|
1
|
|
|
import itertools |
|
|
|
|
|
|
2
|
|
|
import re |
|
3
|
|
|
import logging |
|
4
|
|
|
|
|
5
|
|
|
from topik.tokenizers.simple import _simple_document |
|
6
|
|
|
from topik.tokenizers._registry import register |
|
7
|
|
|
from nltk.collocations import BigramCollocationFinder, TrigramCollocationFinder, QuadgramCollocationFinder |
|
|
|
|
|
|
8
|
|
|
from nltk.metrics.association import BigramAssocMeasures, TrigramAssocMeasures, QuadgramAssocMeasures |
|
|
|
|
|
|
9
|
|
|
|
|
10
|
|
|
# sample_corpus for doctests |
|
11
|
|
|
sample_corpus = [ |
|
|
|
|
|
|
12
|
|
|
("doc1", str(u"Frank the Swank-Tank walked his sassy unicorn, Brony," |
|
13
|
|
|
u" to prancercise class daily. Prancercise was " |
|
14
|
|
|
u"a tremendously popular pastime of sassy " |
|
15
|
|
|
u"unicorns and retirees alike.")), |
|
16
|
|
|
("doc2", str(u"Prancercise is a form of both art and fitniss, " |
|
17
|
|
|
u"originally invented by sassy unicorns. It has " |
|
18
|
|
|
u"recently been popularized by such retired " |
|
19
|
|
|
u"celebrities as Frank The Swank-Tank."))] |
|
20
|
|
|
|
|
21
|
|
|
# TODO: replace min_freqs with freq_bounds like ngrams takes. Unify format across the board. |
|
|
|
|
|
|
22
|
|
|
def _collect_ngrams(raw_corpus, top_n=10000, min_length=1, min_freqs=None, stopwords=None): |
|
23
|
|
|
"""collects bigrams and trigrams from collection of documents. Input to collocation tokenizer. |
|
24
|
|
|
|
|
25
|
|
|
bigrams are pairs of words that recur in the collection; trigrams/quadgrams are triplets/quadruplets. |
|
26
|
|
|
|
|
27
|
|
|
Parameters |
|
28
|
|
|
---------- |
|
29
|
|
|
raw_corpus : iterable of tuple of (doc_id(str/int), doc_text(str)) |
|
30
|
|
|
body of documents to examine |
|
31
|
|
|
top_n : int |
|
32
|
|
|
limit results to this many entries |
|
33
|
|
|
min_length : int |
|
34
|
|
|
Minimum length of any single word |
|
35
|
|
|
min_freqs : iterable of int |
|
36
|
|
|
threshold of when to consider a pair of words as a recognized n-gram, |
|
37
|
|
|
starting with bigrams. |
|
38
|
|
|
stopwords : None or iterable of str |
|
39
|
|
|
Collection of words to ignore as tokens |
|
40
|
|
|
|
|
41
|
|
|
Examples |
|
42
|
|
|
-------- |
|
43
|
|
|
>>> patterns = _collect_ngrams(sample_corpus, min_freqs=[2, 2, 2]) |
|
44
|
|
|
>>> patterns[0].pattern |
|
45
|
|
|
u'(frank swank|swank tank|sassy unicorns)' |
|
46
|
|
|
>>> patterns[1].pattern |
|
47
|
|
|
u'(frank swank tank)' |
|
48
|
|
|
""" |
|
49
|
|
|
|
|
50
|
|
|
# generator of documents, turn each element to its list of words |
|
51
|
|
|
doc_texts = (_simple_document(doc_text, min_length=min_length, stopwords=stopwords) |
|
|
|
|
|
|
52
|
|
|
for doc_id, doc_text in raw_corpus) |
|
|
|
|
|
|
53
|
|
|
|
|
54
|
|
|
# generator, concatenate (chain) all words into a single sequence, lazily |
|
55
|
|
|
words = itertools.chain.from_iterable(doc_texts) |
|
56
|
|
|
|
|
57
|
|
|
words_iterators = itertools.tee(words, 3) |
|
58
|
|
|
bigrams_patterns = _get_bigrams(words_iterators[0], top_n, min_freqs[0]) |
|
59
|
|
|
trigrams_patterns = _get_trigrams(words_iterators[1], top_n, min_freqs[1]) |
|
60
|
|
|
quadgrams_patterns = _get_quadgrams(words_iterators[2], top_n, min_freqs[2]) |
|
61
|
|
|
|
|
62
|
|
|
return (bigrams_patterns, trigrams_patterns, quadgrams_patterns) |
|
63
|
|
|
|
|
64
|
|
|
def _get_bigrams(words, top_n, min_freq): |
|
|
|
|
|
|
65
|
|
|
bcf = BigramCollocationFinder.from_words(iter(words)) |
|
66
|
|
|
bcf.apply_freq_filter(min_freq) |
|
67
|
|
|
bigrams = [' '.join(w) for w in bcf.nbest(BigramAssocMeasures.pmi, top_n)] |
|
|
|
|
|
|
68
|
|
|
return re.compile('(%s)' % '|'.join(bigrams), re.UNICODE) |
|
|
|
|
|
|
69
|
|
|
|
|
70
|
|
|
def _get_trigrams(words, top_n, min_freq): |
|
|
|
|
|
|
71
|
|
|
tcf = TrigramCollocationFinder.from_words(iter(words)) |
|
72
|
|
|
tcf.apply_freq_filter(min_freq) |
|
73
|
|
|
trigrams = [' '.join(w) for w in tcf.nbest(TrigramAssocMeasures.chi_sq, top_n)] |
|
|
|
|
|
|
74
|
|
|
return re.compile('(%s)' % '|'.join(trigrams), re.UNICODE) |
|
|
|
|
|
|
75
|
|
|
|
|
76
|
|
|
def _get_quadgrams(words, top_n, min_freq): |
|
|
|
|
|
|
77
|
|
|
qcf = QuadgramCollocationFinder.from_words(iter(words)) |
|
78
|
|
|
qcf.apply_freq_filter(min_freq) |
|
79
|
|
|
quadgrams = [' '.join(w) for w in qcf.nbest(QuadgramAssocMeasures.chi_sq, top_n)] |
|
|
|
|
|
|
80
|
|
|
return re.compile('(%s)' % '|'.join(quadgrams), re.UNICODE) |
|
|
|
|
|
|
81
|
|
|
|
|
82
|
|
|
|
|
83
|
|
|
def _collocation_document(text, patterns, min_length=1, stopwords=None): |
|
84
|
|
|
"""A text tokenizer that includes collocations(bigrams and trigrams). |
|
85
|
|
|
|
|
86
|
|
|
A collocation is sequence of words or terms that co-occur more often |
|
87
|
|
|
than would be expected by chance. This function breaks a raw document |
|
88
|
|
|
up into tokens based on a pre-established collection of bigrams, trigrams, |
|
89
|
|
|
and trigrams. This collection is derived from a body of many documents, and |
|
90
|
|
|
must be obtained in a prior step using the collect_ngrams |
|
91
|
|
|
function. |
|
92
|
|
|
|
|
93
|
|
|
Uses nltk.collocations.(Bi/Tri/Quad)gramCollocationFinder to |
|
94
|
|
|
find bigrams/trigrams/quadgrams. |
|
95
|
|
|
|
|
96
|
|
|
Parameters |
|
97
|
|
|
---------- |
|
98
|
|
|
text : str |
|
99
|
|
|
A single document's text to be tokenized |
|
100
|
|
|
patterns: tuple of compiled regex object to find n-grams |
|
101
|
|
|
Obtained from collect_ngrams function |
|
102
|
|
|
min_length : int |
|
103
|
|
|
Minimum length of any single word |
|
104
|
|
|
stopwords : None or iterable of str |
|
105
|
|
|
Collection of words to ignore as tokens |
|
106
|
|
|
|
|
107
|
|
|
Examples |
|
108
|
|
|
-------- |
|
109
|
|
|
>>> patterns = _collect_ngrams(sample_corpus, min_freqs=[2, 2, 2]) |
|
110
|
|
|
>>> text = sample_corpus[0][1] |
|
111
|
|
|
>>> tokenized_text = _collocation_document(text,patterns) |
|
112
|
|
|
>>> tokenized_text == [ |
|
113
|
|
|
... u'frank_swank', u'tank', u'walked', u'sassy', u'unicorn', u'brony', |
|
114
|
|
|
... u'prancercise', u'class', u'daily', u'prancercise', u'tremendously', |
|
115
|
|
|
... u'popular', u'pastime', u'sassy_unicorns', u'retirees', u'alike'] |
|
116
|
|
|
True |
|
117
|
|
|
""" |
|
118
|
|
|
text = ' '.join(_simple_document(text, min_length=min_length, stopwords=stopwords)) |
|
119
|
|
|
for pattern in patterns: |
|
120
|
|
|
text = re.sub(pattern, lambda match: match.group(0).replace(' ', '_'), text) |
|
121
|
|
|
return text.split() |
|
122
|
|
|
|
|
123
|
|
|
@register |
|
|
|
|
|
|
124
|
|
|
def ngrams(raw_corpus, min_length=1, freq_bounds=None, top_n=10000, stopwords=None): |
|
125
|
|
|
''' |
|
126
|
|
|
A tokenizer that extracts collocations (bigrams and trigrams) from a corpus |
|
127
|
|
|
according to the frequency bounds, then tokenizes all documents using those |
|
128
|
|
|
extracted phrases. |
|
129
|
|
|
|
|
130
|
|
|
Parameters |
|
131
|
|
|
---------- |
|
132
|
|
|
raw_corpus : iterable of tuple of (doc_id(str/int), doc_text(str)) |
|
133
|
|
|
body of documents to examine |
|
134
|
|
|
min_length : int |
|
135
|
|
|
Minimum length of any single word |
|
136
|
|
|
freq_bounds : list of tuples of ints |
|
137
|
|
|
Currently ngrams supports bigrams and trigrams, so this list should |
|
138
|
|
|
contain two tuples (the first for bigrams, the second for trigrams), |
|
139
|
|
|
where each tuple consists of a (minimum, maximum) corpus-wide frequency. |
|
140
|
|
|
top_n : int |
|
141
|
|
|
limit results to this many entries |
|
142
|
|
|
stopwords: None or iterable of str |
|
143
|
|
|
Collection of words to ignore as tokens |
|
144
|
|
|
|
|
145
|
|
|
Examples |
|
146
|
|
|
-------- |
|
147
|
|
|
>>> tokenized_corpora = ngrams(sample_corpus, freq_bounds=[(2, 100), (2, 100), (2, 100)]) |
|
148
|
|
|
>>> next(tokenized_corpora) == ('doc1', |
|
149
|
|
|
... [u'frank_swank', u'tank', u'walked', u'sassy', u'unicorn', u'brony', |
|
150
|
|
|
... u'prancercise', u'class', u'daily', u'prancercise', u'tremendously', |
|
151
|
|
|
... u'popular', u'pastime', u'sassy_unicorns', u'retirees', u'alike']) |
|
152
|
|
|
True |
|
153
|
|
|
''' |
|
154
|
|
|
if not freq_bounds: |
|
155
|
|
|
freq_bounds=[(50, 10000), (25, 10000), (15, 10000)] |
|
|
|
|
|
|
156
|
|
|
min_freqs = [freq[0] for freq in freq_bounds] |
|
|
|
|
|
|
157
|
|
|
# Tee corpus, since we exhaust it when finding patterns |
|
158
|
|
|
logging.debug("Collecting (bi/tri/quad)grams from corpus") |
|
159
|
|
|
corpus_iterators = itertools.tee(raw_corpus, 2) |
|
160
|
|
|
patterns = _collect_ngrams(corpus_iterators[0], top_n=top_n, min_length=min_length, min_freqs=min_freqs, |
|
161
|
|
|
stopwords=stopwords) |
|
162
|
|
|
logging.debug("Determining collocation on corpus") |
|
163
|
|
|
for doc_id, doc_text in corpus_iterators[1]: |
|
164
|
|
|
yield doc_id, _collocation_document(doc_text, patterns, min_length=min_length, stopwords=stopwords) |
|
165
|
|
|
|
ContinuumIO /
topik
Loading history...
The coding style of this project requires that you add a docstring to this code element. Below, you find an example for methods:
If you would like to know more about docstrings, we recommend to read PEP-257: Docstring Conventions.