def test_decodehmm(self):
hmm = hw5.NamedEntityRecognitionHMM()
hmm.train(self.train_tups)
test = [w[0] for w in self.train_tups[0]]
probs, pointers = hmm.generate_probabilities(test)
# find the actual labels
labels = hw5.decode(test, probs, pointers)
label_answers = [('Comparison', 'O'), ('in', 'O'), ('alkaline', 'B'), ('phosphatases', 'I'), ('in', 'O'), ('5', 'B'), ('-', 'I'), ('nucleotidase', 'I'), ('.', 'O')]
self.assertEqual(label_answers, labels)
def test_hmmprobspointers(self):
hmm = hw5.NamedEntityRecognitionHMM()
hmm.train(self.train_tups)
test = [w[0] for w in self.train_tups[0]]
probs, pointers = hmm.generate_probabilities(test)
# correct shape
self.assertEqual(9, len(probs))
self.assertEqual(9, len(pointers))
for row in probs:
self.assertEqual(3, len(row))
self.assertTrue(type(row) is dict)
self.assertTrue(type(row['O'] is float))
for row in pointers:
self.assertEqual(3, len(row))
self.assertTrue(type(row) is dict)
self.assertTrue(type(row['O'] is str))
# now we'll test the actual values
# pi = 0, .5, .5 (I, O, B)
vocab = 13
p_Comparison_I = (0 + 1) / (5 + vocab)
p_Comparison_0 = (1 + 1) / (8 + vocab)
p_Comparison_B = (0 + 1) / (3 + vocab)
# WARNING THIS CODE DOES NOT FULLY TEST THE PROBABILITIES, ONLY
# THE FIRST TWO COLUMNS
column1 = {
"I": 0 * (p_Comparison_I),
"O": 0.5 * (p_Comparison_0),
"B": 0.5 * (p_Comparison_B)
}
self.assertEqual(column1, probs[0])
p_in_I = (0 + 1) / (5 + vocab)
p_in_O = (3 + 1) / (8 + vocab)
p_in_B = (0 + 1) / (3 + vocab)
p_O_O = 4 / 8
p_O_I = 3 / 5
p_O_B = 0 / 3
p_I_O = 0 / 8
p_I_I = 2 / 5
p_I_B = 3 / 3
p_B_O = 2 / 8
p_B_I = 0 / 5
p_B_B = 0 / 3
# the "missing" 2 for the O states are because the sentences end with O,
# so these are accounted for in the distribution of pi
prev_I_and_I = p_in_I * p_I_I * column1["I"]
prev_O_and_I = p_in_I * p_I_O * column1["O"]
prev_B_and_I = p_in_I * p_I_B * column1["B"]
prev_I_and_O = p_in_O * p_O_I * column1["I"]
prev_O_and_O = p_in_O * p_O_O * column1["O"]
prev_B_and_O = p_in_O * p_O_B * column1["B"]
prev_I_and_B = p_in_B * p_B_I * column1["I"]
prev_O_and_B = p_in_B * p_B_O * column1["O"]
prev_B_and_B = p_in_B * p_B_B * column1["B"]
I_val = max(prev_I_and_I, prev_O_and_I, prev_B_and_I)
O_val = max(prev_I_and_O, prev_O_and_O, prev_B_and_O)
B_val = max(prev_I_and_B, prev_O_and_B, prev_B_and_B)
self.assertAlmostEqual(I_val, probs[1]["I"])
self.assertAlmostEqual(O_val, probs[1]["O"])
self.assertAlmostEqual(B_val, probs[1]["B"])
# ensure that back pointers are correct
point_answers = [{
'O': None,
'I': None,
'B': None
}, {
'O': 'O',
'I': 'B',
'B': 'O'
}, {
'O': 'O',
'I': 'B',
'B': 'O'
}, {
'O': 'O',
'I': 'B',
'B': 'O'
}, {
'O': 'I',
'I': 'I',
'B': 'O'
}, {
'O': 'O',
'I': 'I',
'B': 'O'
}, {
'O': 'O',
'I': 'B',
'B': 'O'
}, {
'O': 'I',
'I': 'I',
'B': 'O'
}, {
'O': 'I',
'I': 'I',
'B': 'O'
}]
self.assertEqual(point_answers, pointers)