Thanks. It is based on http://nlp.stanford.edu/~wcmac/papers/nli-diss.pdf.

I use a simpler alignment representation, and use semantic role types in the predictions.