def _add_tlinks_to_fragment(self, in_fragment, tmp_fragment, out_fragment):
"""Take the links from the merged tlinks and add them into the
fragment. Based on the method with the same name in the
classifier wrapper."""
xmldoc1 = Parser().parse_file(open(in_fragment,'r'))
xmldoc2 = Parser().parse_file(open(tmp_fragment,'r'))
xmldoc1.remove_tags(TLINK)
for tlink in xmldoc2.get_tags(TLINK):
reltype = tlink.attrs[RELTYPE]
id1 = tlink.attrs.get(EVENT_INSTANCE_ID, None)
if not id1:
id1 = tlink.attrs.get(TIME_ID, None)
if not id1:
logger.warn("Could not find id1 in " + tlink.content)
id2 = tlink.attrs.get(RELATED_TO_EVENT_INSTANCE, None)
if not id2:
id2 = tlink.attrs.get(RELATED_TO_TIME, None)
if not id2:
logger.warn("Could not find id2 in " + tlink.content)
#origin = CLASSIFIER + ' ' + tlink.attrs.get(CONFIDENCE,'')
origin = tlink.attrs.get('origin','')
xmldoc1.add_tlink(reltype, id1, id2, origin)
xmldoc1.save_to_file(out_fragment)
class Slink2Tlink (TarsqiComponent):
"""Implements the S2T component of Tarsqi.
S2T takes the output of the Slinket component and applies rules to the
Slinks to create new Tlinks.
Instance variables:
NAME - a string
rules - an S2TRuleDictionary"""
def __init__(self):
"""Set component name and load rules into an S2TRuleDictionary object.
This object knows where the rules are stored."""
self.NAME = S2T
self.rules = read_rules()
def process(self, infile, outfile):
"""Apply all S2T rules to the input file.
Parses the xml file with xml_parser.Parser and converts it to a shallow tree
with converter.FragmentConverter. Then calls createTLinksFromSlinks."""
xmlfile = open(infile, "r")
self.xmldoc = Parser().parse_file(xmlfile)
self.doctree = FragmentConverter(self.xmldoc, infile).convert()
#self.print_doctree(S2T)
self.alinks = self.doctree.alink_list
self.slinks = self.doctree.slink_list
self.tlinks = self.doctree.tlink_list
#self.createTLinksFromALinks()
self.createTLinksFromSLinks()
self.xmldoc.save_to_file(outfile)
def createTLinksFromALinks(self):
"""Calls alink.lookForAtlinks to add Tlinks from Alinks. This is
rather moronic unfortunately because it will never do anything
because at the time of application there are no tlinks in the
document. Needs to be separated out and apply at a later
processing stage, after all other tlinking."""
logger.debug("Number of ALINKs in file: "+str(len(self.alinks)))
for alinkTag in self.alinks:
try:
alink = Alink(self.xmldoc, self.doctree, alinkTag)
alink.lookForAtlinks()
except:
logger.error("Error processing ALINK")
def createTLinksFromSLinks(self):
"""Calls lookForStlinks for a given Slink object."""
logger.debug("Number of SLINKs in file: "+str(len(self.slinks)))
for slinkTag in self.slinks:
try:
slink = Slink(self.xmldoc, self.doctree, slinkTag)
slink.match_rules(self.rules)
except:
logger.error("Error processing SLINK")
class Slink2Tlink(TarsqiComponent):
"""Implements the S2T component of Tarsqi.
S2T takes the output of the Slinket component and applies rules to the
Slinks to create new Tlinks.
Instance variables:
NAME - a string
rules - an S2TRuleDictionary"""
def __init__(self):
"""Set component name and load rules into an S2TRuleDictionary object.
This object knows where the rules are stored."""
self.NAME = S2T
self.rules = read_rules()
def process(self, infile, outfile):
"""Apply all S2T rules to the input file.
Parses the xml file with xml_parser.Parser and converts it to a shallow tree
with converter.FragmentConverter. Then calls createTLinksFromSlinks."""
xmlfile = open(infile, "r")
self.xmldoc = Parser().parse_file(xmlfile)
self.doctree = FragmentConverter(self.xmldoc, infile).convert()
#self.print_doctree(S2T)
self.alinks = self.doctree.alink_list
self.slinks = self.doctree.slink_list
self.tlinks = self.doctree.tlink_list
#self.createTLinksFromALinks()
self.createTLinksFromSLinks()
self.xmldoc.save_to_file(outfile)
def createTLinksFromALinks(self):
"""Calls alink.lookForAtlinks to add Tlinks from Alinks. This is
rather moronic unfortunately because it will never do anything
because at the time of application there are no tlinks in the
document. Needs to be separated out and apply at a later
processing stage, after all other tlinking."""
logger.debug("Number of ALINKs in file: " + str(len(self.alinks)))
for alinkTag in self.alinks:
try:
alink = Alink(self.xmldoc, self.doctree, alinkTag)
alink.lookForAtlinks()
except:
logger.error("Error processing ALINK")
def createTLinksFromSLinks(self):
"""Calls lookForStlinks for a given Slink object."""
logger.debug("Number of SLINKs in file: " + str(len(self.slinks)))
for slinkTag in self.slinks:
try:
slink = Slink(self.xmldoc, self.doctree, slinkTag)
slink.match_rules(self.rules)
except:
logger.error("Error processing SLINK")
def merge_tags(infile1, infile2, merged_file):
"""Merge the tags from infile1, which has all tags from the input,
with tags from infile2, which has only s, lex and TIMEX3 tags. The
lex tags are used as the pivots and it is assumed that both files
contain the same amount of lex tags."""
# create the document objects and add lex_id values to the lex tags
doc1 = Parser().parse_file(open(infile1, "r"))
doc2 = Parser().parse_file(open(infile2, "r"))
_mark_lex_tags(doc1)
_mark_lex_tags(doc2)
# get the timexes and embedded lex tags from infile2, and create
# index of the lex tags of infile1 using lex_id
extended_timexes = _get_timextags_with_contained_lextags(doc2)
lexid_to_lextag = _create_lexid_index(doc1)
for extended_timex in extended_timexes:
# get first and last document element of infile1
timex_tag = extended_timex[0]
first_lex = extended_timex[1][0]
last_lex = extended_timex[1][-1]
first_element = lexid_to_lextag[first_lex]
last_element = lexid_to_lextag[last_lex].get_closing_tag()
# get the entire sequence that is to be embedded in the timex tag
sequence = first_element.get_slice_till(last_element.id)
sequence_string = ''
for el in sequence:
sequence_string = "%s%s" % (sequence_string, el.content)
# check whether this sequence, when embedded in a tag, results
# in well-formed XML, if so, add the new timex tag to infile1,
# otherwise, ignore and print warning
try:
Parser().parse_string("<TAG>%s</TAG>" % sequence_string)
# insert opening and closing timex tags
first_element.insert_element_before(timex_tag)
last_element.insert_element_after(
XmlDocElement('</TIMEX3>', 'TIMEX3'))
except ExpatError:
logger.warn("Could not wrap TIMEX3 tag around\n\t %s" %
sequence_string)
# save the Document object of infile1 as the resulting merged file
doc1.save_to_file(merged_file)
def merge_tags(infile1, infile2, merged_file):
"""Merge the tags from infile1, which has all tags from the input,
with tags from infile2, which has only s, lex and TIMEX3 tags. The
lex tags are used as the pivots and it is assumed that both files
contain the same amount of lex tags."""
# create the document objects and add lex_id values to the lex tags
doc1 = Parser().parse_file(open(infile1,"r"))
doc2 = Parser().parse_file(open(infile2,"r"))
_mark_lex_tags(doc1)
_mark_lex_tags(doc2)
# get the timexes and embedded lex tags from infile2, and create
# index of the lex tags of infile1 using lex_id
extended_timexes = _get_timextags_with_contained_lextags(doc2)
lexid_to_lextag = _create_lexid_index(doc1)
for extended_timex in extended_timexes:
# get first and last document element of infile1
timex_tag = extended_timex[0]
first_lex = extended_timex[1][0]
last_lex = extended_timex[1][-1]
first_element = lexid_to_lextag[first_lex]
last_element = lexid_to_lextag[last_lex].get_closing_tag()
# get the entire sequence that is to be embedded in the timex tag
sequence = first_element.get_slice_till(last_element.id)
sequence_string = ''
for el in sequence:
sequence_string = "%s%s" % (sequence_string, el.content)
# check whether this sequence, when embedded in a tag, results
# in well-formed XML, if so, add the new timex tag to infile1,
# otherwise, ignore and print warning
try:
Parser().parse_string("<TAG>%s</TAG>" % sequence_string)
# insert opening and closing timex tags
first_element.insert_element_before(timex_tag)
last_element.insert_element_after(XmlDocElement('</TIMEX3>', 'TIMEX3'))
except ExpatError:
logger.warn("Could not wrap TIMEX3 tag around\n\t %s" % sequence_string)
# save the Document object of infile1 as the resulting merged file
doc1.save_to_file(merged_file)
def _add_tlinks_to_fragment(self, in_fragment, tmp_fragment, out_fragment):
"""Takes the links created by the classifier and merges them into the
input fragment."""
xmldoc1 = Parser().parse_file(open(in_fragment, 'r'))
xmldoc2 = Parser().parse_file(open(tmp_fragment, 'r'))
for tlink in xmldoc2.get_tags(TLINK):
reltype = tlink.attrs[RELTYPE]
id1 = tlink.attrs.get(EVENT_INSTANCE_ID, None)
if not id1:
id1 = tlink.attrs.get(TIME_ID, None)
if not id1:
logger.warn("Could not find id1 in " + tlink.content)
id2 = tlink.attrs.get(RELATED_TO_EVENT_INSTANCE, None)
if not id2:
id2 = tlink.attrs.get(RELATED_TO_TIME, None)
if not id2:
logger.warn("Could not find id2 in " + tlink.content)
origin = CLASSIFIER + ' ' + tlink.attrs.get(CONFIDENCE, '')
xmldoc1.add_tlink(reltype, id1, id2, origin)
xmldoc1.save_to_file(out_fragment)
def _add_tlinks_to_fragment(self, in_fragment, tmp_fragment, out_fragment):
"""Takes the links created by the classifier and merges them into the
input fragment."""
xmldoc1 = Parser().parse_file(open(in_fragment,'r'))
xmldoc2 = Parser().parse_file(open(tmp_fragment,'r'))
for tlink in xmldoc2.get_tags(TLINK):
reltype = tlink.attrs[RELTYPE]
id1 = tlink.attrs.get(EVENT_INSTANCE_ID, None)
if not id1:
id1 = tlink.attrs.get(TIME_ID, None)
if not id1:
logger.warn("Could not find id1 in " + tlink.content)
id2 = tlink.attrs.get(RELATED_TO_EVENT_INSTANCE, None)
if not id2:
id2 = tlink.attrs.get(RELATED_TO_TIME, None)
if not id2:
logger.warn("Could not find id2 in " + tlink.content)
origin = CLASSIFIER + ' ' + tlink.attrs.get(CONFIDENCE,'')
xmldoc1.add_tlink(reltype, id1, id2, origin)
xmldoc1.save_to_file(out_fragment)
def tlink_inject_with_prior_with_check(no_tlink_file, result_file, tlink_file,
original_file):
"""
"""
"""
Verb event should be a dictionary to map between
an event id and some other lemmas generated from the
initial lemma.
"""
verb_events = {}
"""
EVENT tag sample
<EVENT class="OCCURRENCE" eid="e1000028">
"""
xml_document = Parser().parse_file(open(no_tlink_file, "r"))
xmldoc_original = Parser().parse_file(open(original_file, "r"))
for element in xml_document.get_tags(EVENT):
if element.is_opening_tag():
eid = element.attrs[EID]
event_content = element.next.content
synsets_event = None
if len(wn.synsets(event_content, 'v')) > 0:
synsets_event = wn.synsets(element.next.content,
'v')[0].lemma_names
verb_morphy = wn.morphy(event_content, 'v')
verb_events[eid] = {
MORPHY_LEMMA: verb_morphy,
SYNSET_LEMMA: synsets_event
}
"""
<MAKEINSTANCE eventID="e2" polarity="POS" pos="VERB" eiid="ei2"
tense="PRESENT" aspect="PERFECTIVE">
"""
verb_event_instance = {}
for element in xml_document.get_tags(INSTANCE):
if element.is_opening_tag():
eiid = element.attrs[EIID]
eid = element.attrs[EVENTID]
if eid in verb_events:
verb_event_instance[eiid] = verb_events[eid]
"""
All TLINKs in the original document between two events
Because excepts the TLINKs parts, original and classified
documents should be identical, so they could use the same
verb_event_instance.
"""
original_ee_tlinks = {}
for element in xmldoc_original.get_tags(TLINK):
# keep track of event order here
if element.is_opening_tag():
lid = element.attrs[LID]
if EVENT_INSTANCE_ID in element.attrs:
eiid = element.attrs[EVENT_INSTANCE_ID]
if RELATED_TO_EVENT_INSTANCE in element.attrs:
reiid = element.attrs[RELATED_TO_EVENT_INSTANCE]
if RELTYPE in element.attrs:
if eiid in verb_event_instance and reiid in verb_event_instance:
original_ee_tlinks[(eiid, reiid)] = (
lid, element.attrs[RELTYPE])
with open(result_file, 'r') as result_file:
label_vote_dict = json.load(result_file)
fix_label_counter = 0
worsen_label_counter = 0
for feature_type in label_vote_dict:
for line_counter in label_vote_dict[feature_type]:
result_dict = label_vote_dict[feature_type][line_counter][
RESULT_DICT]
label_vote = label_vote_dict[feature_type][line_counter][VOTE_DICT]
ids = label_vote_dict[feature_type][line_counter][TLINK_IDS_DICT]
raw_relType = label_vote[-1][0]
"""
Have to re calculate the relType here
- Calculate P ( label | lemma_pair, result_vector ) ~ P(label)
x P ( result_vector | label )
x P ( lemma_pair | label )
"""
# if raw_relType == NORELATION or raw_relType == SIMULTANEOUS:
if raw_relType == NORELATION:
pass
else:
def check_event_pair(ids):
for id in ids:
if id[1] == TID:
return False
return True
"""
If the relation is between event pairs, we check
the narrative scheme, else, just use the raw_relType
for TLink between time and event.
"""
new_ids = {}
for id in ids:
if id[1] in [TID, EIID]:
new_ids[id[0]] = id[2]
original_relation = None
if (new_ids['0'], new_ids['1']) in original_ee_tlinks:
original_relation = original_ee_tlinks[(new_ids['0'],
new_ids['1'])][1]
elif (new_ids['1'], new_ids['0']) in original_ee_tlinks:
original_relation = reverse(
original_ee_tlinks[(new_ids['1'], new_ids['0'])][1])
"""
Eleventh try
Only consider main event pairs inter sentences
"""
if check_event_pair(ids):
probability = {}
max_label = None
max_prob = None
"""
Third approach: only consider labels inside the votes
"""
result_prob = {}
label_prob = {}
lemma_pair_prob = {}
for label in [str(label[0]) for label in label_vote]:
if not label in [BEFORE, AFTER, SIMULTANEOUS]:
continue
"""
15th try: only fix BEFORE and AFTER labels
"""
# if not label in [BEFORE, AFTER]:
# continue
# for label in [BEFORE, AFTER, SIMULTANEOUS]:
probability[label] = 1
result_prob[label] = histogram.get_probability_vector(
result_dict, label)
probability[label] *= result_prob[label]
"""
First approach: only use the morphy lemma
"""
morphy_1 = verb_event_instance[
new_ids['0']][MORPHY_LEMMA]
morphy_2 = verb_event_instance[
new_ids['1']][MORPHY_LEMMA]
# lemma_pair_prob[label] = crd.get_lemma_pair_prob((morphy_1,morphy_2,label))
# lemma_pair_prob[label] = crd.get_lemma_pair_prob_smoothing((morphy_1,morphy_2),label)
"""
Tenth approach: desperate try, multiply all of them together
"""
# probability[label] *= lemma_pair_prob[label]
"""
Done first approach
"""
"""
Second approach: use all pairs of lemmas with lemma
in corresponding two synsets
"""
lemma_pair_prob[label] = 0
synset_1 = verb_event_instance[
new_ids['0']][SYNSET_LEMMA]
synset_2 = verb_event_instance[
new_ids['1']][SYNSET_LEMMA]
if synset_1 != None and synset_2 != None:
for l_1, l_2 in itertools.product(
synset_1, synset_2):
lemma_pair_prob[
label] += crd.get_lemma_pair_prob_smoothing(
(l_1, l_2), label)
# lemma_pair_prob[label] += crd.get_lemma_pair_prob((l_1,l_2),label)
"""
Done second approach
"""
"""
Seventh try: turn off lemma pairs
"""
probability[label] *= lemma_pair_prob[label]
label_prob[label] = histogram.get_probability_label(
label)
# """
# 14th try: normalize BEFORE and AFTER labels
# """
# if label == BEFORE or label == AFTER:
# label_prob[label] = (histogram.get_probability_label (BEFORE)
# + histogram.get_probability_label (AFTER))/2
"""
13 rd try: disable label prob
"""
probability[label] *= label_prob[label]
if max_prob == None or max_prob < probability[label]:
max_prob = probability[label]
max_label = label
"""
Forth try:
if max_prob == 0, it means that all probabilities = 0
and we should follow the initialy vote
"""
if max_prob == 0:
relType = raw_relType
else:
relType = max_label
need_to_keep_track = False
if (relType == raw_relType and original_relation != None
and original_relation != relType
and original_relation
in [BEFORE, AFTER, SIMULTANEOUS]):
need_to_keep_track = True
logging.info(
'---------------DOESNT HELP----------------')
if relType != raw_relType and original_relation != None:
need_to_keep_track = True
if (original_relation == relType):
fix_label_counter += 1
if (original_relation == raw_relType):
worsen_label_counter += 1
logging.info(
'---------------MAKE CHANGE----------------')
if need_to_keep_track:
logging.info('Correct relation : %s' %
original_relation)
logging.info('Original classified : %s' % raw_relType)
logging.info('Prior classified : %s' % relType)
logging.info(morphy_1)
logging.info(morphy_2)
logging.info(synset_1)
logging.info(synset_2)
logging.info('--result_prob--')
logging.info(result_prob)
logging.info('--label_prob--')
logging.info(label_prob)
logging.info('--lemma_pair_prob--')
logging.info(lemma_pair_prob)
logging.info(probability)
logging.info('==============================')
else:
relType = raw_relType
xml_document.add_tlink(relType, new_ids['0'], new_ids['1'],
SVM_CLASSIFIER_ORIGIN)
xml_document.save_to_file(tlink_file)
return (fix_label_counter, worsen_label_counter)
class Blinker (TarsqiComponent):
"""Implements the Blinker component of Tarsqi. Blinker takes the
shallow tree implemented in the Document object and applies rules
that capture regularities between events and times as well as
between events.
Instance variables:
NAME - a string
rules - a BlinkerRuleDictionary
rule2_index - a dictionary, quick access to type 2 rules
dct - a string of the form YYYYMMDD, representing the document creation time
xmldoc - an XmlDocument, created by xml_parser.Parser
doctree - a Document, created by converter.FragmentConverter"""
def __init__(self):
"""Set component name and load rules into a BlinkerRuleDictionary
object, this object knows where the rules are stored."""
self.NAME = BLINKER
self.rules = BlinkerRuleDictionary()
self.rule2_index = {}
#self.rules.pp_ruletype(2)
self._populate_rule2_index()
def _populate_rule2_index(self):
"""Rules of type 2 (timex-signal-event) can be simply put in a
hash keyed on the signals."""
for rule in self.rules[2]:
relation = rule.get_attribute('relation')[0] # vals are now lists
signal = rule.get_attribute('signal')[0]
self.rule2_index[signal] = relation
def process(self, infile, outfile, dct):
"""Apply all Blinker rules to the input file. Parses the xml
file with xml_parser.Parser and converts it to a shallow tree
with converter.FragmentConverter. Then applies the Blinker
rules. Curently only applies rules of type 2.
Arguments
infile - an absolute path
outfile - an absolute path
No return value."""
xmlfile = open(infile, "r")
self.dct = dct
self.xmldoc = Parser().parse_file(xmlfile)
self.doctree = FragmentConverter(self.xmldoc, infile).convert(user=BLINKER)
#self.print_doctree(BLINKER)
self._run_blinker()
self.xmldoc.save_to_file(outfile)
def _run_blinker(self):
"""Apply BLinker rules to the sentences in the doctree
variable. Currently only deals with rule type 2, anchoring an
event to a timex in those cases where there is a signal (that
is, a preposition) available. New Tlinks are added just before
the closing tag of the fragment."""
self._run_timex_linking()
self._apply_event_ordering_with_signal_rules()
# variables needed for different rule types are prefixed with r<ruleNum>
r3_event1 = None
# iterate over sentences
for si in range(len(self.doctree)):
sentence = self.doctree[si]
r3_main_event = None
if _DEBUG5: print "processing sentence", si
# iterate over elements within a sentence
for i in range(len(sentence)):
element = sentence[i]
timex = element.get_timex()
event = element.get_event()
# RULE TYPE 2
if timex:
# chunk contains a timex, now try to anchor events to it
self._apply_event_anchoring_rules(sentence, timex, i)
# RULE TYPE 3
if event and element.isChunk() and element.isVerbChunk():
# the first verb event in a sentence is considered the main event
if not r3_main_event:
r3_main_event = event
# if previous sentence contained an event, create a link
if r3_event1:
r3_event2 = r3_main_event
self._apply_type3_rules(r3_event1, r3_event2)
r3_event1 = r3_event2
# else set event1
else:
r3_event1 = r3_main_event
#"""
# RULE TYPE 5
if event and element.isChunk() \
and element.isVerbChunk() \
and event.attrs['class'] == 'REPORTING':
if _DEBUG5:
print "applying type 5 rules"
self._apply_type5_rules(sentence, event, i)
#"""
# R3: if no main event in sentence
if not r3_main_event:
r3_event1 = None
def _run_timex_linking(self):
"""Apply the rules that govern relations between TIMEX3 tags. Only
applies to TIMEX3 tags with a VAL attribute equal to DATE."""
timexes = [timex for timex in self.xmldoc.get_tags(TIMEX)
if timex.attrs['TYPE'] == 'DATE']
for t in timexes:
if t.attrs.get('VAL', None) is None:
logger.warn("Missing VAL: %s" % t.get_content())
for i in range(len(timexes)):
for j in range(len(timexes)):
if i < j:
try:
self._create_timex_link(timexes[i], timexes[j])
except Exception:
logger.error("Error in Timex Linking:\n%s\n%s" % \
(timexes[i].get_content(),
timexes[j].get_content()))
def _create_timex_link(self, timex1, timex2):
"""Try to create a TLINK between two timexes."""
creation_year = self.dct[0:4]
date1 = timex1.attrs.get('VAL', None)
date2 = timex2.attrs.get('VAL', None)
if date1 is None or date2 is None:
return
date1 = fix_timex_val(date1)
date2 = fix_timex_val(date2)
tid1 = timex1.attrs['tid']
tid2 = timex2.attrs['tid']
comment = "Blinker - Timex Linking"
if date1 == date2:
if date1 not in ('PAST_REF', 'FUTURE_REF'):
self.xmldoc.add_tlink('IDENTITY', tid1, tid2, comment)
else:
rel = compare_date(date1, date2, creation_year)
if rel != 'IDENTITY':
self.xmldoc.add_tlink(rel, tid1, tid2, comment)
def _apply_type3_rules(self, event1, event2):
""" Creates a TLINK between two main events """
if _DEBUG3:
print event1.dtrs[0].getText(), event2.dtrs[0].getText()
print event1.dtrs[0].getText(), event1.attrs['class'], \
event1.attrs['tense'], event1.attrs['aspect']
print event2.dtrs[0].getText(), event2.attrs['class'], \
event2.attrs['tense'], event2.attrs['aspect']
for i in range(len(self.rules[3])):
rule = self.rules[3][i]
if _DEBUG3:
print "RULE %s:" % (rule.rule_number)
print rule.attrs['arg1.class'], rule.attrs['arg1.tense'], rule.attrs['arg1.aspect']
print rule.attrs['arg2.class'], rule.attrs['arg2.tense'], rule.attrs['arg2.aspect']
# see tags.py and library.timeMLspec.py for attribute names
if event1.attrs['class'] in rule.attrs['arg1.class'] and \
event2.attrs['class'] in rule.attrs['arg2.class'] and \
event1.attrs['tense'] in rule.attrs['arg1.tense'] and \
event2.attrs['tense'] in rule.attrs['arg2.tense'] and \
event1.attrs['aspect'] in rule.attrs['arg1.aspect'] and \
event2.attrs['aspect'] in rule.attrs['arg2.aspect']:
rel = rule.attrs['relation'][0]
self.xmldoc.add_tlink( rel,
event1.attrs[EIID],
event2.attrs[EIID],
"Blinker - Type 3 (rule %s)" % rule.rule_number)
if _DEBUG3: print "RULE %s fired!" % rule.rule_number
# apply the first matching rule
return
def _apply_type5_rules(self, sentence, event1, position):
""" Creates TLINKs between the reporting event and reported events
Takes as arguments sentence, reporting event constituent, and
position of that constituent within the sentence list"""
# filter out rules with wrong tense
applicable_rules = self.rules[5][:]
applicable_rules = [rule for rule in applicable_rules
if event1.attrs['tense'] in rule.attrs['arg1.tense']]
# reset to opposite when quote is encountered
direct = 'INDIRECT'
# forward
if _DEBUG5:
print "inside rule application function"
sentence.pretty_print()
for i in range(position+1, len(sentence)):
if _DEBUG5: print "processing element", i
element = sentence[i]
# quote
if element.isToken() and element.getText() in QUOTES:
if direct == 'DIRECT': direct = 'INDIRECT'
if direct == 'INDIRECT': direct = 'DIRECT'
# event
event2 = element.get_event()
if event2 and element.isChunk() and element.isVerbChunk():
current_rules = applicable_rules[:]
current_rules = [rule for rule in current_rules if direct in rule.attrs['sentType']]
if _DEBUG5:
print event1.dtrs[0].getText(), event2.dtrs[0].getText()
print event1.dtrs[0].getText(), event1.attrs['class'], \
event1.attrs['tense'], event1.attrs['aspect']
print event2.dtrs[0].getText(), event2.attrs['class'], \
event2.attrs['tense'], event2.attrs['aspect']
for rule in current_rules:
# if attribute not set in the rule, accept any value
for att in ['class', 'tense', 'aspect']:
if not rule.attrs.has_key('arg2.'+att):
rule.attrs['arg2.'+att] = [event2.attrs[att]]
if _DEBUG5:
print "RULE %s (%s):" % (rule.rule_number, rule.attrs['sentType'][0])
print rule.attrs['arg1.class'], rule.attrs['arg1.tense'], \
rule.attrs['arg1.aspect']
print rule.attrs['arg2.class'], rule.attrs['arg2.tense'], \
rule.attrs['arg2.aspect']
# check that specified values match
if event2.attrs['class'] in rule.attrs['arg2.class'] and \
event2.attrs['tense'] in rule.attrs['arg2.tense'] and \
event2.attrs['aspect'] in rule.attrs['arg2.aspect']:
rel = rule.attrs['relation'][0]
self.xmldoc.add_tlink( rel,
event1.attrs['eiid'],
event2.attrs['eiid'],
"Blinker - Type 5 (rule %s)" % rule.rule_number)
if _DEBUG5: print "RULE %s fired!" % rule.rule_number
# apply the first matching rule
return
# backward
# - this creates multiple links for REPORTING to REPORTING
# - may add the appropriate rules to the rule file instead
direct = 'INDIRECT'
for i in range(position-1, -1, -1): # ..,3,2,1,0
if _DEBUG5: print "processing element", i
element = sentence[i]
# quote
if element.isToken() and element.getText() in QUOTES:
if direct == 'DIRECT': direct = 'INDIRECT'
if direct == 'INDIRECT': direct = 'DIRECT'
# event
event2 = element.get_event()
if event2 and element.isChunk() and element.isVerbChunk():
current_rules = applicable_rules[:]
current_rules = [rule for rule in current_rules if direct in rule.attrs['sentType']]
if _DEBUG5:
print event1.dtrs[0].getText(), event2.dtrs[0].getText()
print event1.dtrs[0].getText(), event1.attrs['class'], \
event1.attrs['tense'], event1.attrs['aspect']
print event2.dtrs[0].getText(), event2.attrs['class'], \
event2.attrs['tense'], event2.attrs['aspect']
print "Applying rules for sentence type:", direct, len(current_rules), "rules"
for rule in current_rules:
# if attribute not set in the rule, accept any value
for att in ['class', 'tense', 'aspect']:
if not rule.attrs.has_key('arg2.'+att):
rule.attrs['arg2.'+att] = [event2.attrs[att]]
if _DEBUG5:
print "RULE %s (%s):" % (rule.rule_number, rule.attrs['sentType'][0])
print rule.attrs['arg1.class'], rule.attrs['arg1.tense'], \
rule.attrs['arg1.aspect']
print rule.attrs['arg2.class'], rule.attrs['arg2.tense'], \
rule.attrs['arg2.aspect']
# check that specified values match
if event2.attrs['class'] in rule.attrs['arg2.class'] and \
event2.attrs['tense'] in rule.attrs['arg2.tense'] and \
event2.attrs['aspect'] in rule.attrs['arg2.aspect']:
rel = rule.attrs['relation'][0]
self.xmldoc.add_tlink( rel,
event1.attrs['eiid'],
event2.attrs['eiid'],
"Blinker - Type 5 (rule %s)" % rule.rule_number)
if _DEBUG5: print "RULE %s fired!" % rule.rule_number
# apply the first matching rule
return
def _apply_event_anchoring_rules(self, sentence, timex, i):
"""Anchor events to a given timex that occurs in the sentence
at index i. The method proceeds by looking for some simple
syntactic patterns with and without prepositions. If a pattern
with a preposition occurs, then the preposition is looked up
in self.rule2_index. If no signal is found, then the default
INCLUDES rule will apply (rule 1), this is not yet
implemented."""
# NOTES:
# - Need to add some kind of confidence measures
# PATTERN: [TIMEX EVENT]
# Or, more precisely, an event in the same chunk as the timex
# Example: "October elections"
event = sentence[i].get_event()
if event:
eiid = event.attrs[EIID]
tid = timex.attrs[TID]
self.xmldoc.add_tlink('IS_INCLUDED', eiid, tid, "Blinker - Type 1")
return
# Pattern: [CHUNK-WITH-EVENT] Prep [CHUNK-WITH-TIMEX]
if i > 1:
event = sentence[i-2].get_event()
if sentence[i-1].isPreposition() and event:
signal = sentence[i-1].getText().lower()
rel = self.rule2_index.get(signal)
eiid = event.attrs[EIID]
tid = timex.attrs[TID]
if _DEBUG2:
print "FOUND: [%s] %s [%s] --> %s" % \
(event.dtrs[0].getText(), signal, timex.getText(), rel)
self.xmldoc.add_tlink(rel, eiid, tid, "Blinker - Type 2 (%s)" % signal)
return
# Pattern: [CHUNK-WITH-VERBAL-EVENT] [CHUNK-WITH_TIMEX]
if i > 0:
previous_chunk = sentence[i-1]
if previous_chunk.isVerbChunk():
event = previous_chunk.get_event()
if event:
#if event.attrs[POL] != 'NEG':
eiid = event.attrs[EIID]
tid = timex.attrs[TID]
self.xmldoc.add_tlink('IS_INCLUDED', eiid, tid, "Blinker - Type 1a")
return
def _apply_event_ordering_with_signal_rules(self):
"""Some more rules without using any rules, basically a placeholder
for event ordering rules that use a signal."""
signal_mapping = {
'after': 'AFTER',
'before': 'BEFORE',
'during': 'DURING'
}
for si in range(len(self.doctree)):
sentence = self.doctree[si]
for i in range(len(sentence)):
try:
#print sentence[i:i+4]
(VG1, Prep, NG, VG2) = sentence[i:i+4]
event1 = VG1.get_event()
event2 = VG2.get_event()
# Pattern: [VG +Event] [Prep] [NG -Event] [VG +Event]
if event1 and VG1.isVerbChunk() and \
Prep.isPreposition() and \
NG.isNounChunk() and not NG.get_event() and \
event2 and VG2.isVerbChunk():
#print "[VG +Event] [Prep] [NG -Event] [VG +Event]"
#print Prep
prep_token = Prep.getText().lower()
#print prep_token
rel = signal_mapping.get(prep_token)
#print rel
if rel:
#print 'adding tlink'
eiid1 = event1.attrs[EIID]
eiid2 = event2.attrs[EIID]
self.xmldoc.add_tlink(rel, eiid1, eiid2, "Blinker - Event:Signal:Event")
except:
pass
def incorporate_tlink_with_prior_correction(svm_histogram_file,
no_tlink_directory,
result_directory, tlink_directory,
histogram_class,
correction_method):
histogram = histogram_class.load_histogram(svm_histogram_file)
result_files = glob.glob(
os.path.join(result_directory, '*%s' % RESULT_SUFFIX))
total_fix_label_counter = 0
total_worsen_label_counter = 0
prior = histogram.get_prior()
"""
Should be converging condition
"""
all_relation_collect = {}
for i in xrange(2):
print i
logging.info('======================RUN========================')
logging.info(i)
for feature_type in prior:
logging.info(feature_type)
logging.info(prior[feature_type])
label_prob_collect = {}
for result_file in result_files:
new_relation_collect = correction_method(result_file, prior)
for feature_type in new_relation_collect:
if feature_type not in label_prob_collect:
label_prob_collect[feature_type] = {}
label_prob_collect[feature_type][
result_file] = new_relation_collect[feature_type]
all_relation_collect[i] = label_prob_collect
new_prior = {}
new_prior_count = defaultdict(int)
"""
update prior here
Update prior based on the posterior received
from classifying each sample on test data.
"""
for feature_type in label_prob_collect:
new_prior[feature_type] = defaultdict(float)
for result_file in label_prob_collect[feature_type]:
for line_counter in label_prob_collect[feature_type][
result_file]:
probability, raw_relType, id_0, id_1 =\
label_prob_collect[feature_type][result_file][line_counter]
new_prior_count[feature_type] += 1
max_value = max(probability.values())
#
tf = sorted(probability.values())
if tf[0] == tf[1]:
print probability
for label in probability:
new_prior[feature_type][label] += probability[label]
for label in new_prior[feature_type]:
new_prior[feature_type][label] /= new_prior_count[feature_type]
"""
Second way of update prior:
Update prior on the real label assigned to each
sample each iteration.
It's the extremity version of first prior approach,
by actually assign for each label a posterior of 1
for the most likely label.
"""
# for feature_type in label_prob_collect:
# new_prior[feature_type] = defaultdict(float)
# for result_file in label_prob_collect[feature_type]:
# for line_counter in label_prob_collect[feature_type][result_file]:
# probability, raw_relType, id_0, id_1 =\
# label_prob_collect[feature_type][result_file][line_counter]
# new_prior_count[feature_type] += 1
# max_value = max(probability.values())
#
# tf = sorted(probability.values())
# if tf[0] == tf[-1]:
# new_prior[feature_type][raw_relType] += 1
# else:
# for label in probability:
# if probability[label] == max_value:
# new_prior[feature_type][label] += 1
# break
# for label in new_prior[feature_type]:
# new_prior[feature_type][label] /= new_prior_count[feature_type]
prior = new_prior
"""
Check this part first
"""
# return
result_file_collect = {}
for feature_type in label_prob_collect:
if feature_type in []:
for result_file in all_relation_collect[0][feature_type]:
if result_file not in result_file_collect:
result_file_collect[result_file] = {}
result_file_collect[result_file][feature_type] =\
all_relation_collect[0][feature_type][result_file]
else:
for result_file in label_prob_collect[feature_type]:
if result_file not in result_file_collect:
result_file_collect[result_file] = {}
result_file_collect[result_file][feature_type] =\
label_prob_collect[feature_type][result_file]
for result_file in result_file_collect:
rel_filename = result_file[result_file.rindex(os.path.sep) + 1:]
no_tlink_file = os.path.join(
no_tlink_directory,
'%s%s' % (rel_filename[:-len(RESULT_SUFFIX)], NO_TLINK_SUFFIX))
tlink_file = os.path.join(
tlink_directory,
'%s%s' % (rel_filename[:-len(RESULT_SUFFIX)], ADD_TLINK_SUFFIX))
xml_document = Parser().parse_file(open(no_tlink_file, "r"))
for feature_type in result_file_collect[result_file]:
for line_counter in result_file_collect[result_file][feature_type]:
(probability, raw_relType, id_0, id_1) =\
result_file_collect[result_file][feature_type][line_counter]
tf = sorted(probability.values())
if tf[0] == tf[-1]:
"""
I should fix the label here to the label
guessed by vote dict
"""
relType = raw_relType
else:
if tf[-1] == tf[-2]:
print probability
relType = sorted(probability.items(),
key=lambda x: x[1])[-1][0]
if relType != NORELATION:
xml_document.add_tlink(relType, id_0, id_1,
SVM_CLASSIFIER_ORIGIN)
xml_document.save_to_file(tlink_file)