Baseline fix

[kudkudak]

Turns out it is important to calculate counts only based on tokens of given passage, rather than estimated on all stories. It is important to produce similar results as in https://www.microsoft.com/en-us/research/wp-content/uploads/2016/11/MCTest_EMNLP2013.pdf. Performance changes from

From:

		correct
question_type	subset
multiple	mc160	0.523438
	mc500	0.414634
one	mc160	0.508929
	mc500	0.477941

To:

		correct
question_type	subset
multiple	mc160	0.578125
	mc500	0.539634
one	mc160	0.678571
	mc500	0.544118

Fix is to calculate icounts as _icounts = compute_inverse_counts([tokens]) in SlidingWindow.predict_target

[kudkudak]

Another fix: baseline dist calculation should use float as: closest = np.abs(last_q - last_a) / (float(len(passage)) - 1). I understand stopwords removal is done during data preprocessing

[allanj]

I couldn't obtain the results you list there after the suggested modification.