Fixing random bugs in parser training. I discovered that the execution of sempre logical forms currently does not give us correct answers for logical forms in the DPD output. I added a workaround to this problem in parser evaluation that checks if the predicted logical forms is in the DPD output.
These changes get us to an oracle@5 training accuracy of 82%, which is the best we can do on the training set given the current parser's action space.
Fixing random bugs in parser training. I discovered that the execution of sempre logical forms currently does not give us correct answers for logical forms in the DPD output. I added a workaround to this problem in parser evaluation that checks if the predicted logical forms is in the DPD output.
These changes get us to an oracle@5 training accuracy of 82%, which is the best we can do on the training set given the current parser's action space.