support/language_models.py

   1 import string
   2 import random
   3 import collections
   4 import itertools
   5 from math import log10
   6 import os
   7
   8 import support.norms
   9
  10 def datafile(name, sep='\t'):
  11     """Read key,value pairs from file.
  12     """
  13     with open(os.path.join(os.path.dirname(os.path.realpath(__file__)), name), 'r') as f:
  14         for line in f:
  15             splits = line.split(sep)
  16             yield [splits[0], int(splits[1])]
  17
  18 english_counts = collections.Counter(dict(datafile('count_1l.txt')))
  19 normalised_english_counts = support.norms.normalise(english_counts)
  20
  21 english_bigram_counts = collections.Counter(dict(datafile('count_2l.txt')))
  22 normalised_english_bigram_counts = support.norms.normalise(english_bigram_counts)
  23
  24 english_trigram_counts = collections.Counter(dict(datafile('count_3l.txt')))
  25 normalised_english_trigram_counts = support.norms.normalise(english_trigram_counts)
  26
  27 with open(os.path.join(os.path.dirname(os.path.realpath(__file__)), 'words.txt'), 'r') as f:
  28     keywords = [line.rstrip() for line in f]
  29
  30
  31 def weighted_choice(d):
  32     """Generate random item from a dictionary of item counts
  33     """
  34     target = random.uniform(0, sum(d.values()))
  35     cuml = 0.0
  36     for (l, p) in d.items():
  37         cuml += p
  38         if cuml > target:
  39             return l
  40     return None
  41
  42 def random_english_letter():
  43     """Generate a random letter based on English letter counts
  44     """
  45     return weighted_choice(normalised_english_counts)
  46
  47
  48 def ngrams(text, n):
  49     """Returns all n-grams of a text
  50
  51     >>> ngrams(sanitise('the quick brown fox'), 2) # doctest: +NORMALIZE_WHITESPACE
  52     ['th', 'he', 'eq', 'qu', 'ui', 'ic', 'ck', 'kb', 'br', 'ro', 'ow', 'wn',
  53      'nf', 'fo', 'ox']
  54     >>> ngrams(sanitise('the quick brown fox'), 4) # doctest: +NORMALIZE_WHITESPACE
  55     ['theq', 'hequ', 'equi', 'quic', 'uick', 'ickb', 'ckbr', 'kbro', 'brow',
  56      'rown', 'ownf', 'wnfo', 'nfox']
  57     """
  58     return [text[i:i+n] for i in range(len(text)-n+1)]
  59
  60
  61 class Pdist(dict):
  62     """A probability distribution estimated from counts in datafile.
  63     Values are stored and returned as log probabilities.
  64     """
  65     def __init__(self, data=[], estimate_of_missing=None):
  66         data1, data2 = itertools.tee(data)
  67         self.total = sum([d[1] for d in data1])
  68         for key, count in data2:
  69             self[key] = log10(count / self.total)
  70         self.estimate_of_missing = estimate_of_missing or (lambda k, N: 1./N)
  71     def __missing__(self, key):
  72         return self.estimate_of_missing(key, self.total)
  73
  74 def log_probability_of_unknown_word(key, N):
  75     """Estimate the probability of an unknown word.
  76     """
  77     return -log10(N * 10**((len(key) - 2) * 1.4))
  78
  79 Pw = Pdist(datafile('count_1w.txt'), log_probability_of_unknown_word)
  80 Pl = Pdist(datafile('count_1l.txt'), lambda _k, _N: 0)
  81 P2l = Pdist(datafile('count_2l.txt'), lambda _k, _N: 0)
  82 P3l = Pdist(datafile('count_3l.txt'), lambda _k, _N: 0)
  83
  84 def Pwords(words):
  85     """The Naive Bayes log probability of a sequence of words.
  86     """
  87     return sum(Pw[w.lower()] for w in words)
  88
  89 def Pletters(letters):
  90     """The Naive Bayes log probability of a sequence of letters.
  91     """
  92     return sum(Pl[l.lower()] for l in letters)
  93
  94 def Pbigrams(letters):
  95     """The Naive Bayes log probability of the bigrams formed from a sequence
  96     of letters.
  97     """
  98     return sum(P2l[p] for p in ngrams(letters, 2))
  99
 100 def Ptrigrams(letters):
 101     """The Naive Bayes log probability of the trigrams formed from a sequence
 102     of letters.
 103     """
 104     return sum(P3l[p] for p in ngrams(letters, 3))
 105
 106
 107 def cosine_distance_score(text):
 108     """Finds the dissimilarity of a text to English, using the cosine distance
 109     of the frequency distribution.
 110
 111     >>> cosine_distance_score('abcabc') # doctest: +ELLIPSIS
 112     0.73777...
 113     """
 114     # return support.norms.cosine_distance(english_counts,
 115     #     collections.Counter(sanitise(text)))
 116     return 1 - support.norms.cosine_similarity(english_counts,
 117         collections.Counter(sanitise(text)))
 118
 119
 120 if __name__ == "__main__":
 121     import doctest
 122     doctest.testmod()