Notes from our discussion:

- why was a non-standard evaluation method used? Test set for the shared data
  task was split into set used for training and actual evaluation set, leading 
  to a smaller than usual set. Results are therefore not comparable to other 
  numbers published previously.

- not clear whether OOVs were less of a problem due to training on a subset of
  the test set

- not fair to compare a model trained on some hand-labeled data to a model
  using no hand-labeled data at all: the IBM 4 model could have been trained
  on the hand-labeled data as well  

- use of the bias feature was not clear

- the features that really make the model work are the IBM 4 features - maybe
  these are all that's needed and the other features are not relevant? In that
  case the proposed method would boil down to corrective training applied to 
  the IBM 4 model. The evaluation should have included discriminative matching
  trained only with the IBM-4 features and/or the baseline IBM-4 model with 
  features such as word length etc. included.

- error in dice coefficient equation? Should normally be C_E(e) + C_F(f) in
 the denominator. Is the equation wrong or did they really use a product?

- general unhappiness with the way word alignments are currently 
being evaluated: first, there's too much focus on improving word alignments
when it is not clear to what extent they help the overall system. Second,
the Hansard corpus in particular seems to have a disproportionately high
number of unsure links. Third, multi-to-one word alignments always need
to be resolved arbitrarily and it is sometimes really not possible
to say which word maps to which other word within a phrase;  
perhaps some notion of phrase correspondence should finally be incorporated
into alignment evaluations.