morphemes (as is
traditionally done with morphologically complex languages.)  Such a
representation has three important side effects: (i) the length of the
English sentences reduce substantially as many function words in the
source are now encoded in complex tags; (ii) the abstraction of syntax
mediated by function words as complex tags, allows the standard phrase
extraction schemes to conflate discontinuous and continuous variants
of the phrases; and (iii) since most syntax is now abstracted leaving
behind mostly content words and complex tags, word alignment can
operate on just root words, without noise from function words.

We incrementally explore capturing various syntactic substructures as
complex tags on the English side, and evaluate how our translations
improve in BLEU scores in English-to-Turkish SMT.  We also apply the
approach in the reverse direction but with less success, as the
syntactic transformations on the English side have to be undone after
SMT and this introduces some additional complications.  Currently, the
syntax-to-morphology transformations are manually developed and we are
investigating on extracting these transformation automatically in an
incremental fashion and will briefly present our ideas on this.

This is joint work with Reyyan Yeniterzi of CMU- LTI.