def naivepatternHarvester(title, propertyWorder, wikipediaDump,
naivePredicateStatistics, naiveSubjectStatistics):
language = propertyWorder.getLanguage()
print('Working on `' + title + '`')
# Get triples from DBPedia
print_n_flush('Querying DBPedia...')
iri = '<' + namespaces['dbpedia'] + title + '>'
dbpediaData = fetchSubjectTriples(iri, language)
# End of DBPedia get triples
print 'OK'
sourceWiki = language
# Retrieve Wikipedia article
print_n_flush('Retrieving article from Wikipedia...')
# Obtain a pattern graph for the subject
titleLabelSingleton = getPredicateValues(dbpediaData, 'rdfs:label')
# We are pretty sure right now extendendSubjectLabels is a singleton (i.e. there is only one triple for predicate rdfs:label)
try:
assert (len(titleLabelSingleton) == 1)
except:
return
titleLabel = iter(titleLabelSingleton).next()
# We don't do this anymore
# text = getCurrentWikiArticleText(sourceWiki, title)
# We do this instead
try:
text = wikipediaDump.get_page_contents_by_title(
unidecode(titleLabel)).decode('utf-8')
except KeyError:
print_n_flush('\nCould not find a page with this title: "' +
unidecode(titleLabel) + '", skipping')
return
# End of Wikipedia article retrieving
print 'OK'
# Remove wiki markup
print_n_flush('Getting rid of wiki markup...')
# Preliminary wiki markup cleanup
text = WikiExtractor.clean(text)
# Final wiki markup cleanup (turning text into a list of section titles and paragraphs)
text = WikiExtractor.compact(text)
# End of wiki markup cleaning
print 'OK'
mergedText = u' '.join(text)
naivepatterns.naiveStatistics(title, mergedText, dbpediaData,
propertyWorder, naivePredicateStatistics,
naiveSubjectStatistics, 3, False)
def get_kb_description(self, topic_title):
raw_content = wikipedia_api_util.get_raw_page_text(topic_title)
cleaned = WikiExtractor.clean(raw_content)
compacted = WikiExtractor.compact(cleaned)
desc = ' '.join(compacted)
if desc is None or desc.strip() == '':
return topic_title
return desc
def get_kb_description(self, topic_title):
raw_content = wikipedia_api_util.get_raw_page_text(topic_title)
cleaned = WikiExtractor.clean(raw_content)
compacted = WikiExtractor.compact(cleaned)
desc = ' '.join(compacted)
if desc is None or desc.strip()=='':
return topic_title
return desc
def clean_markups(self, text):
if not text:
return ""
clean_text = WikiExtractor.clean(text)
clean_frags = WikiExtractor.compact(clean_text)
clean_html = [re.sub(HTML_TAG_REGEX, '', frag) for frag in clean_frags]
return "\n".join(clean_html) if len(clean_html) > 0 else ""
def remove_markup(self):
# First fix wiktioanry links that aren't being handled properly
# by the WikiExtractor library.
wikt = r"\[{2,}wikt:[^\|]+\|([^\]]+)\]{2,}"
self.text = re.sub(wikt, r'\1', self.text)
broken_wikt = r"{{broken wikt link\|([^\|}]+)(?:\|([^}]+))?}{2,}"
self.text = re.sub(broken_wikt, r'\1', self.text)
# Use the WikiExtractor library to finish processing
self.text = WikiExtractor.clean(self.text)
self.text = '\n'.join(WikiExtractor.compact(self.text))
def clean_markups(self, text):
if not text:
return ""
clean_text = WikiExtractor.clean(text)
clean_frags = WikiExtractor.compact(clean_text)
clean_html = [re.sub(HTML_TAG_REGEX, '', frag)
for frag in clean_frags]
return "\n".join(clean_html) if len(clean_html) > 0 else ""
def remove_markup(self):
"""Remove wiki markup leaving just the plain-text."""
# First fix wiktioanry links that aren't being handled properly
# by the WikiExtractor library.
wikt = r"\[{2,}wikt:[^\|]+\|([^\]]+)\]{2,}"
self.text = re.sub(wikt, r'\1', self.text)
broken_wikt = r"{{broken wikt link\|([^\|}]+)(?:\|([^}]+))?}{2,}"
self.text = re.sub(broken_wikt, r'\1', self.text)
# Use the WikiExtractor library to finish processing
self.text = WikiExtractor.clean(self.text)
self.text = '\n'.join(WikiExtractor.compact(self.text))
def page_handler(page):
global db_cursor
global db
try:
if 'redirect' in page:
synonym_data = {
'synonym': page['title'] + ';',
'redirect': page['redirect']
}
db_cursor.execute(
"""
UPDATE articles
SET synonyms =
IFNULL(CONCAT(synonyms, %(synonym)s), %(synonym)s)
WHERE title = %(redirect)s
""", synonym_data)
#print('Number of rows inserted: %d' % db_cursor.rowcount)
db.commit()
return
"""Write the right bits to the right files."""
#print(page['title'])
#print(page['title'])
#print("page_id :",page['id'])
#print(page['redirect'])
#print("time :",page['revisions'][-1]['timestamp'])
text = HTMLParser.HTMLParser().unescape(
page['revisions'][-1]['text'])
text = ''.join(BeautifulSoup(text).findAll(text=True))
text = WikiExtractor.clean(text)
text = ''.join(WikiExtractor.compact(text))
#print(text)
article_data = {
'id': page['id'],
'title': page['title'],
'timestamp': page['revisions'][-1]['timestamp'],
'text': text
}
print(page['id'])
db_cursor.execute(
"""
INSERT INTO articles(id, title, timestamp, text)
VALUES (%(id)s, %(title)s, %(timestamp)s, %(text)s)
""", article_data)
#print('Number of rows inserted: %d' % db_cursor.rowcount)
db.commit()
except Exception, e:
print >> sys.stderr, "invoked error. id : %s, %s" % (page['id'], e)
def page_handler(page):
global db_cursor
global db
try:
if 'redirect' in page:
synonym_data = {
'synonym': page['title'] + ';',
'redirect': page['redirect']
}
db_cursor.execute("""
UPDATE articles
SET synonyms =
IFNULL(CONCAT(synonyms, %(synonym)s), %(synonym)s)
WHERE title = %(redirect)s
""", synonym_data)
#print('Number of rows inserted: %d' % db_cursor.rowcount)
db.commit()
return
"""Write the right bits to the right files."""
#print(page['title'])
#print(page['title'])
#print("page_id :",page['id'])
#print(page['redirect'])
#print("time :",page['revisions'][-1]['timestamp'])
text = HTMLParser.HTMLParser().unescape(page['revisions'][-1]['text'])
text = ''.join(BeautifulSoup(text).findAll(text=True))
text = WikiExtractor.clean(text)
text = ''.join(WikiExtractor.compact(text))
#print(text)
article_data = {
'id': page['id'],
'title': page['title'],
'timestamp': page['revisions'][-1]['timestamp'],
'text': text
}
print(page['id'])
db_cursor.execute("""
INSERT INTO articles(id, title, timestamp, text)
VALUES (%(id)s, %(title)s, %(timestamp)s, %(text)s)
""", article_data)
#print('Number of rows inserted: %d' % db_cursor.rowcount)
db.commit()
except Exception, e:
print >> sys.stderr, "invoked error. id : %s, %s" % (page['id'], e)
from time import strftime
from jsonrpclib.jsonrpc import ProtocolError
import logging
import sys
from pairseslib.pickling import pickleDump, pickleLoad
if __name__ == '__main__':
logging.basicConfig(filename=os.path.join(cfg['home'], 'wikiread.log'),
level=logging.DEBUG,
format=cfg['logtimestampformat'])
# Open the Wikipedia dump through wikidump
wikipediaDump = wikiModel.Dump(
'/Volumes/Data/wikidump/enwiki-20130304-pages-articles.xml', False,
False)
# Instantiate the English Wikipedia worder
propertyWorder = EnglishWikipediaModule()
text = wikipediaDump.get_page_contents_by_title('Bern').decode('utf-8')
text = expandTemplates(text, propertyWorder)
# Preliminary wiki markup cleanup
text = WikiExtractor.clean(text)
# Final wiki markup cleanup (turning text into a list of section titles and paragraphs)
text = WikiExtractor.compact(text)
for line in text:
print(line.encode('utf-8'))
def sampler(title, propertyWorder, wikipediaDump, sampleSentences):
language = propertyWorder.getLanguage()
print('Working on `%s`' % title)
projectedTitle = unquote(title.replace('_',' ')).decode(encoding='utf-8')
primaryTitleLabels = {projectedTitle}
print('Going with "%s"' % (projectedTitle))
titleLabel = primaryTitleLabels.pop()
############################################################################
# Retrieve article for subject #
print_n_flush('Retrieving article from Wikipedia...')
# We do this instead, fetching the article from the wikipedia dump
strTitleLabel = unidecode(titleLabel)
try:
rawArticle = wikipediaDump.get_page_contents_by_title(strTitleLabel)
except KeyError:
message = "Could not fetch the article for " + titleLabel
logging.warning(message)
print(message)
return
article = rawArticle.decode('utf-8')
print 'OK'
### Expand relevant templates in the Wikipedia article
print_n_flush('Expanding relevant templates...')
article = removeSectionTitles(article)
article = expandTemplates(article, propertyWorder)
print 'OK'
#END# Templates expansion
### Wiki markup cleaning
print_n_flush('Getting rid of wiki markup...')
# Preliminary cleanup
article = WikiExtractor.clean(article)
# Final cleanup (turning text into a list of section titles and paragraphs)
article = WikiExtractor.compact(article)
print 'OK'
#END# Wiki markup cleaning
for paragraph in article:
""" Account for a bug in the PunktSentenceTokenizer when handling
sentence-ending marks followed by a double quote mark """
paragraph = paragraph.replace('?"', '? "')
paragraph = paragraph.replace('!"', '! "')
paragraph = paragraph.replace('."', '. "')
#TODO: Language-agnostic sentence tokenizer
sentences = tokenize_sentence(paragraph)
for sentence in sentences:
sentence = propertyWorder.adjustText(sentence)
sampleSentences.append(sentence)
def patternHarvester(title, propertyWorder, wikipediaDump):
language = propertyWorder.getLanguage()
sourceWiki = language
print('Working on `%s`' % title)
############################################################################
# Fetch triples for subject #
print_n_flush('Querying DBPedia...')
subjectIRI = expandIRI('dbpedia:' + title)
subjectTriples = fetchSubjectTriples(subjectIRI, language, False, False)
print 'OK'
# End of "Fetch triples for subject" #
############################################################################
# Obtain title for the article (i.e. primary subject name)
primaryTitleLabels = getValuesForPredicate(subjectTriples, 'rdfs:label')
# We are pretty sure right now is a singleton
# (i.e. there is one triple for predicate rdfs:label)
try:
assert (len(primaryTitleLabels) == 1)
except:
projectedTitle = title.replace('_', ' ')
primaryTitleLabels = {unicode(projectedTitle)}
message = "Could not find a primary label for %s, will try %s" % (
title, projectedTitle)
print(message)
titleLabel = primaryTitleLabels.pop()
############################################################################
# Retrieve article for subject #
print_n_flush('Retrieving article from Wikipedia...')
# We don't do this anymore
# article = getCurrentWikiArticleText(sourceWiki, title)
# We do this instead, fetching the article from the wikipedia dump
strTitleLabel = unidecode(titleLabel)
try:
rawArticle = wikipediaDump.get_page_contents_by_title(strTitleLabel)
except KeyError:
message = "Could not fetch the article for " + titleLabel
logging.warning(message)
print(message)
return
article = rawArticle.decode('utf-8')
print 'OK'
# End of "Retrieve article for subject"
############################################################################
subjectWordings = set()
subjectWordings.add(titleLabel)
# Retrieve secondary names (obtained from redirects to the primary article)
# and add them as subject labels
subjectWordings |= otherLabels(subjectIRI, language)
# Filter and get the labels for the classes the subject is an instance of
# (e.g. Los Angeles would have "city" as a label to an object for a
# rdf:type triple)
subjectClasses = getLabelsForPredicate(subjectTriples, 'rdf:type')
wordedClassLabels = set()
for classLabel in subjectClasses:
captlzd, uncptlzd = propertyWorder.getClassLabelWording(classLabel)
wordedClassLabels.add(uncptlzd)
wordedClassLabels.add(captlzd)
subjectWordings |= wordedClassLabels
### Compute and annotate wordings for triple objects
annotatedSubjectWordings = list()
# Cycle through all wordings for the subject and get an annotation
# for each one
for subjectWording in subjectWordings:
try:
(root, words, graph) = annotateText(subjectWording)
except AnnotationError:
continue
annotatedSubjectWordings.append((subjectWording, (root, words, graph)))
### Compute and annotate wordings for objects in each triple
print_n_flush('Finding and annotating wordings for triple objects...')
annotatedObjectWordings = list()
predicateOccurrences = dict()
for triple in subjectTriples:
predicate = triple['p']['value']
if predicate in ignored:
continue
if predicate not in predicateOccurrences:
predicateOccurrences[predicate] = set()
try:
objectWording = getCommonWording(triple, propertyWorder)
except CommonWordingNotFound:
# TODO: Find out if any important data types are left out
"""
if triple['p']['value'] not in notWorded:
notWorded[triple['p']['value']] = list()
notWorded[triple['p']['value']].append(triple)
pprint(triple['p']['value'] + '::' + triple['o']['value'])
pprint(triple)
"""
continue
try:
(root, words, graph) = annotateText(objectWording)
except AnnotationError:
continue
annotatedObjectWordings.append(
(objectWording, (root, words, graph, predicate)))
### END of templates expansion
print 'OK'
### Expand relevant templates in the Wikipedia article
print_n_flush('Expanding relevant templates...')
article = expandTemplates(article, propertyWorder)
print 'OK'
#END# Templates expansion
### Wiki markup cleaning
print_n_flush('Getting rid of wiki markup...')
# Preliminary cleanup
article = WikiExtractor.clean(article)
# Final cleanup (turning text into a list of section titles and paragraphs)
article = WikiExtractor.compact(article)
print 'OK'
#END# Wiki markup cleaning
# Sentence counter
i = 0
j = -1
for paragraph in article:
""" Account for a bug in the PunktSentenceTokenizer when handling
sentence-ending marks followed by a double quote mark """
paragraph = paragraph.replace('?"', '? "')
paragraph = paragraph.replace('!"', '! "')
paragraph = paragraph.replace('."', '. "')
#TODO: Language-agnostic sentence tokenizer
sentences = tokenize_sentence(paragraph)
for sentence in sentences:
sentence = propertyWorder.adjustText(sentence)
# Statistics
for ow, (owRootWord, owWords, owGraph,
predicate) in annotatedObjectWordings:
if ow in sentence:
predicateOccurrences[predicate].add(ow)
i += 1
# Get the graph for this sentence
print_n_flush('PS')
# Parse the sentence through the Stanford NLP Core Tools
try:
(sentenceR, sentenceW, sentenceG, sentence,
sentenceWData) = annotateText(sentence, True)
except AnnotationError:
continue
legalNodeIndices = map(lambda x: int(x[x.rindex("-") + 1:]),
sentenceG.nodes())
rootNode = 'ROOT-0'
# From here on, the initials "sw" refer to "subject wording"
for sw, (swRootWord, swWords, swGraph) in annotatedSubjectWordings:
try:
swRootWordIndex = matchWording(sentence, sentenceW,
sentenceG, legalNodeIndices,
sentenceWData, sw, swWords,
swGraph, swRootWord)
except ValueError as e:
"""No match found for wording in sentence"""
continue
subjectTarget = swRootWord + '-' + unicode(swRootWordIndex)
# Compute and generate subgraph for shortest path to Subject
# s1 will be the nodes from root to subject
try:
s1 = set(
shortestPathFromRoot(sentence, sentenceG,
subjectTarget))
except ShortestPathError:
continue
# From here on, the initials "ow" refer to "object wording"
# Compute and generate subgraph for shortest path to Object
# s2 is the set of nodes from root to object
for ow, (owRootWord, owWords, owGraph,
predicate) in annotatedObjectWordings:
try:
owRootWordIndex = matchWording(
sentence, sentenceW, sentenceG, legalNodeIndices,
sentenceWData, ow, owWords, owGraph, owRootWord)
except ValueError as e:
"""No match found for wording in sentence"""
continue
objectTarget = owRootWord + '-' + unicode(owRootWordIndex)
if objectTarget == subjectTarget:
""" No use for this kind of pattern """
continue
try:
s2 = set(
shortestPathFromRoot(sentence, sentenceG,
objectTarget))
except ShortestPathError:
continue
# At this point, we definitely have a pattern
# Nodes in the spanning tree comprising solely the shortest
# paths to the subject and to the object
s = s1 | s2
# S is the aforementioned spanning tree
S = nx.DiGraph(sentenceG.subgraph(s), name=predicate)
anonRoot = unicode(cfg['roottag'] + '-0')
anonSubject = unicode(cfg['subjecttag'] + '-' +
unicode(swRootWordIndex))
anonObject = unicode(cfg['objecttag'] + '-' +
unicode(owRootWordIndex))
renamings = dict()
renamings[rootNode] = anonRoot
renamings[subjectTarget] = anonSubject
renamings[objectTarget] = anonObject
entities = list()
numerals = 0
try:
for node in S.nodes():
if node not in renamings.keys():
if propertyWorder.partOfProperNoun(node):
""" The word may refer to an entity, in this
case let's abstract from the word and save a
relation for this pattern"""
index = int(node[node.rindex('-') + 1:])
anonEntity = '%s%05d-%d' % (
cfg['entitytagprefix'], len(entities),
index)
renamings[node] = anonEntity
entityWording = associatedWording(
sentence,
node,
sentenceG,
sentenceWData,
allowNestedWordingMatch=True)
entities.append(
(entityWording,
getClasses(entityWording, language)))
elif isNumeric(node):
index = int(node[node.rindex('-') + 1:])
anonNumeral = '%s%05d-%d' % (
cfg['numerictagprefix'], numerals,
index)
numerals += 1
renamings[node] = anonNumeral
except AnnotationError:
continue
# First anonymize subject, object and entities
S = nx.relabel_nodes(S, renamings)
# Remove indices as well
indexlessNodes = map(lambda word: word[0:word.rindex("-")],
S.nodes())
S = nx.relabel_nodes(S, dict(zip(S.nodes(),
indexlessNodes)))
if '' in S.nodes():
""" A bug in either the SCNLP or the python wrapper makes empty nodes out of
schwas and other unicode chars that might be used as a diacritic"""
# TODO: Find a fix for this
message = 'Invalid dependencies for this sentence: ' + sentence
logging.warning(message)
print(message)
continue
# DOT representation of the graph
pydotS = nx.to_pydot(S).to_string().encode(
encoding='UTF-8', errors='strict')
pattern = Pattern(pydotS, predicate, entities, title, sw,
ow, sentence)
try:
saveGraph(S, pattern.hash)
except (TypeError, UnicodeEncodeError):
# TODO: Fix this "TypeError: coercing to
# Unicode: need string or buffer, NoneType found" error
# also : "UnicodeEncodeError: 'ascii' codec can't encode character"
checkLog = True
logging.warning('A graph could not be saved: '
'Sentence: ' + sentence + 'Nodes: ' +
str(S.nodes()) + 'Edges: ' +
str(S.edges(data=True)))
pass
storePredicateOccurrences(title, predicateOccurrences)
