
It has the model that assumes multiple empty words.

It also has the model that assumes a single empty words. I have three options for the distortion: 
1. empty-word: which differentiate empty words and non empty words. We use a scalar k = I for empty words. 
   - good performance, but it seems that lambda_\epsilon is too large. We need k for words that appear more than once in the same sentence.
2. uniform: uniform distortion (same as IBM model 1)
3. phi: f_j has larger probability to align with words that have larger fertility (bad model)
