1 | import numpy as np |
---|
2 | from vsm.corpus import Corpus |
---|
3 | from vsm.extensions.corpusbuilders import * |
---|
4 | from vsm.extensions.corpuscleanup import apply_stoplist_len |
---|
5 | from vsm.extensions.htrc import vol_link_fn, add_link_ |
---|
6 | import os |
---|
7 | import re |
---|
8 | |
---|
9 | __all__ = ['CorpusSent', 'sim_sent_sent', 'sim_sent_sent_across', |
---|
10 | 'file_tokenize', 'file_corpus', 'dir_tokenize', 'dir_corpus', |
---|
11 | 'extend_sdd', 'extend_across'] |
---|
12 | |
---|
13 | |
---|
14 | class CorpusSent(Corpus): |
---|
15 | """ |
---|
16 | A subclass of Corpus whose purpose is to store original |
---|
17 | sentence information in the Corpus |
---|
18 | |
---|
19 | :See Also: :class: Corpus |
---|
20 | """ |
---|
21 | def __init__(self, corpus, sentences, context_types=[], context_data=[], |
---|
22 | remove_empty=False): |
---|
23 | |
---|
24 | super(CorpusSent, self).__init__(corpus, context_types=context_types, |
---|
25 | context_data=context_data, remove_empty=remove_empty) |
---|
26 | |
---|
27 | sentences = [re.sub('\n', ' ', s) for s in sentences] |
---|
28 | self.sentences = np.array(sentences) |
---|
29 | |
---|
30 | |
---|
31 | def __set_words_int(self): |
---|
32 | """ |
---|
33 | Mapping of words to their integer representations. |
---|
34 | """ |
---|
35 | self.words_int = dict((t,i) for i,t in enumerate(self.words)) |
---|
36 | |
---|
37 | |
---|
38 | def apply_stoplist(self, stoplist=[], freq=0): |
---|
39 | """ |
---|
40 | Takes a Corpus object and returns a copy of it with words in the |
---|
41 | stoplist removed and with words of frequency <= `freq` removed. |
---|
42 | |
---|
43 | :param stoplist: The list of words to be removed. |
---|
44 | :type stoplist: list |
---|
45 | |
---|
46 | :type freq: integer, optional |
---|
47 | :param freq: A threshold where words of frequency <= 'freq' are |
---|
48 | removed. Default is 0. |
---|
49 | |
---|
50 | :returns: Copy of corpus with words in the stoplist and words of |
---|
51 | frequnecy <= 'freq' removed. |
---|
52 | |
---|
53 | :See Also: :class:`Corpus` |
---|
54 | """ |
---|
55 | if freq: |
---|
56 | #TODO: Use the TF model instead |
---|
57 | |
---|
58 | print 'Computing collection frequencies' |
---|
59 | cfs = np.zeros_like(self.words, dtype=self.corpus.dtype) |
---|
60 | |
---|
61 | for word in self.corpus: |
---|
62 | cfs[word] += 1 |
---|
63 | |
---|
64 | print 'Selecting words of frequency <=', freq |
---|
65 | freq_stop = np.arange(cfs.size)[(cfs <= freq)] |
---|
66 | stop = set(freq_stop) |
---|
67 | else: |
---|
68 | stop = set() |
---|
69 | |
---|
70 | for t in stoplist: |
---|
71 | if t in self.words: |
---|
72 | stop.add(self.words_int[t]) |
---|
73 | |
---|
74 | if not stop: |
---|
75 | print 'Stop list is empty.' |
---|
76 | return self |
---|
77 | |
---|
78 | print 'Removing stop words' |
---|
79 | f = np.vectorize(lambda x: x not in stop) |
---|
80 | corpus = self.corpus[f(self.corpus)] |
---|
81 | |
---|
82 | print 'Rebuilding corpus' |
---|
83 | corpus = [self.words[i] for i in corpus] |
---|
84 | context_data = [] |
---|
85 | for i in xrange(len(self.context_data)): |
---|
86 | print 'Recomputing token breaks:', self.context_types[i] |
---|
87 | tokens = self.view_contexts(self.context_types[i]) |
---|
88 | spans = [t[f(t)].size for t in tokens] |
---|
89 | tok = self.context_data[i].copy() |
---|
90 | tok['idx'] = np.cumsum(spans) |
---|
91 | context_data.append(tok) |
---|
92 | |
---|
93 | return CorpusSent(corpus, self.sentences, context_data=context_data, |
---|
94 | context_types=self.context_types) |
---|
95 | |
---|
96 | |
---|
97 | @staticmethod |
---|
98 | def load(file): |
---|
99 | """ |
---|
100 | Loads data into a Corpus object that has been stored using |
---|
101 | `save`. |
---|
102 | |
---|
103 | :param file: Designates the file to read. If `file` is a string ending |
---|
104 | in `.gz`, the file is first gunzipped. See `numpy.load` |
---|
105 | for further details. |
---|
106 | :type file: string-like or file-like object |
---|
107 | |
---|
108 | :returns: c : A Corpus object storing the data found in `file`. |
---|
109 | |
---|
110 | :See Also: :class: Corpus, :meth: Corpus.load, :meth: numpy.load |
---|
111 | """ |
---|
112 | print 'Loading corpus from', file |
---|
113 | arrays_in = np.load(file) |
---|
114 | |
---|
115 | c = CorpusSent([], []) |
---|
116 | c.corpus = arrays_in['corpus'] |
---|
117 | c.words = arrays_in['words'] |
---|
118 | c.sentences = arrays_in['sentences'] |
---|
119 | c.context_types = arrays_in['context_types'].tolist() |
---|
120 | |
---|
121 | c.context_data = list() |
---|
122 | for n in c.context_types: |
---|
123 | t = arrays_in['context_data_' + n] |
---|
124 | c.context_data.append(t) |
---|
125 | |
---|
126 | c.__set_words_int() |
---|
127 | |
---|
128 | return c |
---|
129 | |
---|
130 | def save(self, file): |
---|
131 | """ |
---|
132 | Saves data from a CorpusSent object as an `npz` file. |
---|
133 | |
---|
134 | :param file: Designates the file to which to save data. See |
---|
135 | `numpy.savez` for further details. |
---|
136 | :type file: str-like or file-like object |
---|
137 | |
---|
138 | :returns: None |
---|
139 | |
---|
140 | :See Also: :class: Corpus, :meth: Corpus.save, :meth: numpy.savez |
---|
141 | """ |
---|
142 | |
---|
143 | print 'Saving corpus as', file |
---|
144 | arrays_out = dict() |
---|
145 | arrays_out['corpus'] = self.corpus |
---|
146 | arrays_out['words'] = self.words |
---|
147 | arrays_out['sentences'] = self.sentences |
---|
148 | arrays_out['context_types'] = np.asarray(self.context_types) |
---|
149 | |
---|
150 | for i,t in enumerate(self.context_data): |
---|
151 | key = 'context_data_' + self.context_types[i] |
---|
152 | arrays_out[key] = t |
---|
153 | |
---|
154 | np.savez(file, **arrays_out) |
---|
155 | |
---|
156 | |
---|
157 | def sent_int(self, sent): |
---|
158 | """ |
---|
159 | sent : list of strings |
---|
160 | List of sentence tokenization. |
---|
161 | The list could be a subset of existing sentence tokenization. |
---|
162 | """ |
---|
163 | tok = self.view_contexts('sentence', as_strings=True) |
---|
164 | sent_li = [] |
---|
165 | for i in xrange(len(tok)): |
---|
166 | sent_li.append(sent) |
---|
167 | keys = [i for i in xrange(len(tok)) |
---|
168 | if set(sent_li[i]).issubset(tok[i].tolist())] |
---|
169 | |
---|
170 | n = len(keys) |
---|
171 | if n == 0: |
---|
172 | raise Exception('No token fits {0}.'.format(sent)) |
---|
173 | elif n > 1: |
---|
174 | return keys |
---|
175 | return keys[0] |
---|
176 | |
---|
177 | |
---|
178 | def sim_sent_sent(ldaviewer, sent, print_len=10): |
---|
179 | """ |
---|
180 | ldaviewer : ldaviewer object |
---|
181 | sent : sentence index or sentence as a list of words |
---|
182 | |
---|
183 | Returns |
---|
184 | ------- |
---|
185 | sim_sents : numpy array |
---|
186 | (sentence index, probability) as (i, value) pair. |
---|
187 | tokenized_sents : list of arrays |
---|
188 | List containing tokenized sentences as arrays. |
---|
189 | orig_sents : list of strings |
---|
190 | List containing original sentences as strings. |
---|
191 | """ |
---|
192 | from vsm.viewer.ldagibbsviewer import LDAGibbsViewer |
---|
193 | |
---|
194 | corp = ldaviewer.corpus |
---|
195 | ind = sent |
---|
196 | if isinstance(sent, list) and isinstance(sent[0], str): |
---|
197 | ind = corp.sent_int(sent) |
---|
198 | sim_sents = ldaviewer.sim_doc_doc(ind, print_len=print_len) |
---|
199 | lc = sim_sents['doc'][:print_len] |
---|
200 | lc = [s.split(', ') for s in lc] |
---|
201 | lc = [int(s[-1]) for s in lc] |
---|
202 | |
---|
203 | # only returns print_len length |
---|
204 | tokenized_sents, orig_sents = [], [] |
---|
205 | for i in lc: |
---|
206 | tokenized_sents.append(corp.view_contexts('sentence', as_strings=True)[i]) |
---|
207 | orig_sents.append(corp.sentences[i]) |
---|
208 | |
---|
209 | return tokenized_sents, orig_sents, sim_sents |
---|
210 | |
---|
211 | |
---|
212 | def sim_sent_sent_across(ldavFrom, ldavTo, beagleviewer, sent, print_len=10, |
---|
213 | label_fn=vol_link_fn): |
---|
214 | """ |
---|
215 | ldavFrom : ldaviewer object where the sentence is from. |
---|
216 | ldavTo : ldaviewer object to find similar sentences |
---|
217 | beagleviewer : beagleviewer object is used to find |
---|
218 | similar words for words that don't exist in a different corpus. |
---|
219 | sent : sentence index of the corpus that corresponds to ldavFrom, |
---|
220 | or sentence as a list of words |
---|
221 | |
---|
222 | Returns |
---|
223 | ------- |
---|
224 | sim_sents : numpy array |
---|
225 | (sentence index, probability) as (i, value) pair. |
---|
226 | tokenized_sents : list of arrays |
---|
227 | List containing tokenized sentences as arrays. |
---|
228 | orig_sents : list of strings |
---|
229 | List containing original sentences as strings. |
---|
230 | """ |
---|
231 | from vsm.viewer.ldagibbsviewer import LDAGibbsViewer |
---|
232 | from vsm.viewer.beagleviewer import BeagleViewer |
---|
233 | |
---|
234 | def first_in_corp(corp, wordlist): |
---|
235 | """ |
---|
236 | Goes down the list to find a word that's in `corp`. |
---|
237 | Assumes there is a word in the `wordlist` that's in `corp`. |
---|
238 | """ |
---|
239 | for w in wordlist: |
---|
240 | if w in corp.words: |
---|
241 | return w |
---|
242 | |
---|
243 | corp = ldavFrom.corpus # to get sent ind |
---|
244 | ind = sent |
---|
245 | word_list = [] |
---|
246 | if isinstance(sent, list) and isinstance(sent[0], str): |
---|
247 | ind = corp.sent_int(sent) |
---|
248 | word_list = sent |
---|
249 | elif isinstance(sent, list): |
---|
250 | word_list = set() |
---|
251 | for i in sent: |
---|
252 | li = set(ldavFrom.corpus.view_contexts('sentence', |
---|
253 | as_strings=True)[i]) |
---|
254 | word_list.update(li) |
---|
255 | |
---|
256 | else: # if sent is an int index |
---|
257 | word_list = ldavFrom.corpus.view_contexts('sentence', |
---|
258 | as_strings=True)[ind].tolist() |
---|
259 | |
---|
260 | word_list = list(word_list) |
---|
261 | # Before trying ldavTo.sim_word_word, make sure all words |
---|
262 | # in the list exist in ldavTo.corpus. |
---|
263 | wl = [] |
---|
264 | for w in word_list: |
---|
265 | if w not in ldavTo.corpus.words: |
---|
266 | words = beagleviewer.sim_word_word(w)['word'] |
---|
267 | replacement = first_in_corp(ldavTo.corpus, words) |
---|
268 | wl.append(replacement) |
---|
269 | print 'BEAGLE composite model replaced {0} by {1}'.format(w, |
---|
270 | replacement) |
---|
271 | else: |
---|
272 | wl.append(w) |
---|
273 | |
---|
274 | # from ldavFrom:sent -> ldavTo:topics -> ldavTo:sent(doc) |
---|
275 | tops = ldavTo.sim_word_top(wl).first_cols[:(ldavTo.model.K/6)] |
---|
276 | tops = [int(t) for t in tops] |
---|
277 | print "Related topics: ", tops |
---|
278 | # sim_sents = ldavTo.sim_top_doc(tops, print_len=print_len, |
---|
279 | # as_strings=False) |
---|
280 | # lc = sim_sents['i'][:print_len] |
---|
281 | # tokenized_sents, orig_sents = [], [] |
---|
282 | # for i in lc: |
---|
283 | # tokenized_sents.append(ldavTo.corpus.view_contexts('sentence', as_strings=True)[i]) |
---|
284 | # orig_sents.append(ldavTo.corpus.sentences[i]) |
---|
285 | sim_sents = ldavTo.sim_top_doc(tops, print_len=print_len, |
---|
286 | label_fn=label_fn) |
---|
287 | return sim_sents |
---|
288 | |
---|
289 | |
---|
290 | def extend_sdd(args, v, print_len=10): |
---|
291 | """ |
---|
292 | Extend table resulting from sim_doc_doc with |
---|
293 | label_fn = vol_link_fn. Adds an ArgumentMap column. |
---|
294 | """ |
---|
295 | from vsm.viewer.ldagibbsviewer import LDAGibbsViewer |
---|
296 | |
---|
297 | sdd = v.sim_doc_doc(args, label_fn=vol_link_fn, print_len=print_len) |
---|
298 | table_str = sdd._repr_html_() |
---|
299 | rows = table_str.split('</tr>') |
---|
300 | |
---|
301 | rows[0] = re.sub("2", "3", rows[0]) + '</tr>' |
---|
302 | rows[1] += '<th style="text-align: center; background: #EFF2FB;">Argument\ |
---|
303 | Map</th></tr>' |
---|
304 | |
---|
305 | for i in xrange(2,len(rows)-1): |
---|
306 | a = rows[i].split('</a>, ') |
---|
307 | arg = a[1].split(',')[0] |
---|
308 | |
---|
309 | arg_map = find_arg(arg) |
---|
310 | rows[i] += '<td>{0}</td></tr>'.format(arg_map) |
---|
311 | |
---|
312 | return ''.join(rows) |
---|
313 | |
---|
314 | |
---|
315 | def extend_across(vFrom, vTo, beagle_v, args, txtFrom, txtTo, print_len=10): |
---|
316 | """ |
---|
317 | Extend table resulting from sim_sent_sent_across with |
---|
318 | the label_fn= vol_link_fn. Adds ArgumentMap and Novelty columns. |
---|
319 | """ |
---|
320 | from vsm.extensions.htrc import add_link_ |
---|
321 | |
---|
322 | across = sim_sent_sent_across(vFrom, vTo, beagle_v, args, print_len=print_len) |
---|
323 | table_str = across._repr_html_() |
---|
324 | rows = table_str.split('</tr>') |
---|
325 | |
---|
326 | rows[0] = re.sub("2", "4", rows[0]) + '</tr>' |
---|
327 | rows[1] += '<th style="text-align: center; background: #EFF2FB;">\ |
---|
328 | Argument Map</th><th style="text-align: center; background: \ |
---|
329 | #EFF2FB;">Novelty</th></tr>' |
---|
330 | |
---|
331 | for i in xrange(2,len(rows)-1): |
---|
332 | a = rows[i].split('</a>, ') |
---|
333 | arg = a[1].split(',')[0] |
---|
334 | |
---|
335 | novelty = in_ed1(arg, txtTo, txtFrom) |
---|
336 | arg_map = find_arg(novelty) |
---|
337 | |
---|
338 | # add link to novelty when it's found in the corpusFrom. |
---|
339 | if not novelty == 'new': |
---|
340 | li = novelty.split(' ') |
---|
341 | idx = int(li[0]) |
---|
342 | md = vFrom.corpus.view_metadata('sentence')[idx] |
---|
343 | link = add_link_(md['page_urls'], md['sentence_label']) |
---|
344 | li[0] = link |
---|
345 | novelty = ' '.join(li) |
---|
346 | |
---|
347 | rows[i] += '<td>{0}</td><td>{1}</td></tr>'.format(arg_map, novelty) |
---|
348 | |
---|
349 | return ''.join(rows) |
---|
350 | |
---|
351 | |
---|
352 | def in_ed1(idx, difftxt, ed1txt): |
---|
353 | """ |
---|
354 | Only for sim_sent_sent_across. |
---|
355 | Return ind from ed1txt if i has a equal match. |
---|
356 | Else return 'i'th entry in difftxt. |
---|
357 | |
---|
358 | """ |
---|
359 | path = '/var/inphosemantics/data/20131214/Washburn/vsm-data/' |
---|
360 | |
---|
361 | with open(path + ed1txt, 'r') as f1: |
---|
362 | ed1 = f1.read() |
---|
363 | ed1 = ed1.split(',') |
---|
364 | |
---|
365 | with open(path + difftxt, 'r') as f: |
---|
366 | txt = f.read() |
---|
367 | entries = txt.split(',') |
---|
368 | |
---|
369 | for i in xrange(len(entries)): |
---|
370 | if entries[i].startswith(str(idx) + ' '): |
---|
371 | if '=' in entries[i]: |
---|
372 | return ed1[i] |
---|
373 | else: |
---|
374 | prob = entries[i].split(' ')[1] |
---|
375 | return ed1[i] + ' ' + prob |
---|
376 | # didn't find idx in the table. |
---|
377 | return 'new' |
---|
378 | |
---|
379 | def find_arg(i): |
---|
380 | """ |
---|
381 | Find the arg (e.g. '422') if i is one of the analyzed args, |
---|
382 | otherwise return '' |
---|
383 | """ |
---|
384 | import json |
---|
385 | |
---|
386 | path = '/var/inphosemantics/data/20131214/Washburn/vsm-data/' |
---|
387 | |
---|
388 | if i == 'new' or '(' in i: |
---|
389 | return '' |
---|
390 | |
---|
391 | i = int(i) |
---|
392 | with open(path + 'arg_indices.json', 'r') as jsonf: |
---|
393 | indices = json.load(jsonf) |
---|
394 | |
---|
395 | for k in indices: |
---|
396 | if i in indices[k]: |
---|
397 | return str(k) |
---|
398 | return '' |
---|
399 | |
---|
400 | |
---|
401 | def file_tokenize(text): |
---|
402 | """ |
---|
403 | `file_tokenize` is a helper function for :meth:`file_corpus`. |
---|
404 | |
---|
405 | Takes a string that is content in a file and returns words |
---|
406 | and corpus data. |
---|
407 | |
---|
408 | :param text: Content in a plain text file. |
---|
409 | :type text: string |
---|
410 | |
---|
411 | :returns: words : List of words. |
---|
412 | Words in the `text` tokenized by :meth:`vsm.corpus.util.word_tokenize`. |
---|
413 | corpus_data : Dictionary with context type as keys and |
---|
414 | corresponding tokenizations as values. The tokenizations |
---|
415 | are np.arrays. |
---|
416 | """ |
---|
417 | words, par_tokens, sent_tokens, sent_orig = [], [], [], [] |
---|
418 | sent_break, par_n, sent_n = 0, 0, 0 |
---|
419 | |
---|
420 | pars = paragraph_tokenize(text) |
---|
421 | |
---|
422 | for par in pars: |
---|
423 | sents = sentence_tokenize(par) |
---|
424 | |
---|
425 | for sent in sents: |
---|
426 | w = word_tokenize(sent) |
---|
427 | words.extend(w) |
---|
428 | sent_break += len(w) |
---|
429 | sent_tokens.append((sent_break, par_n, sent_n)) |
---|
430 | sent_orig.append(sent) |
---|
431 | sent_n += 1 |
---|
432 | |
---|
433 | par_tokens.append((sent_break, par_n)) |
---|
434 | par_n += 1 |
---|
435 | |
---|
436 | idx_dt = ('idx', np.int32) |
---|
437 | sent_label_dt = ('sentence_label', np.array(sent_n, np.str_).dtype) |
---|
438 | par_label_dt = ('paragraph_label', np.array(par_n, np.str_).dtype) |
---|
439 | |
---|
440 | corpus_data = dict() |
---|
441 | dtype = [idx_dt, par_label_dt] |
---|
442 | corpus_data['paragraph'] = np.array(par_tokens, dtype=dtype) |
---|
443 | dtype = [idx_dt, par_label_dt, sent_label_dt] |
---|
444 | corpus_data['sentence'] = np.array(sent_tokens, dtype=dtype) |
---|
445 | |
---|
446 | return words, corpus_data, sent_orig |
---|
447 | |
---|
448 | |
---|
449 | def file_corpus(filename, nltk_stop=True, stop_freq=1, add_stop=None): |
---|
450 | """ |
---|
451 | `file_corpus` is a convenience function for generating Corpus |
---|
452 | objects from a a plain text corpus contained in a single string |
---|
453 | `file_corpus` will strip punctuation and arabic numerals outside |
---|
454 | the range 1-29. All letters are made lowercase. |
---|
455 | |
---|
456 | :param filename: File name of the plain text file. |
---|
457 | :type plain_dir: string-like |
---|
458 | |
---|
459 | :param nltk_stop: If `True` then the corpus object is masked |
---|
460 | using the NLTK English stop words. Default is `False`. |
---|
461 | :type nltk_stop: boolean, optional |
---|
462 | |
---|
463 | :param stop_freq: The upper bound for a word to be masked on |
---|
464 | the basis of its collection frequency. Default is 1. |
---|
465 | :type stop_freq: int, optional |
---|
466 | |
---|
467 | :param add_stop: A list of stop words. Default is `None`. |
---|
468 | :type add_stop: array-like, optional |
---|
469 | |
---|
470 | :returns: c : a Corpus object |
---|
471 | Contains the tokenized corpus built from the input plain-text |
---|
472 | corpus. Document tokens are named `documents`. |
---|
473 | |
---|
474 | :See Also: :class:`vsm.corpus.Corpus`, |
---|
475 | :meth:`file_tokenize`, |
---|
476 | :meth:`vsm.corpus.util.apply_stoplist` |
---|
477 | """ |
---|
478 | with open(filename, mode='r') as f: |
---|
479 | text = f.read() |
---|
480 | |
---|
481 | words, tok, sent = file_tokenize(text) |
---|
482 | names, data = zip(*tok.items()) |
---|
483 | |
---|
484 | c = CorpusSent(words, sent, context_data=data, context_types=names, |
---|
485 | remove_empty=False) |
---|
486 | c = apply_stoplist(c, nltk_stop=nltk_stop, |
---|
487 | freq=stop_freq, add_stop=add_stop) |
---|
488 | |
---|
489 | return c |
---|
490 | |
---|
491 | |
---|
492 | |
---|
493 | def dir_tokenize(chunks, labels, chunk_name='article', paragraphs=True): |
---|
494 | """ |
---|
495 | """ |
---|
496 | words, chk_tokens, sent_tokens, sent_orig = [], [], [], [] |
---|
497 | sent_break, chk_n, sent_n = 0, 0, 0 |
---|
498 | |
---|
499 | if paragraphs: |
---|
500 | par_tokens = [] |
---|
501 | par_n = 0 |
---|
502 | |
---|
503 | for chk, label in zip(chunks, labels): |
---|
504 | print 'Tokenizing', label |
---|
505 | pars = paragraph_tokenize(chk) |
---|
506 | |
---|
507 | for par in pars: |
---|
508 | sents = sentence_tokenize(par) |
---|
509 | |
---|
510 | for sent in sents: |
---|
511 | w = word_tokenize(sent) |
---|
512 | words.extend(w) |
---|
513 | sent_break += len(w) |
---|
514 | sent_tokens.append((sent_break, label, par_n, sent_n)) |
---|
515 | sent_orig.append(sent) |
---|
516 | sent_n += 1 |
---|
517 | |
---|
518 | par_tokens.append((sent_break, label, par_n)) |
---|
519 | par_n += 1 |
---|
520 | |
---|
521 | chk_tokens.append((sent_break, label)) |
---|
522 | chk_n += 1 |
---|
523 | else: |
---|
524 | for chk, label in zip(chunks, labels): |
---|
525 | print 'Tokenizing', label |
---|
526 | sents = sentence_tokenize(chk) |
---|
527 | |
---|
528 | for sent in sents: |
---|
529 | w = word_tokenize(sent) |
---|
530 | words.extend(w) |
---|
531 | sent_break += len(w) |
---|
532 | sent_tokens.append((sent_break, label, sent_n)) |
---|
533 | sent_orig.append(sent) |
---|
534 | sent_n += 1 |
---|
535 | |
---|
536 | chk_tokens.append((sent_break, label)) |
---|
537 | chk_n += 1 |
---|
538 | |
---|
539 | idx_dt = ('idx', np.int32) |
---|
540 | label_dt = (chunk_name + '_label', np.array(labels).dtype) |
---|
541 | sent_label_dt = ('sentence_label', np.array(sent_n, np.str_).dtype) |
---|
542 | corpus_data = dict() |
---|
543 | dtype = [idx_dt, label_dt] |
---|
544 | corpus_data[chunk_name] = np.array(chk_tokens, dtype=dtype) |
---|
545 | |
---|
546 | if paragraphs: |
---|
547 | par_label_dt = ('paragraph_label', np.array(par_n, np.str_).dtype) |
---|
548 | dtype = [idx_dt, label_dt, par_label_dt] |
---|
549 | corpus_data['paragraph'] = np.array(par_tokens, dtype=dtype) |
---|
550 | dtype = [idx_dt, label_dt, par_label_dt, sent_label_dt] |
---|
551 | corpus_data['sentence'] = np.array(sent_tokens, dtype=dtype) |
---|
552 | else: |
---|
553 | dtype = [idx_dt, label_dt, sent_label_dt] |
---|
554 | corpus_data['sentence'] = np.array(sent_tokens, dtype=dtype) |
---|
555 | |
---|
556 | return words, corpus_data, sent_orig |
---|
557 | |
---|
558 | |
---|
559 | |
---|
560 | def dir_corpus(plain_dir, chunk_name='article', paragraphs=True, word_len=2, |
---|
561 | nltk_stop=True, stop_freq=1, add_stop=None, corpus_sent=True, |
---|
562 | ignore=['.log', '.pickle', '.xml']): |
---|
563 | """ |
---|
564 | `dir_corpus` is a convenience function for generating Corpus |
---|
565 | objects from a directory of plain text files. |
---|
566 | |
---|
567 | `dir_corpus` will retain file-level tokenization and perform |
---|
568 | sentence and word tokenizations. Optionally, it will provide |
---|
569 | paragraph-level tokenizations. |
---|
570 | |
---|
571 | It will also strip punctuation and arabic numerals outside the |
---|
572 | range 1-29. All letters are made lowercase. |
---|
573 | |
---|
574 | :param plain_dir: String containing directory containing a |
---|
575 | plain-text corpus. |
---|
576 | :type plain_dir: string-like |
---|
577 | |
---|
578 | :param chunk_name: The name of the tokenization corresponding |
---|
579 | to individual files. For example, if the files are pages |
---|
580 | of a book, one might set `chunk_name` to `pages`. Default |
---|
581 | is `articles`. |
---|
582 | :type chunk_name: string-like, optional |
---|
583 | |
---|
584 | :param paragraphs: If `True`, a paragraph-level tokenization |
---|
585 | is included. Defaults to `True`. |
---|
586 | :type paragraphs: boolean, optional |
---|
587 | |
---|
588 | :param word_len: Filters words whose lengths are <= word_len. |
---|
589 | Default is 2. |
---|
590 | :type word_len: int, optional |
---|
591 | |
---|
592 | :param nltk_stop: If `True` then the corpus object is masked |
---|
593 | using the NLTK English stop words. Default is `False`. |
---|
594 | :type nltk_stop: boolean, optional |
---|
595 | |
---|
596 | :param stop_freq: The upper bound for a word to be masked on |
---|
597 | the basis of its collection frequency. Default is 1. |
---|
598 | :type stop_freq: int, optional |
---|
599 | |
---|
600 | :param corpus_sent: If `True` a CorpusSent object is returned. |
---|
601 | Otherwise Corpus object is returned. Default is `True`. |
---|
602 | :type corpus_sent: boolean, optional |
---|
603 | |
---|
604 | :param add_stop: A list of stop words. Default is `None`. |
---|
605 | :type add_stop: array-like, optional |
---|
606 | |
---|
607 | :param ignore: The list containing suffixes of files to be filtered. |
---|
608 | The suffix strings are normally file types. Default is ['.json', |
---|
609 | '.log', '.pickle']. |
---|
610 | :type ignore: list of strings, optional |
---|
611 | |
---|
612 | :returns: c : Corpus or CorpusSent |
---|
613 | Contains the tokenized corpus built from the input plain-text |
---|
614 | corpus. Document tokens are named `documents`. |
---|
615 | |
---|
616 | :See Also: :class: Corpus, :class: CorpusSent, :meth: dir_tokenize, |
---|
617 | :meth: apply_stoplist |
---|
618 | """ |
---|
619 | chunks = [] |
---|
620 | filenames = os.listdir(plain_dir) |
---|
621 | filenames = filter_by_suffix(filenames, ignore) |
---|
622 | filenames.sort() |
---|
623 | |
---|
624 | for filename in filenames: |
---|
625 | filename = os.path.join(plain_dir, filename) |
---|
626 | with open(filename, mode='r') as f: |
---|
627 | chunks.append(f.read()) |
---|
628 | |
---|
629 | words, tok, sent = dir_tokenize(chunks, filenames, chunk_name=chunk_name, |
---|
630 | paragraphs=paragraphs) |
---|
631 | names, data = zip(*tok.items()) |
---|
632 | |
---|
633 | if corpus_sent: |
---|
634 | c = CorpusSent(words, sent, context_data=data, context_types=names, |
---|
635 | remove_empty=False) |
---|
636 | else: |
---|
637 | c = Corpus(words, context_data=data, context_types=names) |
---|
638 | c = apply_stoplist_len(c, nltk_stop=nltk_stop, add_stop=add_stop, |
---|
639 | word_len=word_len, freq=stop_freq) |
---|
640 | |
---|
641 | return c |
---|