def __init__(self, max_edit_distance=2, dictionary=Dictionary(),
word_counts="data/count_brown_reuters.txt",
bigram_counts="data/count_2w_brown_reuters.txt"):
""" Create a new Spellchecker
Keyword arguments:
max_edit_distance -- Maximum edit distance we should consider for words.
Anything above 2 is too computationally intensive for most systems.
dictionary -- Dictionary to use
word_counts -- Filename of word counts file
bigram_counts -- Filename of bigram counts file
"""
self.max_edit_distance = max_edit_distance
self.dictionary = dictionary
self.freq = FreqDist()
self.freq.load(open(word_counts))
self.bfreq = FreqDist()
self.bfreq.load(open(bigram_counts))
self.cmatrix = ConfusionMatrix()
class Spellchecker(object):
""" Spellchecker which can process an arbitrary text """
def __init__(self, max_edit_distance=2, dictionary=Dictionary(),
word_counts="data/count_brown_reuters.txt",
bigram_counts="data/count_2w_brown_reuters.txt"):
""" Create a new Spellchecker
Keyword arguments:
max_edit_distance -- Maximum edit distance we should consider for words.
Anything above 2 is too computationally intensive for most systems.
dictionary -- Dictionary to use
word_counts -- Filename of word counts file
bigram_counts -- Filename of bigram counts file
"""
self.max_edit_distance = max_edit_distance
self.dictionary = dictionary
self.freq = FreqDist()
self.freq.load(open(word_counts))
self.bfreq = FreqDist()
self.bfreq.load(open(bigram_counts))
self.cmatrix = ConfusionMatrix()
def prob(self, word, edit, context=None):
""" Calculate probability of candidate word given edit info and context.
Context can be one of (word_before, None), (None, word_after),
(word_before, word_after) or None.
"""
p_word = self.freq.freq(word)
p_edit = self.cmatrix.get_prob(edit) + 0.5
p_context = 0.5
if context:
word_before, word_after = context
if word_before or word_after:
if word_before:
p_context += self.bfreq[(word_before, word)]
if word_after:
p_context += self.bfreq[(word, word_after)]
return p_word * p_edit * p_context
def check(self, text):
""" Check a text for spelling errors
Returns a SpellcheckResult with information about misspelled words
and candidates.
"""
sents = [[w for w in word_tokenize(sent) if w.isalpha()] for sent in sent_tokenize(text)]
results = SpellcheckResult(sents)
# Loop through sentences, words
for sentence_pos, sentence in enumerate(sents):
for word_pos, word in enumerate(sentence):
# Skip non-alphanumeric characters -- TODO: how to deal with this?
if not word.isalpha():
continue
# Keep case mask and lowercase word
mask = CaseMask(word)
word = word.lower()
# Check if word is in our dictionary
if self.dictionary.has_word(word):
continue
# Determine context
word_before = word_after = None
if word_pos > 0:
word_before = sentence[word_pos - 1]
if word_pos < len(sentence) - 1:
word_after = sentence[word_pos + 1]
context = (word_before, word_after)
for n in range(1, self.max_edit_distance + 1):
# Make a list of candidates of distance N
# edits_meta -> (word, (meta))
word_edits = edits_meta(n, word)
# Find words that are in dictionary
dict_cands = set([w for w in word_edits if self.dictionary.has_word(w[0])])
# Find unknown words from corpus
non_words = word_edits - dict_cands
#non_words = set([w for w in word_edits if not w[0] in cands])
corpus_cands = set([w for w in non_words if self.freq.has_key(w[0])])
cands = dict_cands.union(corpus_cands)
# Calculate probabilities for each candidate
# dict_cands: word -> probability
probs = {}
for cand in cands:
w, edit = cand
p = self.prob(w, edit, context)
if probs.has_key(w):
probs[w] += p
else:
probs[w] = p
# Normalize probabilities
if probs:
probs = normalize(probs)
# Sort candidates
candidates = sorted(probs.keys(),
cmp = lambda c1, c2: cmp(probs[c1], probs[c2]),
reverse = True)
if candidates:
# Found candidates, don't try any larger edit distance
break
# Apply case mask to candidates
candidates = map(mask.apply, candidates)
# Add result
result = Correction(mask.apply(word), candidates, probs, sentence_pos, word_pos)
results.append(result)
return results
def correct(self, text):
""" Correct misspelled word in text.
Returns the corrected text. """
raise NotImplementedError
