def convert(self, id, result, format):
assert format == "csv"
_int = lambda x: None if x is None else int(x)
naf = KafNafParser(BytesIO(result.encode("utf-8")))
deps = {dep.get_to(): (dep.get_function(), dep.get_from())
for dep in naf.get_dependencies()}
tokendict = {token.get_id(): token for token in naf.get_tokens()}
s = StringIO()
w = csv.writer(s)
w.writerow(["id", "token_id", "offset", "sentence", "para", "word", "term_id",
"lemma", "pos", "pos1", "parent", "relation"])
for term in naf.get_terms():
tokens = [tokendict[id] for id in term.get_span().get_span_ids()]
for token in tokens:
tid = term.get_id()
pos = term.get_pos()
pos1 = POSMAP[pos]
row = [id, token.get_id(), _int(token.get_offset()), _int(token.get_para()), token.get_text(),
tid, term.get_lemma(), pos, pos1]
if tid in deps:
rel, parent = deps[tid]
row += [parent, rel.split("/")[-1]]
else:
row += [None, None]
w.writerow(row)
return s.getvalue()
def test_create_terms():
"""
Can we create_terms via the create_{term,token} functions?
"""
naf = KafNafParser(type="NAF")
sent=1; offset=0
input = [(u'dit', u'dit', u'O', u'VNW'),
(u'is', u'zijn', u'V', u'WW'),
(u'een', u'een', u'D', u'LID'),
(u'test', u'test', u'N', u'N')]
offset = 0
for (word, lemma, pos, morph) in input:
token = naf.create_wf(word, 1, offset)
offset += len(word)
term = naf.create_term(lemma, pos, morph, [token])
tokens = {t.get_id(): t for t in naf.get_tokens()}
assert_equal(len(tokens), 4)
result = {}
for term in naf.get_terms():
for token_id in term.get_span().get_span_ids():
token = tokens[token_id]
result[term.get_id()] = (token.get_text(), term.get_lemma(),
term.get_pos(), term.get_morphofeat())
result = [result[tid] for tid in sorted(result.keys())]
assert_equal(input, result)
def from_naf(self, article, naf):
def _int(x):
return None if x is None else int(x)
naf = KafNafParser(BytesIO(naf.encode("utf-8")))
deps = {dep.get_to(): (dep.get_function(), dep.get_from())
for dep in naf.get_dependencies()}
tokendict = {token.get_id(): token for token in naf.get_tokens()}
for term in naf.get_terms():
tokens = [tokendict[id] for id in term.get_span().get_span_ids()]
for token in tokens:
tid = term.get_id()
tok = {"aid": article,
"token_id": token.get_id(),
"offset": _int(token.get_offset()),
"sentence": _int(token.get_sent()),
"para": _int(token.get_para()),
"word": token.get_text(),
"term_id": tid,
"lemma": term.get_lemma(),
"pos": term.get_pos()}
if tid in deps:
rel, parent = deps[tid]
tok['parent'] = parent
tok['relation'] = rel.split("/")[-1]
yield tok
def process_single_file(self,file):
try:
xml_obj = KafNafParser(file)
except:
print>>sys.stderr,'Error parsing',file,': skipped'
return
print>>sys.stderr,'Processing file', os.path.basename(file), 'Type:',xml_obj.get_type()
self.langs[xml_obj.get_language()] += 1
sentences = []
current_sent = []
this_sent = None
pos_for_wid = {} ## For each token id (wid) the pos of it
for term in xml_obj.get_terms():
w_ids = term.get_span().get_span_ids()
pos = term.get_pos()
for wid in term.get_span().get_span_ids():
pos_for_wid[wid] = pos
for token in xml_obj.get_tokens():
wid = token.get_id()
value = token.get_text()
if self.convert_to_lowercase:
value = value.lower()
if value in self.punctuation:
value = 'PUN'
if value == '*':
value = 'STAR'
sentence = token.get_sent()
if this_sent is not None and sentence != this_sent: ## There is a new sent
sentences.append(current_sent)
current_sent = []
current_sent.append((wid,value))
this_sent = sentence
## Add the last sentence as well
sentences.append(current_sent)
for sentence in sentences:
if self.include_sentence_delimiters:
sentence.insert(0,('xxx','<S>'))
sentence.append(('xxx','</S>'))
for idx in range(0,len(sentence)):
for ngramlen in range(self.min_ngram_len,self.max_ngram_len+1):
file_desc = self.get_file_desc_for_ngram(ngramlen)
start = idx
end = start + ngramlen
if end <= len(sentence):
this_ngram = '\t'.join(value for wid, value in sentence[start:end])
this_ngram_pos = '\t'.join(pos_for_wid.get(wid,'X') for wid, value in sentence[start:end])
file_desc.write(this_ngram.encode('utf-8')+'\t'+DELIMITER+'\t'+this_ngram_pos+'\n')
def test_create_terms():
"""
Can we create_terms via the create_{term,token} functions?
"""
naf = KafNafParser(type="NAF")
sent = 1
offset = 0
input = [(u'dit', u'dit', u'O', u'VNW'), (u'is', u'zijn', u'V', u'WW'),
(u'een', u'een', u'D', u'LID'), (u'test', u'test', u'N', u'N')]
offset = 0
for (word, lemma, pos, morph) in input:
token = naf.create_wf(word, 1, offset)
offset += len(word)
term = naf.create_term(lemma, pos, morph, [token])
tokens = {t.get_id(): t for t in naf.get_tokens()}
assert_equal(len(tokens), 4)
result = {}
for term in naf.get_terms():
for token_id in term.get_span().get_span_ids():
token = tokens[token_id]
result[term.get_id()] = (token.get_text(), term.get_lemma(),
term.get_pos(), term.get_morphofeat())
result = [result[tid] for tid in sorted(result.keys())]
assert_equal(input, result)
def from_naf(self, naf):
naf = KafNafParser(BytesIO(naf.encode("utf-8")))
tokendict = {token.get_id(): token for token in naf.get_tokens()}
for term in naf.get_terms():
tokens = [tokendict[id] for id in term.get_span().get_span_ids()]
for token in tokens:
yield {"aid": article.pk,
"token_id": token.get_id(),
"offset": token.get_offset(),
"sentence": token.get_sent(),
"para": token.get_para(),
"word": token.get_text(),
"term_id": term.get_id(),
"lemma": term.get_lemma(),
"pos": term.get_pos()}
def read_training_data(file_name):
"""
read kaf/naf and matches the aspects with the words
"""
parser = KafNafParser(PATH_ANNOTATED_DATA + file_name)
terms = list(parser.get_terms())
# create token dictionairy containing naf info
tokens_container = dict()
for token_el in parser.get_tokens():
token_node = token_el.node
token_id = token_node.get('wid').replace('w', 't')
token_info = token_node.attrib
tokens_container[token_id] = token_info
properties = list(parser.get_properties())
handled_properties, term_dict = handle_properties(properties, terms,
tokens_container)
return terms, properties, handled_properties, term_dict, tokens_container
def convert(self, id, result, format):
assert format == "csv"
naf = KafNafParser(BytesIO(result.encode("utf-8")))
memo = self._csv_memo(naf)
tokendict = {token.get_id(): token for token in naf.get_tokens()}
s = StringIO()
w = csv.writer(s)
w.writerow(self._csv_header())
for term in naf.get_terms():
tokens = [tokendict[id] for id in term.get_span().get_span_ids()]
for token in tokens:
tid = term.get_id()
pos = term.get_pos()
pos1 = POSMAP[pos]
row = [id] + list(self._csv_row(memo, term, token))
w.writerow(row)
return s.getvalue()
def process_file(this_file,token_freq):
xml_obj = KafNafParser(this_file)
print>>sys.stderr,'Processing file',this_file
token_for_wid = {}
order_for_wid = {}
opinion_expressions = []
opinion_targets = []
whole_text = ' '
for n, token in enumerate(xml_obj.get_tokens()):
text = token.get_text().lower()
token_freq[text] += 1
token_for_wid[token.get_id()] = text
order_for_wid[token.get_id()] = n
whole_text += text + ' '
wids_for_tid = {}
lemma_for_wid = {}
pos_for_wid = {}
for term in xml_obj.get_terms():
tid = term.get_id()
wids = term.get_span().get_span_ids()
wids_for_tid[tid] = wids
for wid in wids:
lemma_for_wid[wid] = term.get_lemma()
pos_for_wid[wid] = term.get_pos()
##Properties!
aspects = [] ## [(label,term_span)...]
for property in xml_obj.get_properties():
for refs in property.get_references():
for span in refs:
aspects.append((property.get_type(),span.get_span_ids()))
already_counted = {EXP:set(), TAR:set()}
for opinion in xml_obj.get_opinions():
for this_type, opinion_obj in [(EXP,opinion.get_expression()),(TAR,opinion.get_target())]:
if this_type is EXP and opinion_obj.get_polarity()=='NON-OPINIONATED':
continue
if opinion_obj is not None:
span = opinion_obj.get_span()
if span is not None:
list_wids = []
for tid in span.get_span_ids():
list_wids.extend(wids_for_tid.get(tid,[]))
list_wids.sort(key=lambda wid: order_for_wid[wid]) ##Sorted according the the order of the tokens
string_wids = '#'.join(list_wids)
opinion_tokens = ' '.join( token_for_wid[wid] for wid in list_wids)
opinion_lemmas = ' '.join( lemma_for_wid[wid] for wid in list_wids)
opinion_pos = ' '.join( pos_for_wid[wid] for wid in list_wids)
if string_wids not in already_counted[this_type]:
if this_type == EXP:
polarity = (opinion_obj.get_polarity()).lower()
opinion_expressions.append((opinion_tokens,polarity,opinion_lemmas,opinion_pos))
else:
##Calculate the aspect type
possible_aspects = []
target_ids = span.get_span_ids()
for aspect_label, aspect_span in aspects:
num_in_common = len(set(target_ids) & set(aspect_span))
if num_in_common != 0:
possible_aspects.append((aspect_label,num_in_common,len(aspect_span)))
aspect_for_target = 'unknown'
if len(possible_aspects) != 0:
##Sorting by the number in common first, and by the lengtgh of the aspect secondly
aspect_for_target = sorted(possible_aspects,key=lambda t: (t[1],t[2]), reverse=True)[0][0]
opinion_targets.append((opinion_tokens,aspect_for_target, opinion_lemmas,opinion_pos))
already_counted[this_type].add(string_wids)
del xml_obj
print>>sys.stderr,'\tNumber of opinion expressions:',len(opinion_expressions)
print>>sys.stderr,'\tNumber of opinion targets:',len(opinion_targets)
print>>sys.stderr,'\tNumber of characters of the text:',len(whole_text)
return opinion_expressions, opinion_targets, whole_text
def create_training_sentences(folder_tag_in, folder_kaf_in, opinion_layers,
non_opinion, folder_out):
#Remove the outputfolder if exists and create it again
if os.path.exists(folder_out):
shutil.rmtree(folder_out)
os.mkdir(folder_out)
total_sents_opi = total_sents_no_opi = 0
for tag_file in glob.glob(os.path.join(folder_tag_in, '*.tag')):
basename = os.path.basename(tag_file).replace('.tag', '')
kaf_file = os.path.join(folder_kaf_in, basename + '.kaf')
if os.path.exists(kaf_file):
##From the tag file we extract the token ids for opinions and for non opinionated
opinion_wids = set() #token ids annotated as opinions
no_opinion_wids = set() #token ids annotated as no opinions
fd = open(tag_file, 'rb')
for line in fd:
fields = line.strip().split('\t')
wid = fields[0]
for opinion_idx in opinion_layers:
if fields[opinion_idx] == 'Opinion':
opinion_wids.add(wid)
if non_opinion is not None and fields[
non_opinion] == 'NON-OPINIONATED':
no_opinion_wids.add(wid)
fd.close()
#########
###
# Obtain the sentences that are opinionated (positive) and not (negative)
# The negatives are:
# If there are non-opinionated: just the non opinionated
# If not --> all the rest that are not positive
#####
sentences = {}
all_sent_ids = set()
sent_for_token_id = {}
kaf_obj = KafNafParser(kaf_file)
for token in kaf_obj.get_tokens():
token_id = token.get_id()
sent_id = token.get_sent()
token_value = token.get_text()
if sent_id not in sentences:
sentences[sent_id] = []
sentences[sent_id].append(token_value)
all_sent_ids.add(sent_id)
sent_for_token_id[token_id] = sent_id
###
positive_sents = set()
negative_sents = set()
##Positive sents are the sentences for the opinion_ids
for token_id in opinion_wids:
positive_sents.add(sent_for_token_id[token_id])
####
#Negative sents
if non_opinion is not None:
#In this case the negative are just the sentence of the no_opinion_wids
for token_id in no_opinion_wids:
negative_sents.add(sent_for_token_id[token_id])
else:
#In this case the negative are all the sentences but the positive ones
negative_sents = all_sent_ids - positive_sents
#Free some memory
del opinion_wids
del no_opinion_wids
del kaf_obj
##Store the results in the file
output_file = os.path.join(folder_out, basename + '.sents')
fd_out = open(output_file, 'w')
fd_out.write('#' + tag_file + '\n')
for sent_id in sorted(list(positive_sents)):
text = ' '.join(sentences[sent_id])
fd_out.write('+ ' + text.encode('utf-8') + '\n')
for sent_id in sorted(list(negative_sents)):
text = ' '.join(sentences[sent_id])
fd_out.write('- ' + text.encode('utf-8') + '\n')
fd_out.close()
#print 'Processed ',basename
#print ' Subjective sents:',len(positive_sents)
#print ' Non subje. sents:',len(negative_sents)
total_sents_opi += len(positive_sents)
total_sents_no_opi += len(negative_sents)
else:
print 'KAF FILE NOT FOUND', kaf_file
return total_sents_opi, total_sents_no_opi
treetagger_cmd = complete_path_to_treetagger+'/cmd/tree-tagger-spanish' mapping_file = this_folder +'/mappings/spanish.map.treetagger.kaf.csv' model = 'Spanish models' else: ## Default is dutch print>>sys.stderr,'Language',my_lang,'not supported by this wrapper' sys.exit(0) map_tt_to_kaf = loadMapping(mapping_file) ## Create the input text for reference_tokens = [] sentences = [] prev_sent='-200' aux = [] for token in input_obj.get_tokens(): sent_id = token.get_sent() word = token.get_text() w_id = token.get_id() if sent_id != prev_sent: if len(aux) != 0: sentences.append(aux) aux = [] aux.append((word,w_id)) prev_sent = sent_id if len(aux)!=0: sentences.append(aux) num_term = 0
def get_terms_in_sentence(naf: KafNafParser, sent: int) -> Iterable[Cterm]:
tokens = sort_tokens(t for t in naf.get_tokens() if t.get_sent() == sent)
tokenids = [t.get_id() for t in tokens]
return sort_terms(
naf, [naf.get_term(tid) for tid in naf.map_tokens_to_terms(tokenids)])
def process_file(this_file, token_freq):
xml_obj = KafNafParser(this_file)
print >> sys.stderr, 'Processing file', this_file
token_for_wid = {}
order_for_wid = {}
opinion_expressions = []
opinion_targets = []
whole_text = ' '
for n, token in enumerate(xml_obj.get_tokens()):
text = token.get_text().lower()
token_freq[text] += 1
token_for_wid[token.get_id()] = text
order_for_wid[token.get_id()] = n
whole_text += text + ' '
wids_for_tid = {}
lemma_for_wid = {}
pos_for_wid = {}
for term in xml_obj.get_terms():
tid = term.get_id()
wids = term.get_span().get_span_ids()
wids_for_tid[tid] = wids
for wid in wids:
lemma_for_wid[wid] = term.get_lemma()
pos_for_wid[wid] = term.get_pos()
##Properties!
aspects = [] ## [(label,term_span)...]
for property in xml_obj.get_properties():
for refs in property.get_references():
for span in refs:
aspects.append((property.get_type(), span.get_span_ids()))
already_counted = {EXP: set(), TAR: set()}
for opinion in xml_obj.get_opinions():
for this_type, opinion_obj in [(EXP, opinion.get_expression()),
(TAR, opinion.get_target())]:
if this_type is EXP and opinion_obj.get_polarity(
) == 'NON-OPINIONATED':
continue
if opinion_obj is not None:
span = opinion_obj.get_span()
if span is not None:
list_wids = []
for tid in span.get_span_ids():
list_wids.extend(wids_for_tid.get(tid, []))
list_wids.sort(key=lambda wid: order_for_wid[
wid]) ##Sorted according the the order of the tokens
string_wids = '#'.join(list_wids)
opinion_tokens = ' '.join(token_for_wid[wid]
for wid in list_wids)
opinion_lemmas = ' '.join(lemma_for_wid[wid]
for wid in list_wids)
opinion_pos = ' '.join(pos_for_wid[wid]
for wid in list_wids)
if string_wids not in already_counted[this_type]:
if this_type == EXP:
polarity = (opinion_obj.get_polarity()).lower()
opinion_expressions.append(
(opinion_tokens, polarity, opinion_lemmas,
opinion_pos))
else:
##Calculate the aspect type
possible_aspects = []
target_ids = span.get_span_ids()
for aspect_label, aspect_span in aspects:
num_in_common = len(
set(target_ids) & set(aspect_span))
if num_in_common != 0:
possible_aspects.append(
(aspect_label, num_in_common,
len(aspect_span)))
aspect_for_target = 'unknown'
if len(possible_aspects) != 0:
##Sorting by the number in common first, and by the lengtgh of the aspect secondly
aspect_for_target = sorted(possible_aspects,
key=lambda t:
(t[1], t[2]),
reverse=True)[0][0]
opinion_targets.append(
(opinion_tokens, aspect_for_target,
opinion_lemmas, opinion_pos))
already_counted[this_type].add(string_wids)
del xml_obj
print >> sys.stderr, '\tNumber of opinion expressions:', len(
opinion_expressions)
print >> sys.stderr, '\tNumber of opinion targets:', len(opinion_targets)
print >> sys.stderr, '\tNumber of characters of the text:', len(whole_text)
return opinion_expressions, opinion_targets, whole_text
def load_naf_stdin():
"""Load a dataset in NAF format.
Use this function to create a new ConlluDataset from a NAF file,
read from stdin.
NOTE: you can only add to NAF files, not create one from scratch.
"""
my_parser = KafNafParser(sys.stdin)
my_dataset = ConlluDataset()
# a big look-up table: for any NAF id, return a hash with
# {sent_id, token_id} in the ConlluDataset
naf2conll_id = {}
# collect the sentences in a hash, indexed by token_obj.get_sent()
sentences = {}
# iterate over the tokens to get: ID, FORM
for token_obj in my_parser.get_tokens():
# (string) identifier of the sentence
sent_id = token_obj.get_sent()
if sent_id in sentences:
sentence = sentences[sent_id]
else:
sentence = Sentence(sent_id=sent_id)
sentences[sent_id] = sentence
# (string) number of the token in the sentence, starting at '1'
token_id = '{}'.format(len(sentence) + 1) # ID
new_token = Token([
token_id, # ID
token_obj.get_text(), # FORM
'_', # LEMMA
'_', # UPOS
'_', # XPOS
'_', # FEATS
'0', # HEAD -> to be overwritten later
'root', # DEPREL -> to be overwritten later
'_', # DEPS
'_' # MISC
])
sentence.add(new_token)
# to match a NAF span to conll tokens, we need sent_id and token_id
naf2conll_id[token_obj.get_id()] = {
'sent_id': sent_id,
'token_id': token_id
}
# iterate over the term to get: LEMMA, XPOS, UPOS, FEATS, sent_id, nafid
for term_obj in my_parser.get_terms():
# span
# TODO: for now, assume terms map one-on-one on tokens
nafid = term_obj.get_span().get_span_ids()
if len(nafid) > 1:
logging.error('Multi-word tokens not implemented yet.')
return
nafid = nafid[0]
conllid = naf2conll_id[nafid]
sent_id = conllid['sent_id']
sentence = sentences[sent_id]
token_id = conllid['token_id']
token = sentence[token_id]
# store the identifier of the NAF term on the token, so we can add
# information to the NAF later.
token.nafid = term_obj.get_id()
token.LEMMA = term_obj.get_lemma()
# NAF pos='' is in lower case, UD UPOS is upper case
token.UPOS = term_obj.get_pos().upper()
# naf: A(B,C) -> ud: A|B|C
xpos = term_obj.get_morphofeat()
if xpos:
token.XPOS = xpos.replace('(', '|').replace(')',
'').replace(',', '|')
if token.XPOS[-1] == '|':
token.XPOS = token.XPOS[:-1]
# look for an external reference containing FEATS
for ext_ref in term_obj.get_external_references():
if ext_ref.get_reftype() == 'FEATS':
token.FEATS = ext_ref.get_reference()
# to match NAF dependencies to conll tokens, we need sent_id and token_id
naf2conll_id[term_obj.get_id()] = {
'sent_id': sent_id,
'token_id': token_id
}
# iterate over the dependencies to get: HEAD, DEPREL
for dep_obj in my_parser.get_dependencies():
# from
conllid = naf2conll_id[dep_obj.get_from()]
sent_id = conllid['sent_id']
sentence = sentences[sent_id]
token_id = conllid['token_id']
token_from = sentence[token_id]
# to
conllid = naf2conll_id[dep_obj.get_to()]
sent_id = conllid['sent_id']
sentence = sentences[sent_id]
token_id = conllid['token_id']
token_to = sentence[token_id]
# function
depfunc = dep_obj.get_function()
token_to.HEAD = token_from.ID
token_to.DEPREL = depfunc
# A final conversion of our list of sentences to a ConlluDataset
for sent_id in sentences:
sentence = sentences[sent_id]
# construct the sentence.full_text
raw_tokens = []
for token in sentence:
raw_tokens.append(token.FORM)
sentence.full_text = ' '.join(raw_tokens)
# add to the dataset
my_dataset.add(sentence)
my_dataset.naf2conll_id = naf2conll_id
return my_dataset, my_parser
def create_training_sentences(folder_tag_in,folder_kaf_in, opinion_layers,non_opinion,folder_out):
#Remove the outputfolder if exists and create it again
if os.path.exists(folder_out):
shutil.rmtree(folder_out)
os.mkdir(folder_out)
total_sents_opi = total_sents_no_opi = 0
for tag_file in glob.glob(os.path.join(folder_tag_in,'*.tag')):
basename = os.path.basename(tag_file).replace('.tag','')
kaf_file = os.path.join(folder_kaf_in,basename+'.kaf')
if os.path.exists(kaf_file):
##From the tag file we extract the token ids for opinions and for non opinionated
opinion_wids = set() #token ids annotated as opinions
no_opinion_wids = set() #token ids annotated as no opinions
fd = open(tag_file,'rb')
for line in fd:
fields = line.strip().split('\t')
wid = fields[0]
for opinion_idx in opinion_layers:
if fields[opinion_idx] == 'Opinion':
opinion_wids.add(wid)
if non_opinion is not None and fields[non_opinion] == 'NON-OPINIONATED':
no_opinion_wids.add(wid)
fd.close()
#########
###
# Obtain the sentences that are opinionated (positive) and not (negative)
# The negatives are:
# If there are non-opinionated: just the non opinionated
# If not --> all the rest that are not positive
#####
sentences = {}
all_sent_ids = set()
sent_for_token_id = {}
kaf_obj = KafNafParser(kaf_file)
for token in kaf_obj.get_tokens():
token_id = token.get_id()
sent_id = token.get_sent()
token_value = token.get_text()
if sent_id not in sentences:
sentences[sent_id] = []
sentences[sent_id].append(token_value)
all_sent_ids.add(sent_id)
sent_for_token_id[token_id] = sent_id
###
positive_sents = set()
negative_sents = set()
##Positive sents are the sentences for the opinion_ids
for token_id in opinion_wids:
positive_sents.add(sent_for_token_id[token_id])
####
#Negative sents
if non_opinion is not None:
#In this case the negative are just the sentence of the no_opinion_wids
for token_id in no_opinion_wids:
negative_sents.add(sent_for_token_id[token_id])
else:
#In this case the negative are all the sentences but the positive ones
negative_sents = all_sent_ids - positive_sents
#Free some memory
del opinion_wids
del no_opinion_wids
del kaf_obj
##Store the results in the file
output_file = os.path.join(folder_out,basename+'.sents')
fd_out = open(output_file,'w')
fd_out.write('#'+tag_file+'\n')
for sent_id in sorted(list(positive_sents)):
text = ' '.join(sentences[sent_id])
fd_out.write('+ '+text.encode('utf-8')+'\n')
for sent_id in sorted(list(negative_sents)):
text = ' '.join(sentences[sent_id])
fd_out.write('- '+text.encode('utf-8')+'\n')
fd_out.close()
#print 'Processed ',basename
#print ' Subjective sents:',len(positive_sents)
#print ' Non subje. sents:',len(negative_sents)
total_sents_opi += len(positive_sents)
total_sents_no_opi += len(negative_sents)
else:
print 'KAF FILE NOT FOUND',kaf_file
return total_sents_opi, total_sents_no_opi
def add_file(filename, data_lexelt, reftype='lexical_key'):
obj = KafNafParser(filename)
tokens_per_sent = {}
sent_for_token = {}
sents_in_order = []
for token in obj.get_tokens():
sentid = token.get_sent()
if sentid not in sents_in_order:
sents_in_order.append(sentid)
sent_for_token[token.get_id()] = sentid
if sentid not in tokens_per_sent: tokens_per_sent[sentid] = []
tokens_per_sent[sentid].append((token.get_id(), token.get_text()))
annotated_lemmas = [] # LIST of (full_id, token ids, lemma,pos,synset)
for term in obj.get_terms():
synset_label = None
for ext_ref in term.get_external_references():
if ext_ref.get_reftype() == 'lexical_key':
synset_label = term.get_lemma() + '%' + ext_ref.get_reference()
elif ext_ref.get_reftype() == 'sense' and ext_ref.get_resource(
) == 'WordNet-3.0':
synset_label = ext_ref.get_reference()
if synset_label is not None:
break
if synset_label is not None:
annotated_lemmas.append(
(filename + '#' + term.get_id(),
term.get_span().get_span_ids(), term.get_lemma(),
term.get_pos(), synset_label))
for full_id, token_ids, lemma, pos, synset_label in annotated_lemmas:
#CREATE NEW INSTANCE
this_key = lemma + '.' + pos.lower()[0]
if this_key not in data_lexelt:
data_lexelt[this_key] = Clexelt(this_key, pos)
if not data_lexelt[this_key].exists(full_id):
#Create the new instance
new_instance = Cinstance()
new_instance.id = full_id
new_instance.docsrc = filename
new_instance.key = synset_label
tokens = []
target_indexes = []
this_sent = sent_for_token[token_ids[0]]
index = sents_in_order.index(this_sent)
start_idx = max(index - 2, 0)
end_idx = min(index + 2, len(sents_in_order) - 1)
selected_sents = sents_in_order[start_idx:end_idx + 1]
num_token = 0
for current_sent in selected_sents:
for token_id, token_text in tokens_per_sent[str(current_sent)]:
tokens.append(token_text)
if token_id in token_ids:
target_indexes.append(num_token)
num_token += 1
new_instance.tokens = tokens[:]
new_instance.index_head = target_indexes[:]
data_lexelt[this_key].add_instance(new_instance)
def process_single_file(self, file):
try:
xml_obj = KafNafParser(file)
except:
print >> sys.stderr, 'Error parsing', file, ': skipped'
return
print >> sys.stderr, 'Processing file', os.path.basename(
file), 'Type:', xml_obj.get_type()
self.langs[xml_obj.get_language()] += 1
sentences = []
current_sent = []
this_sent = None
pos_for_wid = {} ## For each token id (wid) the pos of it
for term in xml_obj.get_terms():
w_ids = term.get_span().get_span_ids()
pos = term.get_pos()
for wid in term.get_span().get_span_ids():
pos_for_wid[wid] = pos
for token in xml_obj.get_tokens():
wid = token.get_id()
value = token.get_text()
if self.convert_to_lowercase:
value = value.lower()
if value in self.punctuation:
value = 'PUN'
if value == '*':
value = 'STAR'
sentence = token.get_sent()
if this_sent is not None and sentence != this_sent: ## There is a new sent
sentences.append(current_sent)
current_sent = []
current_sent.append((wid, value))
this_sent = sentence
## Add the last sentence as well
sentences.append(current_sent)
for sentence in sentences:
if self.include_sentence_delimiters:
sentence.insert(0, ('xxx', '<S>'))
sentence.append(('xxx', '</S>'))
for idx in range(0, len(sentence)):
for ngramlen in range(self.min_ngram_len,
self.max_ngram_len + 1):
file_desc = self.get_file_desc_for_ngram(ngramlen)
start = idx
end = start + ngramlen
if end <= len(sentence):
this_ngram = '\t'.join(
value for wid, value in sentence[start:end])
this_ngram_pos = '\t'.join(
pos_for_wid.get(wid, 'X')
for wid, value in sentence[start:end])
file_desc.write(
this_ngram.encode('utf-8') + '\t' + DELIMITER +
'\t' + this_ngram_pos + '\n')
