def test_count_3gram(self):
sents = [
'el gato come pescado .'.split(),
'la gata come salmón .'.split(),
'unaria'.split(),
]
ngram = NGram(3, sents)
counts = {
('<s>', 'el'): 1,
('el', 'gato'): 1,
('gato', 'come'): 1,
('come', 'pescado'): 1,
('pescado', '.'): 1,
('<s>', 'la'): 1,
('la', 'gata'): 1,
('gata', 'come'): 1,
('come', 'salmón'): 1,
('salmón', '.'): 1,
('<s>', 'unaria'): 1,
('<s>', 'el', 'gato'): 1,
('el', 'gato', 'come'): 1,
('gato', 'come', 'pescado'): 1,
('come', 'pescado', '.'): 1,
('<s>', 'la', 'gata'): 1,
('la', 'gata', 'come'): 1,
('gata', 'come', 'salmón'): 1,
('come', 'salmón', '.'): 1,
('salmón', '.', '</s>'): 1,
('<s>', 'unaria', '</s>'): 1,
}
for gram, c in counts.items():
self.assertEqual(ngram.count(gram), c)
def test_count_3gram(self):
ngram = NGram(3, self.sents)
counts = {
('<s>', 'el'): 1,
('el', 'gato'): 1,
('gato', 'come'): 1,
('come', 'pescado'): 1,
('pescado', '.'): 1,
('<s>', '<s>'): 2,
('<s>', 'la'): 1,
('la', 'gata'): 1,
('gata', 'come'): 1,
('come', 'salmón'): 1,
('salmón', '.'): 1,
('<s>', '<s>', 'el'): 1,
('<s>', 'el', 'gato'): 1,
('el', 'gato', 'come'): 1,
('gato', 'come', 'pescado'): 1,
('come', 'pescado', '.'): 1,
('pescado', '.', '</s>'): 1,
('<s>', '<s>', 'la'): 1,
('<s>', 'la', 'gata'): 1,
('la', 'gata', 'come'): 1,
('gata', 'come', 'salmón'): 1,
('come', 'salmón', '.'): 1,
('salmón', '.', '</s>'): 1,
}
for gram, c in counts.items():
self.assertEqual(ngram.count(gram), c)
def test_generate_sent_3and4gram(self):
ngram = NGram(3, self.sents4)
ngram2 = NGram(4, self.sents4)
generator = NGramGenerator(ngram)
generator2 = NGramGenerator(ngram2)
# all the possible generated sentences for 3 or 4-grams:
sents = [
'la casa se construye y el corre y la gata come ensalada',
'el corre y la gata come pescado y duerme',
'la casa se construye y el corre y la gata come ensalada',
'la casa se construye y el corre y la gata come pescado y duerme',
'la casa se construye y el corre',
'la gata come pescado y duerme',
'el corre y la gata come ensalada',
'el corre',
'la gata come ensalada',
'la casa se construye y el corre',
'la gata come pescado y duerme',
]
for i in range(1000):
sent = generator.generate_sent()
sent2 = generator2.generate_sent()
self.assertTrue(' '.join(sent) in sents)
self.assertTrue(' '.join(sent2) in sents)
def test_count_2gram(self):
ngram = NGram(2, self.sents)
counts = {
('<s>',): 2,
('el',): 1,
('gato',): 1,
('come',): 2,
('pescado',): 1,
('.',): 2,
('la',): 1,
('gata',): 1,
('salmón',): 1,
('<s>', 'el'): 1,
('el', 'gato'): 1,
('gato', 'come'): 1,
('come', 'pescado'): 1,
('pescado', '.'): 1,
('.', '</s>'): 2,
('<s>', 'la'): 1,
('la', 'gata'): 1,
('gata', 'come'): 1,
('come', 'salmón'): 1,
('salmón', '.'): 1,
}
for gram, c in counts.items():
self.assertEqual(ngram.count(gram), c)
def test_cond_prob_2gram(self):
ngram = NGram(2, self.sents)
probs = {
('pescado', 'come'): 0.5,
('salmón', 'come'): 0.5,
('salame', 'come'): 0.0,
}
for (token, prev), p in probs.items():
self.assertAlmostEqual(ngram.cond_prob(token, (prev,)), p)
def test_cond_prob_1gram(self):
ngram = NGram(1, self.sents)
probs = {
'pescado': 1 / 12.0,
'come': 2 / 12.0,
'salame': 0.0,
}
for token, p in probs.items():
self.assertEqual(ngram.cond_prob(token), p)
def test_cond_prob_1gram(self):
ngram = NGram(1, self.sents)
probs = {
'pescado': 1 / 12.0,
'come': 2 / 12.0,
'salame': 0.0,
}
for token, p in probs.items():
self.assertAlmostEqual(ngram.cond_prob(token), p)
def test_cond_prob_2gram(self):
ngram = NGram(2, self.sents)
probs = {
('pescado', 'come'): 0.5,
('salmón', 'come'): 0.5,
('salame', 'come'): 0.0,
}
for (token, prev), p in probs.items():
self.assertEqual(ngram.cond_prob(token, [prev]), p)
def test_sent_prob_1gram(self):
ngram = NGram(1, self.sents)
sents = {
# 'come', '.' and '</s>' have prob 1/6, the rest have 1/12.
'el gato come pescado .': (1 / 6.0)**3 * (1 / 12.0)**3,
'la gata come salmón .': (1 / 6.0)**3 * (1 / 12.0)**3,
'el gato come salame .': 0.0, # 'salame' unseen
'la la la': (1 / 6.0)**1 * (1 / 12.0)**3,
}
for sent, prob in sents.items():
self.assertAlmostEqual(ngram.sent_prob(sent.split()), prob, msg=sent)
def test_sent_prob_1gram(self):
ngram = NGram(1, self.sents)
sents = {
# 'come', '.' and '</s>' have prob 1/6, the rest have 1/12.
'el gato come pescado .': (1 / 6.0)**3 * (1 / 12.0)**3,
'la gata come salmón .': (1 / 6.0)**3 * (1 / 12.0)**3,
'el gato come salame .': 0.0, # 'salame' unseen
'la la la': (1 / 6.0)**1 * (1 / 12.0)**3,
}
for sent, prob in sents.items():
self.assertAlmostEqual(ngram.sent_prob(sent.split()), prob, msg=sent)
def test_sent_log_prob_1gram(self):
ngram = NGram(1, self.sents)
def log2(x): return log(x, 2)
sents = {
# 'come', '.' and '</s>' have prob 1/6, the rest have 1/12.
'el gato come pescado .': 3 * log2(1 / 6.0) + 3 * log2(1 / 12.0),
'la gata come salmón .': 3 * log2(1 / 6.0) + 3 * log2(1 / 12.0),
'el gato come salame .': -inf, # 'salame' unseen
'la la la': log2(1 / 6.0) + 3 * log2(1 / 12.0),
}
for sent, prob in sents.items():
self.assertAlmostEqual(ngram.sent_log_prob(sent.split()), prob, msg=sent)
def test_cond_prob_3gram(self):
ngram = NGram(3, self.sents)
probs = {
('pescado', ('gato', 'come')): 1,
('salmón', ('gata', 'come')): 1,
('salame', ('gato', 'come')): 0.0,
('gato', ('<s>', 'el')): 1,
('gata', ('<s>', 'el')): 0.0,
}
for (token, prev), p in probs.items():
self.assertEqual(ngram.cond_prob(token, list(prev)), p)
def test_sent_prob_3gram(self):
ngram = NGram(3, self.sents)
sents = {
# after '<s>, <s>': 'el' and 'la' have prob 0.5.
'el gato come pescado .': 0.5 * 1 * 1 * 1 * 1,
'la gata come salmón .': 0.5 * 1 * 1 * 1 * 1,
'el gato come salmón .': 0.5 * 1 * 1 * 0 * 1 * 1, # prob('gato come salmon') = 0
'el gato come salame .': 0.0, # 'salame' unseen
'la la la': 0.0, # 'la' after 'la' unseen
}
for sent, prob in sents.items():
self.assertAlmostEqual(ngram.sent_prob(sent.split()), prob, msg=sent)
def test_cond_prob_3gram(self):
ngram = NGram(3, self.sents)
probs = {
('pescado', 'gato', 'come'): 1,
('salmón', 'gata', 'come'): 1,
('salame', 'gato', 'come'): 0.0,
('salmón', 'gato', 'come'): 0.0,
('el', '<s>', '<s>'): 0.5,
('</s>', 'pescado', '.'): 1,
}
for (token, prev1, prev2), p in probs.items():
self.assertEqual(ngram.cond_prob(token, [prev1, prev2]), p)
def test_cond_prob_3gram(self):
ngram = NGram(3, self.sents)
probs = {
('pescado', 'gato', 'come'): 1,
('salmón', 'gata', 'come'): 1,
('salame', 'gato', 'come'): 0.0,
('salmón', 'gato', 'come'): 0.0,
('el', '<s>', '<s>'): 0.5,
('</s>', 'pescado', '.'): 1,
}
for (token, prev1, prev2), p in probs.items():
self.assertEqual(ngram.cond_prob(token, [prev1, prev2]), p)
def test_sent_log_prob_1gram(self):
ngram = NGram(1, self.sents)
log2 = lambda x: log(x, 2)
sents = {
# 'come', '.' and '</s>' have prob 1/6, the rest have 1/12.
'el gato come pescado .': 3 * log2(1 / 6.0) + 3 * log2(1 / 12.0),
'la gata come salmón .': 3 * log2(1 / 6.0) + 3 * log2(1 / 12.0),
'el gato come salame .': float('-inf'), # 'salame' unseen
'la la la': log2(1 / 6.0) + 3 * log2(1 / 12.0),
}
for sent, prob in sents.items():
self.assertAlmostEqual(ngram.sent_log_prob(sent.split()), prob, msg=sent)
def test_cond_prob_4gram(self):
ngram = NGram(4, self.sents2)
probs = {
('pescado', ('el', 'gato', 'come')): 1,
('salmón', ('la', 'gata', 'come')): 0.0,
('salame', ('el', 'gato', 'come')): 0.0,
('gato', ('<s>', '<s>', 'el')): 1,
('viejo', ('gato', 'come', 'pescado')): 0.5,
('fresco', ('gato', 'come', 'pescado')): 0.5,
}
for (token, prev), p in probs.items():
self.assertEqual(ngram.cond_prob(token, list(prev)), p)
def test_sent_prob_2gram(self):
ngram = NGram(2, self.sents)
sents = {
# after '<s>': 'el' and 'la' have prob 0.5.
# after 'come': 'pescado' and 'salmón' have prob 0.5.
'el gato come pescado .': 0.5 * 0.5,
'la gata come salmón .': 0.5 * 0.5,
'el gato come salmón .': 0.5 * 0.5,
'la gata come pescado .': 0.5 * 0.5,
'el gato come salame .': 0.0, # 'salame' unseen
'la la la': 0.0, # 'la' after 'la' unseen
}
for sent, prob in sents.items():
self.assertAlmostEqual(ngram.sent_prob(sent.split()), prob, msg=sent)
def test_sent_prob_2gram(self):
ngram = NGram(2, self.sents)
sents = {
# after '<s>': 'el' and 'la' have prob 0.5.
# after 'come': 'pescado' and 'salmón' have prob 0.5.
'el gato come pescado .': 0.5 * 0.5,
'la gata come salmón .': 0.5 * 0.5,
'el gato come salmón .': 0.5 * 0.5,
'la gata come pescado .': 0.5 * 0.5,
'el gato come salame .': 0.0, # 'salame' unseen
'la la la': 0.0, # 'la' after 'la' unseen
}
for sent, prob in sents.items():
self.assertAlmostEqual(ngram.sent_prob(sent.split()), prob, msg=sent)
def test_sent_log_prob_2gram(self):
ngram = NGram(2, self.sents)
log2 = lambda x: log(x, 2)
sents = {
# after '<s>': 'el' and 'la' have prob 0.5.
# after 'come': 'pescado' and 'salmón' have prob 0.5.
'el gato come pescado .': 2 * log2(0.5),
'la gata come salmón .': 2 * log2(0.5),
'el gato come salmón .': 2 * log2(0.5),
'la gata come pescado .': 2 * log2(0.5),
'el gato come salame .': float('-inf'), # 'salame' unseen
'la la la': float('-inf'), # 'la' after 'la' unseen
}
for sent, prob in sents.items():
self.assertAlmostEqual(ngram.sent_log_prob(sent.split()), prob, msg=sent)
def test_sent_prob_3gram(self):
ngram = NGram(3, self.sents)
sents = {
# after '<s>, <s>': 'el' and 'la' have prob 0.5.
'el gato come pescado .': 0.5 * 1 * 1 * 1 * 1,
'la gata come salmón .': 0.5 * 1 * 1 * 1 * 1,
'el gato come salmón .':
0.5 * 1 * 1 * 0 * 1 * 1, # prob('gato come salmon') = 0
'el gato come salame .': 0.0, # 'salame' unseen
'la la la': 0.0, # 'la' after 'la' unseen
}
for sent, prob in sents.items():
self.assertAlmostEqual(ngram.sent_prob(sent.split()),
prob,
msg=sent)
def test_init_3gram(self):
ngram = NGram(3, self.sents)
generator = NGramGenerator(ngram)
probs = {
('<s>', '<s>'): {'el': 0.5, 'la': 0.5},
('<s>', 'el',): {'gato': 1.0},
('el', 'gato'): {'come': 1.0},
('gato', 'come'): {'pescado': 1.0},
('come', 'pescado'): {'.': 1.0},
('pescado', '.'): {'</s>': 1.0},
('<s>', 'la'): {'gata': 1.0},
('la', 'gata'): {'come': 1.0},
('gata', 'come'): {'salmón': 1.0},
('come', 'salmón'): {'.': 1.0},
('salmón', '.'): {'</s>': 1.0},
}
sorted_probs = {
('<s>', '<s>'): [('el', 0.5), ('la', 0.5)],
('<s>', 'el',): [('gato', 1.0)],
('el', 'gato'): [('come', 1.0)],
('gato', 'come'): [('pescado', 1.0)],
('come', 'pescado'): [('.', 1.0)],
('pescado', '.'): [('</s>', 1.0)],
('<s>', 'la'): [('gata', 1.0)],
('la', 'gata'): [('come', 1.0)],
('gata', 'come'): [('salmón', 1.0)],
('come', 'salmón'): [('.', 1.0)],
('salmón', '.'): [('</s>', 1.0)],
}
self.assertEqual(dict(generator.probs), probs)
self.assertEqual(generator.sorted_probs, sorted_probs)
def test_sent_log_prob_2gram(self):
ngram = NGram(2, self.sents)
def log2(x): return log(x, 2)
sents = {
# after '<s>': 'el' and 'la' have prob 0.5.
# after 'come': 'pescado' and 'salmón' have prob 0.5.
'el gato come pescado .': 2 * log2(0.5),
'la gata come salmón .': 2 * log2(0.5),
'el gato come salmón .': 2 * log2(0.5),
'la gata come pescado .': 2 * log2(0.5),
'el gato come salame .': -inf, # 'salame' unseen
'la la la': -inf, # 'la' after 'la' unseen
}
for sent, prob in sents.items():
self.assertAlmostEqual(ngram.sent_log_prob(sent.split()), prob, msg=sent)
def test_count_1gram(self):
ngram = NGram(1, self.sents)
counts = {
(): 12,
('el',): 1,
('gato',): 1,
('come',): 2,
('pescado',): 1,
('.',): 2,
('</s>',): 2,
('la',): 1,
('gata',): 1,
('salmón',): 1,
}
for gram, c in counts.items():
self.assertEqual(ngram.count(gram), c)
def test_sent_log_prob_3gram(self):
ngram = NGram(3, self.sents)
log2 = lambda x: log(x, 2)
sents = {
# after '<s>': 'el' and 'la' have prob 0.5.
# after 'come': 'pescado' and 'salmón' have prob 0.5.
'el gato come pescado .': log2(0.5),
'la gata come salmón .': log2(0.5),
'el gato come salmón .': float('-inf'),
'la gata come pescado .': float('-inf'),
'el gato come salame .': float('-inf'), # 'salame' unseen
'la la la': float('-inf'), # 'la' after 'la' unseen
}
for sent, prob in sents.items():
self.assertAlmostEqual(ngram.sent_log_prob(sent.split()),
prob,
msg=sent)
def test_generate_sent_1gram(self):
ngram = NGram(1, self.sents)
generator = NGramGenerator(ngram)
voc = {'el', 'gato', 'come', 'pescado', '.', 'la', 'gata', 'salmón'}
for i in range(100):
sent = generator.generate_sent()
self.assertTrue(set(sent).issubset(voc))
def test_generate_token_3and4gram(self):
ngram = NGram(3, self.sents3)
ngram2 = NGram(4, self.sents3)
generator = NGramGenerator(ngram)
generator2 = NGramGenerator(ngram2)
for i in range(100):
# after 'come pescado' always comes 'y'
token = generator.generate_token(('come', 'pescado'))
self.assertEqual(token, 'y')
# after 'come pescado y' always comes 'duerme'
token = generator2.generate_token(('come', 'pescado', 'y'))
self.assertEqual(token, 'duerme')
# sentence may come start with 'el' or 'la'
token = generator.generate_token(('<s>', '<s>'))
self.assertTrue(token in ['el', 'la'])
token = generator2.generate_token(('<s>', '<s>', '<s>'))
self.assertTrue(token in ['el', 'la'])
def test_generate_token(self):
ngram = NGram(2, self.sents)
generator = NGramGenerator(ngram)
for i in range(100):
# after 'el' always comes 'gato':
token = generator.generate_token(('el',))
self.assertEqual(token, 'gato')
# after 'come' may come 'pescado' or 'salmón'
token = generator.generate_token(('come',))
self.assertTrue(token in ['pescado', 'salmón'])
def test_perplexity_calculation(self):
model = NGram(3, self.sents)
model2 = NGram(4, self.sents)
perplexity = 2**(-(log2(0.5)*2)/10.0)
self.assertEqual(perplexity, model.perplexity(self.sents))
self.assertEqual(perplexity, model2.perplexity(self.sents))
def test_generate_sent_2gram(self):
ngram = NGram(2, self.sents)
generator = NGramGenerator(ngram)
# all the possible generated sentences for 2-grams:
sents = [
'el gato come pescado .',
'la gata come salmón .',
'el gato come salmón .',
'la gata come pescado .',
]
for i in range(100):
sent = generator.generate_sent()
self.assertTrue(' '.join(sent) in sents, sent)
def test_init_1gram(self):
ngram = NGram(1, self.sents)
counts = {
(): 10,
('el',): 1,
('gato',): 1,
('come',): 2,
('pescado',): 1,
('.',): 2,
('la',): 1,
('gata',): 1,
('salmón',): 1,
}
self.assertEqual(dict(ngram.counts), counts)
def test_init_1gram(self):
ngram = NGram(1, self.sents)
generator = NGramGenerator(ngram)
probs = {
(): {
'el': 1 / 12.0,
'gato': 1 / 12.0,
'come': 2 / 12.0,
'pescado': 1 / 12.0,
'.': 2 / 12.0,
'</s>': 2 / 12.0,
'la': 1 / 12.0,
'gata': 1 / 12.0,
'salmón': 1 / 12.0,
}
}
self.assertEqual(generator._probs, probs)
def test_sent_log_prob_3and4gram(self):
ngram = NGram(3, self.sents2)
ngram2 = NGram(4, self.sents2)
sents = {
'el gato come pescado nuevo .': float('-inf'), # 'nuevo' unseen
'la la la': float('-inf'), # 'la' after 'la' unseen
# after 'pescado': 'viejo' and 'fresco' have prob 0.5.
'el gato come pescado fresco . ': log2(0.5),
'el gato come pescado viejo . ': log2(0.5)
}
for sent, prob in sents.items():
self.assertAlmostEqual(ngram.sent_log_prob(sent.split()),
prob, msg=sent)
self.assertAlmostEqual(ngram2.sent_log_prob(sent.split()),
prob, msg=sent)
def test_init_2gram(self):
ngram = NGram(2, self.sents)
counts = {
('el',): 1,
('gato',): 1,
('come',): 2,
('pescado',): 1,
# ('.',): 2,
('la',): 1,
('gata',): 1,
('salmón',): 1,
('el', 'gato'): 1,
('gato', 'come'): 1,
('come', 'pescado'): 1,
('pescado', '.'): 1,
('la', 'gata'): 1,
('gata', 'come'): 1,
('come', 'salmón'): 1,
('salmón', '.'): 1,
}
self.assertEqual(dict(ngram.counts), counts)
def test_sent_prob_3and4gram(self):
ngram = NGram(3, self.sents3)
ngram2 = NGram(4, self.sents3)
sents = {
'el gato come pescado y ronca .': 0.0, # 'ronca' unseen
'la la la': 0.0, # 'la' after 'la' unseen
# la probabilidad se da por el principio,
# si empieza con 'la' o 'el'
'el gato come pescado y duerme . ': 0.5,
'la gata come pescado y duerme . ': 0.5
}
for sent, prob in sents.items():
self.assertAlmostEqual(ngram.sent_prob(sent.split()),
prob, msg=sent)
self.assertAlmostEqual(ngram2.sent_prob(sent.split()),
prob, msg=sent)
def test_prob_1gram(self):
ngram = NGram(1, self.sents)
self.assertEqual(ngram.prob('pescado'), 0.1)
self.assertEqual(ngram.prob('come'), 0.2)
def test_prob(self):
ngram = NGram(2, self.sents)
self.assertEqual(ngram.prob('pescado', ['come']), 0.5)
self.assertEqual(ngram.prob('salmón', ['come']), 0.5)
# from languagemodeling.ngram import NGram, AddOneNGram, InterpolatedNGram
# models = {
# 'ngram': NGram,
# 'addone': AddOneNGram,
# 'inter': InterpolatedNGram,
# }
if __name__ == '__main__':
opts = docopt(__doc__)
# load the data
# WORK HERE!! LOAD YOUR TRAINING CORPUS
#sents = gutenberg.sents(['/Users/juliamilanese/Desktop/corpus_airbnb/txt'])
corpus = PlaintextCorpusReader('path',
'*.txt') #('corpus_airbnb/txt', '.*\.txt')
sents = list(corpus.sents())
# train the model
n = int(opts['-n'])
model = NGram(n, sents)
# model_class = models[opts['-m']]
# model = model_class(n, sents)
# save it
filename = opts['-o']
f = open(filename, 'wb')
pickle.dump(model, f)
f.close()
