language_models.py

   1 import string
   2 import norms
   3 import random
   4 import collections
   5 import unicodedata
   6 import itertools
   7 from math import log10
   8
   9 unaccent_specials = ''.maketrans({"’": "'"})
  10
  11 def letters(text):
  12     """Remove all non-alphabetic characters from a text
  13     >>> letters('The Quick')
  14     'TheQuick'
  15     >>> letters('The Quick BROWN fox jumped! over... the (9lazy) DOG')
  16     'TheQuickBROWNfoxjumpedoverthelazyDOG'
  17     """
  18     return ''.join([c for c in text if c in string.ascii_letters])
  19
  20 def unaccent(text):
  21     """Remove all accents from letters.
  22     It does this by converting the unicode string to decomposed compatability
  23     form, dropping all the combining accents, then re-encoding the bytes.
  24
  25     >>> unaccent('hello')
  26     'hello'
  27     >>> unaccent('HELLO')
  28     'HELLO'
  29     >>> unaccent('héllo')
  30     'hello'
  31     >>> unaccent('héllö')
  32     'hello'
  33     >>> unaccent('HÉLLÖ')
  34     'HELLO'
  35     """
  36     translated_text = text.translate(unaccent_specials)
  37     return unicodedata.normalize('NFKD', translated_text).\
  38         encode('ascii', 'ignore').\
  39         decode('utf-8')
  40
  41 def sanitise(text):
  42     """Remove all non-alphabetic characters and convert the text to lowercase
  43
  44     >>> sanitise('The Quick')
  45     'thequick'
  46     >>> sanitise('The Quick BROWN fox jumped! over... the (9lazy) DOG')
  47     'thequickbrownfoxjumpedoverthelazydog'
  48     >>> sanitise('HÉLLÖ')
  49     'hello'
  50     """
  51     # sanitised = [c.lower() for c in text if c in string.ascii_letters]
  52     # return ''.join(sanitised)
  53     return letters(unaccent(text)).lower()
  54
  55
  56 def datafile(name, sep='\t'):
  57     """Read key,value pairs from file.
  58     """
  59     with open(name, 'r') as f:
  60         for line in f:
  61             splits = line.split(sep)
  62             yield [splits[0], int(splits[1])]
  63
  64 english_counts = collections.Counter(dict(datafile('count_1l.txt')))
  65 normalised_english_counts = norms.normalise(english_counts)
  66
  67 english_bigram_counts = collections.Counter(dict(datafile('count_2l.txt')))
  68 normalised_english_bigram_counts = norms.normalise(english_bigram_counts)
  69
  70 english_trigram_counts = collections.Counter(dict(datafile('count_3l.txt')))
  71 normalised_english_trigram_counts = norms.normalise(english_trigram_counts)
  72
  73 with open('words.txt', 'r') as f:
  74     keywords = [line.rstrip() for line in f]
  75
  76
  77 def weighted_choice(d):
  78         """Generate random item from a dictionary of item counts
  79         """
  80         target = random.uniform(0, sum(d.values()))
  81         cuml = 0.0
  82         for (l, p) in d.items():
  83                 cuml += p
  84                 if cuml > target:
  85                         return l
  86         return None
  87
  88 def random_english_letter():
  89         """Generate a random letter based on English letter counts
  90         """
  91         return weighted_choice(normalised_english_counts)
  92
  93
  94 def ngrams(text, n):
  95     """Returns all n-grams of a text
  96
  97     >>> ngrams(sanitise('the quick brown fox'), 2) # doctest: +NORMALIZE_WHITESPACE
  98     ['th', 'he', 'eq', 'qu', 'ui', 'ic', 'ck', 'kb', 'br', 'ro', 'ow', 'wn',
  99      'nf', 'fo', 'ox']
 100     >>> ngrams(sanitise('the quick brown fox'), 4) # doctest: +NORMALIZE_WHITESPACE
 101     ['theq', 'hequ', 'equi', 'quic', 'uick', 'ickb', 'ckbr', 'kbro', 'brow',
 102      'rown', 'ownf', 'wnfo', 'nfox']
 103     """
 104     return [text[i:i+n] for i in range(len(text)-n+1)]
 105
 106
 107 class Pdist(dict):
 108     """A probability distribution estimated from counts in datafile.
 109     Values are stored and returned as log probabilities.
 110     """
 111     def __init__(self, data=[], estimate_of_missing=None):
 112         data1, data2 = itertools.tee(data)
 113         self.total = sum([d[1] for d in data1])
 114         for key, count in data2:
 115             self[key] = log10(count / self.total)
 116         self.estimate_of_missing = estimate_of_missing or (lambda k, N: 1./N)
 117     def __missing__(self, key):
 118         return self.estimate_of_missing(key, self.total)
 119
 120 def log_probability_of_unknown_word(key, N):
 121     """Estimate the probability of an unknown word.
 122     """
 123     return -log10(N * 10**((len(key) - 2) * 1.4))
 124
 125 Pw = Pdist(datafile('count_1w.txt'), log_probability_of_unknown_word)
 126 Pl = Pdist(datafile('count_1l.txt'), lambda _k, _N: 0)
 127 P2l = Pdist(datafile('count_2l.txt'), lambda _k, _N: 0)
 128
 129 def Pwords(words):
 130     """The Naive Bayes log probability of a sequence of words.
 131     """
 132     return sum(Pw[w.lower()] for w in words)
 133
 134 def Pletters(letters):
 135     """The Naive Bayes log probability of a sequence of letters.
 136     """
 137     return sum(Pl[l.lower()] for l in letters)
 138
 139 def Pbigrams(letters):
 140     """The Naive Bayes log probability of the bigrams formed from a sequence
 141     of letters.
 142     """
 143     return sum(P2l[p] for p in ngrams(letters, 2))
 144
 145
 146 def cosine_distance_score(text):
 147     """Finds the dissimilarity of a text to English, using the cosine distance
 148     of the frequency distribution.
 149
 150     >>> cosine_distance_score('abcabc') # doctest: +ELLIPSIS
 151     0.370847405...
 152     """
 153     return norms.cosine_distance(english_counts,
 154         collections.Counter(sanitise(text)))
 155
 156
 157 if __name__ == "__main__":
 158     import doctest
 159     doctest.testmod()