def __init__(self, trie, keep_discarded=False, keep_filtered=False, outs=None):
"""
Initializes the callback. If specified, the training sentences are
written to @c outs; otherwise they are written to a separate output
file for each input file called <input_file_name>.ner.
@param trie the trie (not) that stores links seen in the page so far.
@param keep_discarded if @c True, the discarded sentences will be
written to a file called <input_file_name>.discarded.
@param outs the output stream to write the data to.
"""
DefaultConllCallback.__init__(self)
self._sent = SentenceData()
self._title = [] # title of the current page
self._cat = None # category of the current page
self._gold_map = {} # page: category
self._redirect_map = {} # page: redirects
self._trie = trie # the NERs mentioned on the page
self._mode = NERTrainingCallback.NO_LINK # state machine state
self._punct = set(u'.?!:') # Sentences should end with one of these
self._link_punct = set(u'.?!:,;') # To strip from the end of links
self._keep_discarded = keep_discarded
self._keep_filtered = keep_filtered
self._out = None
self._discarded = None
self._filtered = None
self._quot = misc.quotationMarks | misc.brackets
def fileStart(self, file_name):
"""Opens the output file."""
DefaultConllCallback.fileStart(self, file_name)
self._out = open(file_name + '.ner', 'w')
if self._keep_discarded:
self._discarded = open(file_name + '.discarded', 'w')
if self._keep_filtered:
self._filtered = open(file_name + '.filtered', 'w')
def fileEnd(self):
"""Closes the output file."""
if self._sent.num_train > 0:
sys.stderr.write("Written {0} sentences out of {1}, with avg length = {2} for file {3}.\n".format(
self._sent.num_train, self._sent.num_sentences,
float(self._sent.num_words) / self._sent.num_train,
self.cc_file))
self._sent.clear_statistics()
DefaultConllCallback.fileEnd(self)
self._out.close()
self._out = None
if self._discarded is not None:
self._discarded.close()
self._discarded = None
if self._filtered is not None:
self._filtered.close()
self._filtered = None
def fieldStart(self, field):
# print "NERTrainingCallback:fieldStart"
DefaultConllCallback.fieldStart(self, field.lower())
# We get the category from the title, and also need the words it
# contains before the last '(' (after which it only contains Wikipedia-
# specific information, such as disambiguation, etc.) for the link
# inference trie
if (self.cc_field.lower() == 'title'):
print "PAGE", self.cc_title.encode('utf-8')
self._title = []
self._title_pars = self.cc_title.count('(')
if self._title_pars == 0:
self._title_pars = 1
self._trie.clear()
self._cat = self._gold_map.get(self.cc_title.lower(), None)
if self._cat is not None:
self._trie.add_title(self.cc_title, self._cat)
self.__add_redirects(self.cc_title, self._cat)
elif self.cc_field.lower() == 'body':
self.first = True
self._mode = NERTrainingCallback.NO_LINK
self.tmp = []
self.sentence_words = []
