That's correct observation. I'm guessing it has to do with whether the words after Open are indicative enough to the model that they should be brought in togeth

That's correct observation. I'm guessing it has to do with whether the words after Open are indicative enough to the model that they should be brought in together with Open. As I said in other comments, with more training data this issue will likely go away. And these tough comments are the best candidates.