Notes and ideas from today:

- bilmes: What is the commonality between phrase-based machine
  translation evaluation and models, and speech synthesis. Speech
  synthesis achieves its quality by piecing together pieces of
  previously uttered speech. Speech synthesis has a cost for adjoining
  successive pieces of speech. These phrase-based machine translation
  systems are not dissimilar, where "pieces" (phrases, etc.) of
  previously human-produced text are pieced together in a source
  language in order to match other pieces of previously human-produced
  target language text. Blue-score doesn't have a "concatenation cost"
  built in, and the MT models have a score on a language over the
  entire re-constructed text rather than just at the boundaries of the
  "segments" that are pieced together. This might bias down the
  potential sources of problem since the concatenations are only part
  of the score, and the rest of the score are within human-produced
  phrases which ideally should get high LM scores.  The idea is to
  bring a concatenation cost into MT.