Lexical Selection for Hybrid MT with Sequence Labeling
Alex Rudnick, Michael Gasser · 2013
We present initial work on an inexpensive approach for building largevocabulary lexical selection modules for hybrid RBMT systems by framing lexical selection as a sequence labeling problem. We submit that Maximum Entropy Markov Models (MEMMs) are a sensible formalism for this problem, due to their ability to take into account many features of the source text, and show how we can build a combination MEMM/HMM system that allows MT system implementors flexibility regarding which words have their lexical choices modeled with classifiers. We present initial results showing successful use of this system both in translating