def wordnetLinking(self):
for trp in self.triples:
sub, rel, obj = trp['triple']
raw_sub, raw_rel, raw_obj = trp['raw_triple']
sub_id, rel_id, obj_id = self.ent2id[sub], self.rel2id[rel], self.ent2id[obj]
for sentence in trp['src_sentences']:
# sent = [wrd.lower() for wrd in sentence.split()]
sent = sentence.split()
''' 92 is the length of list returned by dir when lesk is successful '''
self.ent2wnet[sub_id] = self.ent2wnet.get(sub_id, set())
res = lesk(sent, raw_sub)
if len(dir(res)) == 92: self.ent2wnet[sub_id].add(res.name())
self.ent2wnet[obj_id] = self.ent2wnet.get(obj_id, set())
res = lesk(sent, raw_obj)
if len(dir(res)) == 92: self.ent2wnet[obj_id].add(res.name())
self.rel2wnet[rel_id] = self.rel2wnet.get(rel_id, set())
res = lesk(sent, raw_rel)
if len(dir(res)) == 92: self.rel2wnet[rel_id].add(res.name())
self.setHeading('Wordnet Entity Clusters')
self.printCluster(self.ent2wnet, self.id2ent, 'm2ol')
# for ent in self.ent_list: self.ent2wnet[self.ent2id[ent]] = [ele.name() for ele in lesk(ent)]
# for rel in self.rel_list: self.rel2wnet[self.rel2id[rel]] = [ele.name() for ele in wordnet.synsets(rel)]
self.setHeading('Wordnet Relation Clusters')
self.printCluster(self.rel2wnet, self.id2rel, 'm2ol')
def caseC(question_string, question_string_pos_tagged,
noun_phrases_in_question, likely_sentence_string,
likely_sentence_string_tokenized, noun_phrases_in_likely_sentence):
second_word_in_question_tag = question_string_pos_tagged[1][1]
if (second_word_in_question_tag in noun_tags):
second_word_in_question_supersense = lesk(
question_string, question_string_pos_tagged[1][0], 'n')
second_word_in_question_supersense = second_word_in_question_supersense.lexname(
)
super_sense_matches_in_likely_sentence = []
for word in likely_sentence_string_tokenized:
temp_supersense = lesk(likely_sentence_string, word, 'n')
if (temp_supersense is None):
continue
temp_supersense = temp_supersense.lexname()
if (temp_supersense == second_word_in_question_supersense):
super_sense_matches_in_likely_sentence.append(word)
if (len(super_sense_matches_in_likely_sentence) == 1):
return super_sense_matches_in_likely_sentence[0]
if (len(super_sense_matches_in_likely_sentence) > 1):
first_np_in_likely_sentence = noun_phrases_in_likely_sentence[0]
answer = closest_phrase_absolute(
likely_sentence_string, first_np_in_likely_sentence,
super_sense_matches_in_likely_sentence)
return answer
return None
else:
return None
def createAllNyms(self,sentence):
wordTokens = word_tokenize(sentence)
pos = pos_tag(wordTokens)
hyper = {}
hypo = {}
mero = {}
holo = {}
index = 0
for tokenOrig in wordTokens:
#print(tokenOrig)
if(pos[index][1] in self.wordnet_tag_map):
token = lesk(wordTokens, tokenOrig, self.wordnet_tag_map[pos[index][1]])
else:
token = lesk(wordTokens, tokenOrig)
index = index+1
#print(token)-
hyper[token] = []
hypo[token] = []
mero[token] = []
holo[token] = []
if (token):
if token.hypernyms():
hyper[token] = token.hypernyms()
if token.hyponyms():
hypo[token] = token.hyponyms()
mero[token] = token.part_meronyms()
holo[token] = token.part_holonyms()
return hyper, hypo, mero, holo
def simplified_lesk(word, sent3, sent1, sent2):
a = lesk(sent1, 'bank', 'n')
b = lesk(sent2, 'bank', 'n')
c = lesk(sent3, 'bank', 'n')
if a == b and b == c:
best_sence = b
elif a == b and a == c:
best_sence = a
elif b == c and a == c:
best_sence = c
max_overlap = 0
count_final = 0
for word in sent3:
count = 0
if word in sent1:
count = count + 1
if count > count_final:
max_overlap = 1
for word in sent3:
count = 0
if word in sent2:
count = count + 1
if count > count_final:
max_overlap = 2
return max_overlap
def my_lesk(tagged_strings, desired_word):
"""
@tagged_string: the string of words in the format 'word1_pos1 word2_pos2...'
@desired_word: the word we want disambiguated
Returns: - synset returned by lesk with part of speech (more accurate)
- synset returned by lesk without pos specified if no pos (less accurate)
- None if lesk returns nothing
"""
normal_string = ''
desired_tag = ''
for tagged_string in tagged_strings:
word, tag = tagged_string.rsplit("_", 1)
# Reject non-ASCII characters
try:
word = word.decode('ascii')
except (UnicodeDecodeError, UnicodeEncodeError):
continue
if word == desired_word:
desired_tag = tag
normal_string += word + ' '
# ignore proper nouns and punctuation
if desired_tag == 'NNP' or desired_tag == 'NNPS' or desired_tag in string.punctuation:
return None
# if the POS can be resolved to a wordnet POS, call lesk with POS
# else call lesk without POS
wn_pos = reduce_pos_tagset(desired_tag)
if not wn_pos:
return lesk(normal_string,desired_word)
else:
return lesk(normal_string, desired_word, wn_pos)
def createAllNyms(self, sentence):
wordTokens = word_tokenize(sentence)
pos = pos_tag(wordTokens)
hyper = {}
hypo = {}
mero = {}
holo = {}
index = 0
hyperd = {}
hypod = {}
merod = {}
holod = {}
for tokenOrig in wordTokens:
#print(tokenOrig)
if (pos[index][1] in self.wordnet_tag_map):
token = lesk(wordTokens, tokenOrig,
self.wordnet_tag_map[pos[index][1]])
else:
token = lesk(wordTokens, tokenOrig)
index = index + 1
#print(token)-
hyper[token] = []
hypo[token] = []
mero[token] = []
holo[token] = []
if (token):
print("")
print(tokenOrig)
print(token)
if token.hypernyms():
#hyper.append(token.hypernyms())
hyperd[token] = token.hypernyms()
print("hypernyms")
print(token.hypernyms())
print("")
else:
print("there are no hypernyms")
if token.hyponyms():
#hypo[token] = token.hyponyms()
hypod[token] = token.hyponyms()
print("hyponyms")
print(token.hyponyms())
print("")
else:
print("there are no hyponyms")
#mero[token] = token.part_meronyms()
merod[token] = token.part_meronyms()
print("meronyms")
print(token.part_meronyms())
print("")
#holo[token] = token.part_holonyms()
holod[token] = token.part_holonyms()
print("holonyms")
print(token.part_holonyms())
print("")
return hyperd, hypod, merod, holod
def lesk_similarity(triple_1, triple_2):
triple_1 = triple_1.lower().split(' ')
triple_2 = triple_2.lower().split(' ')
triple_1 = [
wnl.lemmatize(w.replace('^-1', '')) for w in triple_1 if w not in stop
]
triple_2 = [
wnl.lemmatize(w.replace('^-1', '')) for w in triple_2 if w not in stop
]
triple_1 = [x for x in triple_1 if x != '']
triple_2 = [x for x in triple_2 if x != '']
count = 0
for wrd_1 in triple_1:
if count == 1:
break
for wrd_2 in triple_2:
if wrd_1 == wrd_2 or wrd_1 in wrd_2 or wrd_2 in wrd_1:
count += 1
index_1 = triple_1.index(wrd_1)
index_2 = triple_2.index(wrd_2)
break
if count == 0:
return 0
syn_1 = lesk(triple_1, triple_1[index_1])
syn_2 = lesk(triple_2, triple_2[index_2])
if syn_1 == syn_2:
return 1
else:
return 0
def wsd(sent, subj, obj):
'''
Questa funzione effettua la WordSenseDisambiguation mediante il metodo lesk() fornito da WordNet restituendo il
synset delle parole disambiguate.
:param sent: frase del corpus;
:param subj: soggetto della frase;
:param obj: oggetto della frase;
:return: synset relativi al soggetto ed oggetto.
'''
possible_subj = ["i", "you", "he", "she", "it", "we", "they"]
sing_subj = ["i", "he", "she", "it"]
plural_subj = ["we", "you" "they"]
if subj in possible_subj or subj is None:
if subj in plural_subj:
ris = wn.synsets('people')[0]
else:
ris = wn.synsets('person')[0]
elif subj is not None:
ris = lesk(sent, subj)
if ris is None:
ris = wn.synsets('people')[0]
else:
ris = None
if obj is not None:
ris1 = lesk(sent, obj)
if ris1 is None:
#ris1 = wn.synsets('thing')[0]
ris1 = wn.synsets('food')[0]
else:
ris1 = wn.synsets('food')[0]
#ris1 = wn.synsets('thing')[0]
#print("Ris ", ris, "Ris1 ", ris1)
return ris, ris1
def compute_supersense(tuple, sentence):
if tuple[2] in personal_pronoun_tags:
return 'noun.person' if tuple[0] != 'it' else 'noun.entity'
if tuple[0] == 'who':
return 'noun.person'
if tuple[0] == 'what':
return 'noun.entity'
# Ex. in sentence: interest in how what people eat affects their health
# ('what', 'obj', 'WP') what sense returns no sense
# force to entity
if tag_to_wnpos_map[tuple[2]] is not None:
#ottieni la lista di wn pos abbinata al tag penn e calcola il sinset con lesk per ognuno di questi
possible_syns_by_pos = [
x for x in [
lesk(sentence, tuple[0], pos=pos)
for pos in tag_to_wnpos_map[tuple[2]]
] if x is not None
]
#usa lesk per disambiguare tra i synset trovati
filler_sense = lesk(sentence, tuple[0], synsets=possible_syns_by_pos)
else:
filler_sense = lesk(sentence, tuple[0])
filler_supersense = filler_sense.lexname(
) if filler_sense is not None else None
return filler_supersense
def __setup_lesk(cls):
try:
lesk(word_tokenize('This device is used to jam the signal'), 'jam')
except LookupError:
nltk.download('punkt')
nltk.download('wordnet')
nltk.download('omw-1.4')
lesk(word_tokenize('This device is used to jam the signal'), 'jam')
def assign_word_senses(self, word, partOfSpeech, sentenceIndex):
if partOfSpeech == VERBS:
wordSense = wn.synsets(word.word, pos=wn.VERB)
if wordSense:
word.wordSense = lesk(self.sentences[sentenceIndex], word.word, 'v')
elif partOfSpeech == NOUNS:
wordSense = wn.synsets(word.word, pos=wn.NOUN)
if wordSense:
word.wordSense = lesk(self.sentences[sentenceIndex], word.word, 'n')
def disambiguate():
"""Returns the best synset if a word is ambiguous"""
lesks = []
nouns = o_tag()
text = one_line()
for noun in nouns:
if len(wn.synsets(noun[0])) > 1:
if lesk(text, noun[0], 'n'):
lesks.append((noun[0], lesk(text, noun[0], 'n')))
elif lesk(text, noun[0], 'n'):
lesks.append((noun[0], wn.synsets(noun[0])))
return lesks
def apply_lesk(offset_sentic_dict):
print('Appling Lesk Algorithms Started...')
# making another offset_sentic_dict for rest of concepts and concat it to the existed one.
# input: offset_sentic_dict["00044455-n"] = ['0.1', '0.1', '0.1', '0.1', #joy', '#surprise', 'positive', '0.726', 'appearance', 'start', 'casus_belli', 'beginning', 'egress'] // semantics might not included
# output: last_offset_sentic_dict["00044455-n"] = ['0.1', '0.1', '0.1', '0.1', '#joy', '#surprise', 'positive', '0.726', 'appearance', 'start', 'casus_belli', 'beginning', 'egress']
with open('vocabulary/affectnet_dict.pkl', 'rb') as f:
aff_fr_dict = pickle.load(f)
fr_offset_dict = dict()
deleted_offset = []
#direct mapped words were deleted before
for word, value in senticnet.items():
context = word
found = False
for i in range(8, 13):
context += senticnet[word][i] + ' '
try:
synset = lesk(context, word)
offset = str(synset.offset()).zfill(8) + '-' + synset.pos()
found = True
except AttributeError:
# not found
# sequentially, because it is arranged by c. similarity
for v in value[8:13]:
try:
synset = lesk(context, v)
offset = str(synset.offset()).zfill(8) + '-' + synset.pos()
found = True
except AttributeError:
continue
if found == False:
continue
# Direct mapped offset is not considered
if offset in offset_sentic_dict:
continue
if offset in deleted_offset:
continue
if offset not in offset_sentic_dict:
offset_sentic_dict[offset] = value
fr_offset_dict[offset] = aff_fr_dict[word]
else:
if offset_sentic_dict[offset][6] != value[6]:
del offset_sentic_dict[offset]
deleted_offset.append(offset)
continue
else:
offset_sentic_dict[offset] = weighted_sentic(offset_sentic_dict[offset], fr_offset_dict[offset], value, aff_fr_dict[word])
fr_offset_dict[offset] += aff_fr_dict[word]
return offset_sentic_dict
def wsd(listOfWords, word, pos = None): if pos == None: syn = lesk(listOfWords, word) return [syn] if isSynset(syn) else [] else: syns = [] for x in pos: syn = lesk(listOfWords, word, x) if isSynset(syn): syns.append(syn) return syns
def disambiguate():
"""Returns the best synset if a word is ambiguous"""
lesks = []
nouns = o_tag()
text = one_line()
for noun in nouns:
if len(wn.synsets(noun[0])) > 1:
if lesk(text, noun[0], 'n'):
lesks.append((noun[0], lesk(text, noun[0], 'n')))
elif lesk(text, noun[0], 'n'):
lesks.append((noun[0], wn.synsets(noun[0])))
return lesks
def batch_matrix(self, dataset, begin_idx, end_idx):
indice = range(begin_idx, end_idx)
r = np.zeros((end_idx - begin_idx, self.sequence_length, self.sequence_length, 5))
premise = np.array([dataset[i]['sentence1_binary_parse_index_sequence'] for i in indice])
hypothesis = np.array([dataset[i]['sentence2_binary_parse_index_sequence'] for i in indice])
premise_def, hypothesis_def = np.zeros((2, end_idx - begin_idx, self.sequence_length, self.sequence_length))
mask_p, mask_h = np.ones((2,end_idx - begin_idx, self.sequence_length))
mask_p_def, mask_h_def = np.ones((2,end_idx - begin_idx, self.sequence_length, self.sequence_length))
labels = np.array([dataset[i]['label'] for i in indice])
genres = np.array([dataset[i]['genre'] for i in indice])
for i in range(begin_idx, end_idx):
pre, hyp = dataset[i]['sentence1'].split()[:self.sequence_length], dataset[i]['sentence2'].split()[:self.sequence_length]
mask_p[i - begin_idx, :len(pre)] = 1
mask_h[i - begin_idx, :len(hyp)] = 1
for i_p, p in enumerate(pre):
# lesk: find the most possible sysnet of word p from word p and whole sentence pre
tmp = lesk(pre, p)
if tmp is not None:
# get p's definition
tmp = tmp.definition().strip('\'()').split()
else:
continue
# turn definition words to corresponding indices
premise_def[i - begin_idx][i_p][:len(tmp)] = ([word_indices[i] if i in word_indices else 0 for i in tmp])[:self.sequence_length]
mask_p_def[i - begin_idx][i_p][:len(tmp)] = 1
for i_h, h in enumerate(hyp):
tmp = lesk(hyp, h)
if tmp is not None:
tmp = tmp.definition().strip('\'()').split()
else:
continue
hypothesis_def[i - begin_idx][i_h][:len(tmp)] = ([word_indices[i] if i in word_indices else 0 for i in tmp])[:self.sequence_length]
mask_h_def[i - begin_idx][i_h][:len(tmp)] = 1
for i_p, p in enumerate(pre):
p_syn = lesk(pre, p)
if p_syn is None:
continue
for i_h, h in enumerate(hyp):
h_syn = lesk(hyp, h)
if h_syn is None:
continue
r[i - begin_idx][i_p][i_h][0] = is_synonym(p_syn, h_syn)
r[i - begin_idx][i_p][i_h][1] = is_antonym(p_syn, h_syn)
r[i - begin_idx][i_p][i_h][2] = is_hypernym([p_syn], h_syn, 1)
r[i - begin_idx][i_h][i_p][3] = is_hypernym([p_syn], h_syn, 1)
r[i - begin_idx][i_p][i_h][4] = is_co_hypernym(p_syn, h_syn)
return r, premise, hypothesis, premise_def, hypothesis_def ,labels, mask_p, mask_h, mask_p_def, mask_h_def, genres
def sentiment_analysis_wsd(text_n_tagged_text):
pos_tagged_text = text_n_tagged_text[0]
text = text_n_tagged_text[1]
pos_arr = []
neg_arr = []
subj_arr = []
for obj in pos_tagged_text:
if return_pos_sentiwordnet(obj[1]) == 0:
continue
pos = return_pos_sentiwordnet(obj[1])
if lesk(text, obj[0], pos):
syn = lesk(text, obj[0], pos)
polarity = polarity_score_1(syn)
subj = subjectivity_score_1(syn)
else:
polarity = polarity_score_2(obj[0], pos)
subj = subjectivity_score_2(obj[0], pos)
subj_arr.append(subj)
if polarity > 0.0:
pos_arr.append(polarity)
elif polarity < 0.0:
neg_arr.append(polarity)
else:
continue
#print pos_arr
if np.array(pos_arr).size == 0:
pos_mean_score = 0.0
else:
pos_mean_score = round(np.mean(np.array(pos_arr)), 1)
if np.array(neg_arr).size == 0:
neg_mean_score = 0.0
else:
neg_mean_score = round(np.mean(np.array(neg_arr)), 1)
subj_mean_score = round(np.mean(np.array(subj_arr)), 1)
temp_neg_score = neg_mean_score * -1.0
#if (pos_mean_score,neg_mean_score,subj_mean_score):
#return (pos_mean_score,neg_mean_score,subj_mean_score)
#else:
#return (0,0.0)
if pos_mean_score > temp_neg_score:
return ('1', pos_mean_score + neg_mean_score, subj_mean_score)
elif pos_mean_score < temp_neg_score:
return ('-1', pos_mean_score + neg_mean_score, subj_mean_score)
else:
return ('0', 0.0, subj_mean_score)
def get_semantic_features(tagged_tok, line):
'''
return features like synonyms, hypernyms, hyponyms, meronyms, holonymns
extracted from each word of sentence
'''
lemma_sen = set()
hyper_sen = set()
hypo_sen = set()
mero_sen = set()
holo_sen = set()
for word, tag in tagged_tok:
if tag[:2] in WN_TAG_LIST:
sense = lesk(line, word, pos=WN_TAG_LIST.get(tag[:2]))
if not sense:
continue
for lem in sense.lemmas():
lemma_sen.add(lem.name())
for hyper in sense.hypernyms()[:30]:
hyper_sen.add(hyper.name())
for hypo in sense.hyponyms()[:30]:
hypo_sen.add(hypo.name())
for mero in sense.part_meronyms()[:30]:
mero_sen.add(mero.name())
for holo in sense.member_holonyms()[:30]:
holo_sen.add(holo.name())
return (' '.join(lemma_sen), ' '.join(hyper_sen), ' '.join(hypo_sen),
' '.join(mero_sen), ' '.join(holo_sen))
def getFeatures(tokensTagged, line):
lemmaS = set()
hyperS = set()
hypoS = set()
meroS = set()
holoS = set()
for word, tag in tokensTagged:
if tag[:2] in nltkWnMap: # and tag != 'NNP':
sense = lesk(line, word, pos=nltkWnMap.get(tag[:2]))
# sense = lesk(line, word)
if not sense:
continue
for lem in sense.lemmas():
lemmaS.add(lem.name())
for hyper in sense.hypernyms()[:featureMaxLimit]:
hyperS.add(hyper.name())
for hypo in sense.hyponyms()[:featureMaxLimit]:
hypoS.add(hypo.name())
for mero in sense.part_meronyms()[:featureMaxLimit]:
meroS.add(mero.name())
for holo in sense.member_holonyms()[:featureMaxLimit]:
holoS.add(holo.name())
return (' '.join(lemmaS), ' '.join(hyperS), ' '.join(hypoS),
' '.join(meroS), ' '.join(holoS))
def word_synonyms(word, sentence):
synset = lesk(sentence, word)
if synset is None:
return None
lemmas = synset.lemmas()
lemmas = [str(lemma.name()) for lemma in lemmas]
return lemmas
def __subtree__(self, row):
"""
Find subtree pattern like X ->R<- Y and generation heuristic based score
where
X: Subject
R:Relation
Y:Object
"""
corpus = row["corpus"]
output = []
tokens = row["tokens"]
# iterate through all the tokens in the input sentence
for i, sent in enumerate(corpus):
r = ''
for tok in sent:
# # extract subject
# if tok.dep_.find("subjpass") == True:
# y = tok.text
# # extract object
# if tok.dep_.endswith("obj") == True:
# x = tok.text
# # extract relation
if tok.dep_ == "ROOT":
r = lesk(tokens[i], tok.text, get_wordnet_pos(tok.tag_))
if r is not None:
r = r.name()
break
output.append(r)
return output
def get_replacement(self, token_with_tag, tokens):
word = token_with_tag[0]
synset_lock = Lock()
synset_lock.acquire()
synset = lesk(tokens, word)
synset_lock.release()
if synset is None:
return None
lemmas = synset.lemmas()
# find a replacement
replacement = None
for lemma in lemmas:
lemma_name = lemma.name()
lemma_pos = lemma.synset().pos()
if (lemma_pos == 'n') and lemma_name != word:
replacement = lemma_name
break
if replacement is None:
#replacement = word
return None
return replacement.replace("_", " ")
def canonicalise(extractions):
for ext in extractions:
ext = extractions[ext]
# relation synsets
rel_synsets = set([])
rel_root = set([])
if ext['object']:
sentence = ext['subject'] + ' ' + ext['relation'] + ext['object']
else:
sentence = ext['subject'] + ' ' + ext['relation']
doc = nlp(ext['relation'])
for token in doc:
if token.pos_ == 'VERB' and token.text not in [
'will', 'shall', 'may', 'must', 'can', 'could'
]:
try:
rel_synsets.add(lesk(sentence, token.text, 'v').name())
rel_root.add(token.lemma_)
except:
print(
'ERROR:',
token.lemma_,
)
ext['rel_synsets'] = list(rel_synsets)
# entity canonicalisation
ext['subject'] = entity_canonicalisation(ext['subject'])
ext['object'] = entity_canonicalisation(ext['object'])
for m in ext['modifiers'] + ext['subject_modifiers']:
m['m_obj'] = entity_canonicalisation(m['m_obj'])
return extractions
def findCategories(tokens, tags, nouns, verbs):
for word in tokens:
if not lesk(tokens, word):
continue
if lesk(tokens, word).pos() == 'n':
category = lesk(tokens, word).lexname()
if category not in nouns.keys():
nouns[category] = 1
else:
nouns[category] += 1
elif lesk(tokens, word).pos() == 'v':
category = lesk(tokens, word).lexname()
if category not in verbs.keys():
verbs[category] = 1
else:
verbs[category] += 1
def get_tokens_POS(sentence_token_complexity_pairs):
stemmer = SnowballStemmer(LANGUAGE)
tokens_POS = [(sentence, lesk(sentence, str(token)), complexity) for sentence, token, complexity in sentence_token_complexity_pairs]
for i in range(len(tokens_POS)):
sentence, token, complexity = tokens_POS[i]
if token is None:
word = sentence_token_complexity_pairs[i][1]
token = lesk(sentence, stemmer.stem(word))
if token is None:
tokens_POS[i] = (sentence, word, complexity)
else:
tokens_POS[i] = (sentence, token, complexity)
return [(sentence, token, complexity) for sentence, token, complexity in tokens_POS]
def transform_tag(tag, word, words):
synset = lesk(words, word, "n")
if synset:
if tag == "ORGANIZATION" or tag == "PERSON":
return tag[:3]
elif tag == "LOCATION":
paths = synset.hypernym_paths()
for path in paths:
for synset in path:
name = synset.name()
if "city" in name or "town" in name:
return "CIT"
elif "country" in name or "state" in name:
return "COU"
return "NAT"
elif tag == "MISC":
paths = synset.hypernym_paths()
for path in paths:
for synset in path:
name = synset.name()
if "animal" in name:
return "ANI"
elif "sport" in name:
return "SPO"
elif "entertainment" in name:
return "ENT"
return ""
else:
return ""
else:
return ""
def query_expanded(list_of_words, weight, word_similarity = None): """ Expand the query using various word relations (synonyms, hypernyms or hyponyms) """ count = 0 expanded_query = [] for x in list_of_words: expanded_query.extend([x for i in range(weight)]) # # WSD syn = lesk(list_of_words, x) try: for l in syn.lemmas() : # if(count<3): if l.name() not in expanded_query: expanded_query.append(l.name()) count+=1 # for hyp in syn.hypernyms(): # for hyp_lemma in hyp.lemmas(): # if hyp_lemma.name() not in expanded_query: # expanded_query.append(hyp_lemma.name()) # count+=1 # for hyp in syn.hyponyms(): # for hyp_lemma in hyp.lemmas(): # if hyp_lemma.name() not in expanded_query: # expanded_query.append(hyp_lemma.name()) # count+=1 except: pass return expanded_query
def ExtendText(fileName, tagger=PerceptronTagger()):
with io.open(fileName, 'r') as w:
text = TextBlob(w.read(), pos_tagger=tagger)
extended_text = []
for sent in text.sentences:
for word in sent.pos_tags:
#word = "bank"
penn_tags = ['JJ', 'NN', 'V']
extending = False
for tag in penn_tags:
if tag in word[1]:
extending = True
pos = tag[0].lower()
try:
l = lesk(sent.string, word[0].lower(), pos)
syns = l._lemma_names
for syn in syns:
extended_text.append(syn)
break
except:
extended_text.append(word[0].lower())
if not extending:
extended_text.append(word[0].lower())
extended_text = ' '.join([
word for word in extended_text if word not in cachedStopWords
]).lstrip()
return extended_text
def main(questionLoc, categoriesLoc, featuresLoc):
categories, subCategories = [],[]
categories, subCategories = getCategories(categoriesLoc)
questions = loadQuestions(questionLoc)
features = getFeatureLists(featuresLoc)
finalString = ''
for i in range(len(questions)):
disambedSens = None
whWord = features[i][0]
headWord = features[i][1]
if headWord == 'null':
headWord = None
label = features[i][2]
label = convertPOS(label)
question = questions[i].strip()
uni = GetUnigram(question)
bi = GetBigram(question)
tri = GetTrigram(question)
wordShape = GetWordShape(uni)
if headWord != None and ':' not in headWord:
disambigSense = wsd.lesk(question, headWord, pos=label)
if disambigSense:
directHypernym = getHypernym(question, 5, disambigSense)
indirectHypernym = mostSimilarCategory(disambigSense, subCategories)
print whWord, headWord#, disambigSense, wordShape, uni, bi, tri
def desambiguar(men):
# Para que no den problemas los carateres los reemplazo
global cont
# Separo el texto en frases
frases = sentence_tokenizer.tokenize(men)
for frase in frases:
nfrase = frase #Nueva frase
# Proceso la frase
doc = nlp(frase)
cont = 0
# Busco las palabras ambiguas de la frase
for token in doc:
cont = nfrase.find(token.text,cont)
# Intento desambiguar
synset = lesk(doc, token.text)
if synset and len(wn.synsets(token.text)) > 1:
nfrase = creaPildora(token.text, synset.name(), nfrase)
else:
cont += len(token.text)
men = men.replace(frase,nfrase,1)
return men
def synonym_paraphrase(words, xpos):
verb_labels = ['VB', 'VBD', 'VBG', 'VBN', 'VBP', 'VBZ']
adj_labels = ['JJ', 'JJR', 'JJS']
noun_labels = ['NN', 'NNP', 'NNS', 'NNPS']
adv_labels = ['RB', 'RBR', 'RBS']
synonyms = {}
for i in range(len(words)):
### get pos tag
pos_tag = None
if (xpos[i] in verb_labels):
pos_tag = 'v'
if (xpos[i] in adj_labels):
pos_tag = 'a'
if (xpos[i] in noun_labels):
pos_tag = 'n'
if (xpos[i] in adv_labels):
pos_tag = 'r'
if (pos_tag == None):
continue
### get morphy of word
morphy = wn.morphy(words[i], pos_tag)
if (morphy == None):
continue
### get all synonyms
meaning = lesk(' '.join(words), morphy, pos_tag)
if (str(meaning) == 'None'):
continue
syns = meaning.lemma_names()
synonyms[morphy] = syns
return synonyms
def obtener_mejor_definicion(tokens, sustantivo):
definiciones = encontrar_definicion(sustantivo)
if len(definiciones) == 1:
return definiciones[0]['definicion']
else:
synset = lesk(tokens, sustantivo, 'n')
return synset.definition()
def mostcommonsyns(row): #print "Evaluating new row" record, stop_words = row text = record['text'] #print "------------------\n\n", text #print "hi len: ", len(text) stopWords = get_stopwords() sentList = nltk.sent_tokenize(text) #print "hi2" wordsInSentsPos = [nltk.pos_tag(nltk.word_tokenize(s)) for s in sentList] #print "hi3" wordsInSentsWnPos = [[(w[0],penn2morphy(w[1])) for w in s if w[0].lower() not in stopWords] for s in wordsInSentsPos] #print "h4" #the above returns a list of sentences where each sentence is a list of #(word-as-string, pos tag) tuples. Stop words are removed here because pos_tag #uses grammatical structure but lesk does not. #This would also be the place to lemmatize, which will help lesk out. synsetsList = [lesk(s,w[0],w[1]) for s in wordsInSentsWnPos for w in s] #print "go" #print synsetsList, "\n\n" #return [3,2,2,2] #res = FreqDist([x for x in synsetsList if x is not None]) res = [s.name() for s in synsetsList if s is not None] return res
def main():
wikiDict = pickle.load(open('wikis.pickle','rb'))
amountList = []
for value in wikiDict.values():
value = re.sub(r'\[[0-9]*\]',"",value)
text = sent_tokenize(value)
for sent in text:
tokenized = word_tokenize(sent)
pos = pos_tag(tokenized)
for token, tag in pos:
if (tag == 'NNPS' or tag == 'NNP' or tag == 'NNS' or tag == 'NN'):
tag = "n"
if len(wn.synsets(token, tag)) > 1:
print(token, lesk(sent, token, tag))
print("All possible senses:")
n = 0
for ss in wn.synsets(token, tag):
print(ss, ss.definition())
n += 1
amountList.append(int(n))
#print()
#print(amountList)
c2 = Counter(amountList)
print(c2)
def ExtendText(fileName,tagger=PerceptronTagger()): with io.open(fileName, 'r') as w: text = TextBlob(w.read(), pos_tagger=tagger) extended_text = [] for sent in text.sentences: for word in sent.pos_tags: #word = "bank" penn_tags = ['JJ','NN','V'] extending = False for tag in penn_tags: if tag in word[1]: extending = True pos = tag[0].lower() try: l = lesk(sent.string, word[0].lower(), pos) syns = l._lemma_names for syn in syns: extended_text.append(syn) break except: extended_text.append(word[0].lower()) if not extending: extended_text.append(word[0].lower()) extended_text = ' '.join([word for word in extended_text if word not in cachedStopWords]).lstrip() return extended_text
def abstraction_score(text, uselesk = False):
"""Takes in a list of sentences, tags the parts of speech from the word tokenized sentence,
looks for part of speech beginning with N (a type of noun) returns the mean abstraction score by
calling calculate_scores.
Optional, uses the default lesk algorithm from NLTK
"""
nouns = []
for sent in text:
nouns.append([token for token, pos in pos_tag(word_tokenize(sent)) if pos.startswith('N')])
# print nouns
nouns = [item for sublist in nouns for item in sublist]
scores=[]
for i in nouns:
if uselesk:
y = lesk(sent, i)
if y is not None:
#print i, calculate_scores(y)
scores.append(calculate_scores(y))
else:
by_word=[]
synsets = wn.synsets(i, "n")
for synset in synsets:
by_word.append(calculate_scores(synset))
absword = np.mean(by_word)
scores.append(absword)
#print absword
scores = np.array(scores)
#print scores
return scores[~np.isnan(scores)].mean()
def analyser(doc):
expand_tokens = []
expand_Lemms = []
synsets = []
# apply the preprocessing and tokenzation steps
sent_text = nltk.sent_tokenize(doc)
doc_clean = self.build_preprocessor()(doc)
tokens = self.build_tokenizer()(doc_clean)
# use CountVectorizer's _word_ngrams built in method
# to remove stop words and extract n-grams
n_grams = list(set(self._word_ngrams(tokens,self.get_stop_words())))
for x in n_grams:
my_regex = r"\b(?=\w)" + re.escape(x) + r"\b(?!\w)"
matched_sent = [s for s in sent_text if len(re.findall(my_regex, s, re.IGNORECASE)) > 0]
expand_tokens.append(x)
expand_Lemms.append(matched_sent)
Expanded_Token_Lemmd = pd.DataFrame({'Word': expand_tokens,'Sentences': expand_Lemms})
for i in range(0, len(Expanded_Token_Lemmd.index)):
for row_len in range(0, len(Expanded_Token_Lemmd.iloc[i]['Sentences'])):
x = lesk(Expanded_Token_Lemmd.iloc[i]['Sentences'][row_len],Expanded_Token_Lemmd.iloc[i]['Word'].lower().replace(' ','_'))
if not x:
synsets.append(Expanded_Token_Lemmd.iloc[i]['Word'].lower())
else:
synsets.append(str(x)[8:-2])
return(synsets)
def lexsense(word, context='') :
if context:
sense = lesk(context,word,pos=wn.NOUN);
if sense:
return {sense}
return set(wn.synsets(word, pos=wn.NOUN))
def calculate_probable_synset(self, sentence: List[str]):
if self.probable_synset:
return
self.probable_synset = lesk(sentence,
self.token.text,
pos=_token_pos_to_nltk_pos(self.token))
def main(file):
# Stuff for unicode decode errors
reload(sys)
sys.setdefaultencoding("utf-8")
# Get the text of the URL
text = getText(file[1])
tokens = nltk.word_tokenize(text)
# POS tag words
taggedWords = nltk.pos_tag(tokens)
# Filter all nouns
nouns = [(word, tag) for word, tag in taggedWords if tag.startswith("N")]
# Set values for the answers
amountOfPolysemousWords = 0
amountOfSenses = 0
listofsenses = []
# Get wordnet synsets
for word, tag in nouns:
# Get the amount of senses
senses = len(wordnet.synsets(word, "n"))
# Check if the word is polysemous
if senses > 1:
# Count the polysemous words and senses
amountOfPolysemousWords = amountOfPolysemousWords + 1
amountOfSenses = amountOfSenses + senses
listofsenses.append(senses)
# Answer for question 1
print("For this file, there are {} polysemous word".format(amountOfPolysemousWords))
# Answer for question 3
averageSenses = amountOfSenses/amountOfPolysemousWords
print("For this file, the average senses are {} per polysemous word".format(averageSenses))
# Answer for question 4
result = Counter(listofsenses)
print(result)
# Answer for question 5
words = ["cars", "quantity", "carbon", "states", "change", "life"]
pos = "n"
textObject = nltk.Text(tokens)
for sent in sent_tokenize(text):
for word in words:
context = textObject.concordance(word)
print("\n\n" + str(context))
print ("\n\n The result of algorithm is: " + str(lesk(sent, word, pos)))
print("\n\n All possible senses for " + word + ":")
for ss in wordnet.synsets(word, "n"):
print(ss, ss.definition())
def disambiguate(context, word, pos):
"""
Word sense disambiguation using Lesk algorithm
@context: a string containing the word whose meaning we want to disambiguate
@word: the word we want to disambiguate
@pos: the part of speech
Returns: the Synset of the most likely meaning of the word
"""
return lesk(context, word, pos)
def leskify(file):
with(open(file, 'r')) as f:
contents = f.read()
words, describingWord = wordify(file)
meaning = {}
for word in words:
meaning[word] = lesk(contents, word)
# meaning.append((word, lesk(contents, word)))
return meaning, describingWord
def define_func(nick,match,target):
word=match.group('word')
sentence=match.group('sentence')
meaning=match.group('meaning')
lang=match.group('lang')
sentiment=match.group('sentiment')!=None
if sentence:
separator=re.compile("[.\s,]+")
words=re.split(separator,sentence)
for w in words:
synsets=wn.synsets(w)
if synsets!=[]:
word=w
break
if synsets!=[]:
synset=lesk(words,word)
define_synset(nick,word,synset,target,lang,sentiment)
else:
mb.tell(nick+": what the flying f**k does that even mean",target)
return
if word:
synsets=wn.synsets(word)
if synsets==[]:
mb.tell(nick+': no idea what "'+word+'" is. probably something gay. like you',target)
elif len(synsets)==1:
define_synset(nick,word,synsets[0],target,lang,sentiment)
else:
choose=[]
temp_choose=[]
for synset in synsets:
name=None
for hypernym in synset.hypernyms():
for lemma in hypernym.lemmas():
name=lemma.name()
break
break
if not name:
for lemma in synset.lemmas():
name=lemma.name()
break
temp_name=name
if name in temp_choose:
name=name+"("+str(temp_choose.count(name))+")"
if meaning:
if meaning.upper()==name.upper():
define_synset(nick,word,synset,target,lang,sentiment)
return
temp_choose.append(temp_name)
choose.append(name)
words="|".join([re.escape(x).replace("_","[_\s]+") for x in choose])
message=", ".join(choose)
pattern="^(?:murderb[o0]t[,\s:!]+)?(?:(?:(?:(?:who|what)(?:\s+am|\s+is|\s+are|'s|'re|'m|s|re)(?:\s+a|\s+the)?)|(?:define))\s+)?(?:"+word+"\s+)?(?:as\s+in\s+)?(?P<word>{0})$".format(words);
clarify=re.compile(pattern,flags=re.IGNORECASE)
response={'nick':nick,'func':clarify_func,'pattern':clarify,'param':{'sent':sentiment,'lang':lang,'synsets':synsets,'words':[x.upper().replace("_","") for x in choose],'word':word},'target':target}
mb.responses['define']=response
mb.tell(nick+": "+word+" as in "+message+"?",target)
return
def extract_nouns_info(sentence):
nouns_info = []
tokens = word_tokenize(sentence)
tagged_words = pos_tag(tokens)
for (word, tag) in tagged_words:
if('NN' in tag):
pos = 'n'
wsd = lesk(tokens, word, pos)
syns = wsd.lemma_names()
nouns_info.append( (word, wsd, syns) )
return nouns_info
def lexclass(word, context=''):
rslt = set()
if context:
sense = lesk(context,word,pos=wn.NOUN);
if sense:
rslt.add(sense.lexname())
return rslt
sets = wn.synsets(word, pos=wn.NOUN)
for s in sets:
rslt.add(s.lexname())
return rslt
def disambiguate(self, wsd_instance): """ Disambiguates the given instance, returning the predicted lemma sense key. @param wsd_instance - WSDInstance to disambiguate @return Sense key as a string if a sense could be found, None otherwise """ syn = lesk(wsd_instance.context, wsd_instance.lemma, 'n') if syn is not None: return to_sense_key(syn) return None
def getWordSentimentTuple(self, word, pos, wordlist):
if wordlist != "sentiwordnet":
raise InvalidDictionaryException("Invalid dictionary " + wordlist + \
"please use sentiwordnet")
else:
simplePOS = convertPOSTagToSimplePOS(pos)
if pos:
wordSense = lesk(self.tokens, word, simplePOS)
if wordSense:
sentiSynsetWord = swn.senti_synset(wordSense.name())
if sentiSynsetWord:
return (sentiSynsetWord.pos_score(),
sentiSynsetWord.neg_score(),
sentiSynsetWord.obj_score())
return (0, 0, 0)
def synonym_picker3(word, wpos, sentence, npos):
#print(word, pos, sentence)
#print(word, " ", npos)
start_synset = lesk(sentence, word, wpos)
synonyms = []
#start_synset.res_similarity()
if (start_synset):
#print(start_synset.lemmas())
for l in start_synset.lemmas():
if l.name() not in synonyms and l.name() != word:
synonyms.append(l.name().replace('_', ' '))
if len(synonyms) > 0:
return random.choice(list(synonyms))
else:
return word
else:
return word
def analyze(self, doc):
res = []
for sentence in self.normalizer.sent_tokenize(doc):
tagged_sentence = self.tagger.tag(self.normalizer.split_and_normalize(sentence))
lemmatized_doc = []
for w, pos in tagged_sentence:
try:
pos_ = pos[:1]
wn_postag = self.translation_dict[pos_]
except KeyError:
wn_postag = None
if wn_postag:
lemmatized_doc.append(self.lem.lemmatize(w, wn_postag))
for w in lemmatized_doc:
sense = wsd.lesk(lemmatized_doc, w)
if sense:
res.append(sense.name())
return res
def get_Candidate_Frequency_from_wordnet(word, tag, context):
wordnet_tag = get_wordnet_pos(tag)
sent = list(context)
#print("context is")
#print(sent)
#syns = lesk(sent, word, wordnet_tag)
syns = lesk(sent, word)
res = []
if syns:
for l in syns.lemmas():
if l:
lemma_name = str(l.name())
st = LancasterStemmer()
if ((st.stem(word) != st.stem(lemma_name))):
res.append(lemma_name)
candidate_list = list(set(res))
#print("for word: " + word)
#print(candidate_list)
'''
if candidate_list:
for c in candidate_list:
allsyns1 = set(ss for ss in wordnet.synsets(c))
print(allsyns1)
allsyns2 = set(ss for ss in wordnet.synsets(word))
print(allsyns2)
best = max((wordnet.wup_similarity(s1, s2) or 0, s1, s2) for s1, s2 in product(allsyns1, allsyns2))
print(best)
'''
'''
for c in candidate_list:
wordsFromList1 = wordnet.synsets(word)
wordsFromList2 = wordnet.synsets(c)
if wordsFromList1 and wordsFromList2: # Thanks to @alexis' note
s = wordsFromList1[0].wup_similarity(wordsFromList2[0])
similarity_list.append(s)
'''
return candidate_list
def main():
path = "group9/"
dirs = ["p34", "p35"]
number_of_ss = 0
synsets = []
for directory in dirs:
for directory2 in os.listdir(path+directory):
for filename in os.listdir(path+directory+"/"+directory2):
if filename.endswith(".tok.off.pos.ent"):
with open(os.path.join(path, directory+"/"+directory2, filename), 'r') as fname:
output = open(os.path.join(path, directory+"/"+directory2, "output_disambiguation.txt"), 'w')
ambiguous_words = []
ambiguous_lines = []
text_words = []
for line in fname:
split_line = line.split()
text_words.append(split_line)
l = line.split()
if l[4] == "NN" or l[4] == "NNP":
if len(wordnet.synsets(l[3], "n")) > 1:
number_of_ss += len(wordnet.synsets(l[3], "n"))
synsets.append(len(wordnet.synsets(l[3], "n")))
ambiguous_words.append((l[2], l[3]))
for word in ambiguous_words:
start = int(str(word[0][0]) + "001")
end = start + 999
lines = []
for l in text_words:
if start <= int(l[2]) <= end:
lines.append(l[3])
ambiguous_lines.append(lines)
for i in range(len(ambiguous_words)):
ss = lesk(ambiguous_lines[i], ambiguous_words[i][1], "n")
outputwrite = str((ss, ss.definition())) + "\n"
output.write(outputwrite)
c = Counter(synsets)
print(sorted(c.items(), key=lambda pair: pair[0], reverse=True))
print(number_of_ss)
def mostSimilarCategory(theSense, theCategories):
maxValue = 0
catName = ''
posList = None
for cat in theCategories:
for key, values in cat.iteritems():
posList = nltk.pos_tag(values)
index = 0
contextSentence = ' '.join(values)
for word in values:#this is each word in the category
catSense = wsd.lesk(contextSentence, word)
if catSense:
similarityValue = wn.path_similarity(catSense, theSense)
if similarityValue!=None:
if similarityValue > maxValue:
maxValue = similarityValue
catName = key
index += 1
return catName
def getSenseLocs(words, sentence):
senseLocs = {}
token_sent = nltk.word_tokenize(sentence)
positions = matcher.getPositions(words, token_sent)
tagged = [(word, get_wordnet_pos(pos)) for word, pos in nltk.pos_tag(token_sent)]
for key, value in positions.items():
word = tagged[value][0]
pos = tagged[value][1]
#print(word)
#print(pos)
syns = wordnet.synsets(word, pos=pos)
#print(syns)
#print(token_sent)
if len(syns) > 0:
sense = lesk(token_sent, word, pos)
if sense:
senseLocs[str(pos)+"_senseLoc"] = syns.index(sense)
return senseLocs
def combineTags(sentence):
""" Adds the tags to the testfile """
#pbar = ProgressBar()
refDict=defaultdict(list)
sentList = []
#for sentence in taggedText:
for i, words in enumerate(sentence):
if i != 0:
prevword = sentence[i-1][0]
prevtag = sentence[i-1][1]
else:
prevword=prevtag=''
if words[1] == prevtag:
newTuple = (newTuple[0]+' '+words[0],words[1])
sentList.pop()
sentList.append(newTuple)
refDict[i+1,words].append(newTuple)
else:
newTuple = (words[0],words[1])
sentList.append(newTuple)
refDict[i+1,words].append(newTuple)
for i, words in enumerate(sentList):
if i != 0 and i < len(sentList)-1:
sent=sentList[i-1][0],words[0],sentList[i+1][0]
if words[1]!= 'O':
mwords=words[0].replace(' ','_')
if len(wn.synsets(mwords, 'n')) > 1:
leskDec=lesk(word_tokenize(' '.join(sent)), mwords, 'n')
for value in refDict.values():
if value[0] == words and len(value) < 2:
value.append([mwords,leskDec,leskDec.definition()])
else:
for ss in wn.synsets(mwords, 'n'):
for value in refDict.values():
if value[0] == words and len(value) < 2:
value.append([mwords,ss, ss.definition()])
return refDict
import sys
import pprint
import nltk
from nltk.corpus import wordnet as wn
from nltk.wsd import lesk
from collections import Counter
meaning = []
# with open('resume.txt', 'r') as resume:
# for line in resume:
# #pprint.pprint(line.split())
# # tokens = nltk
# pprint.pprint(1)
if __name__ == '__main__':
job_des_words = []
resume_words = []
if(3 == len(sys.argv)):
script, job_des_file, resume_file = sys.argv
with(open(job_des_file)) as jdf:
description = jdf.read()
job_des_words = description.split(" ")
with(open(resume_file)) as rf:
resume = rf.read()
resume_words = resume.split(" ")
print(job_des_words[0])
print(lesk(description, job_des_words[0]))
else:
print("No input present, exiting.")
def disAmbi(sents, ambi_nouns):
noun_dict = dict()
for noun in ambi_nouns:
noun_dict[noun] = lesk(sents, noun, "n")
return noun_dict
def similarity(string1,string2):
#split the string into sentences and sentences into words
sentences1=[d for d in re.split('\.\W',string1)]
sentences2=[d for d in re.split('\.\W',string2)]
#sentences1=[d for d in string1.split('|')]
#sentences2=[d for d in string2.split('|')]
sentences1=[d.split('|') for d in sentences1]
sentences1=[d for e in sentences1 for d in e]
sentences2=[d.split('|') for d in sentences2]
sentences2=[d for e in sentences2 for d in e]
#remove stop words
stop=stopwords.words('english')
#clean_sentence=[ j for j in sentence if j not in stop]
#sentence1=[j for j in sentence1 if j not in stop]
#sentence2=[j for j in sentence2 if j not in stop]
#remove Punctuation
regex=re.compile('[%s]' % re.escape(string.punctuation))
#remove_punct_map=dict((ord(char),None) for char in string.punctuation)
#regex.sub('', s)
#clean_sentence_npunct=[c.translate(remove_punct_map) for c in clean_sentence]
sentences1=[regex.sub('', c) for c in sentences1]
sentences2=[regex.sub('', c) for c in sentences2]
sentences1=[c.replace('\n','') for c in sentences1]
sentences2=[c.replace('\n','') for c in sentences2]
sentences1=[c.replace('\t','') for c in sentences1]
sentences2=[c.replace('\t','') for c in sentences2]
#print(sentences1)
#print(sentences2)
totalsimilarity=0;
for sent1 in sentences1:
sentence1=sent1.split()
for k,ktag in nltk.pos_tag(sentence1):
maxsimilarity=0
syn1=lesk(sentence1,k)
#print(k+" and"+ktag)
if syn1 is None:
#print("syn1 is none");
for i,sym1 in enumerate(wn.synsets(k)):
syn1=sym1
if syn1 is not None:
for sent2 in sentences2:
sentence2=sent2.split()
for j,jtag in nltk.pos_tag(sentence2):
# print(syn1.name()+" and "+syn2.name())
sim=0
a=0
#print(j+ " and " +jtag)
# if(ktag is jtag):
if(a==0):
syn2=lesk(sentence2,j)
#print(syn1.name()+" and "+syn2.name())
if syn2 is None:
#print("syn2 is none")
for i,sym2 in enumerate(wn.synsets(j)):
syn2=sym2
if syn2 is not None:
#print(j+"is none")
#else:
# print(syn1.name()+" and "+syn2.name())
ps=syn1.path_similarity(syn2)
ws=syn1.wup_similarity(syn2)
ls=0
if(syn1.name().split('.')[1] == syn2.name().split('.')[1]):
# print(syn1.name()+" and "+syn2.name())
ls=syn1.lch_similarity(syn2)
if ps is not None:
sim=sim+ps
if ws is not None:
sim=sim+ws
if ls is not None:
sim=sim+ls
if maxsimilarity < sim :
maxsimilarity=sim
totalsimilarity=totalsimilarity+maxsimilarity
#print ("\nscore"+str(totalsimilarity));
return totalsimilarity
def wiki_lookup(search_pass, tag_pass):
"""
This function looks up a word or bigram with a tag on wikipedia and returns the best possible results
:param search_pass: the word or bigram to lookup
:param tag_pass: The tag that belongs to the search_pass
:return: Returns a list with 3 elements, 3 links or less.
"""
search = search_pass
tag = tag_pass
search_lower = search.lower()
# These tags will return one link
tagcheck = ["COUNTRY", "STATE", "CITY", "TOWN", "NATURAL_PLACE", "PERSON", "ORGANISATION", "ANIMAL", "SPORT"]
# Since the link returned with president, is wrong often. This prevents president from being linked.
if search_lower != "president":
# If the search contains just one word.
if len(search.split(" ")) == 1:
# Try to get synset of the search, if not possible set synset to None
try:
search_syn = wordnet.synsets(search, pos="n")[0]
search_syn = str(search_syn)
except IndexError:
search_syn = None
# If the search contains multiple words, replace the spaces with _
else:
search_clean = search.split(" ")
search_clean = "_".join(search_clean)
syn = wordnet.synsets(search_clean, pos="n")
if len(syn) == 0:
search_syn = None
else:
search_syn = str(wordnet.synsets(search_clean, pos="n")[0])
wiki_results = []
url_list = []
result_syns = []
to_return = []
# These tags wont be added to the wiki lookup
if tag != "NATURAL_PLACE" and tag != "ANIMAL" and tag != "ENTERTAINMENT" and tag != "COUNTRY" and tag != "CITY":
search = search+" "+tag
search_results = wikipedia.search(search)
else:
search_results = wikipedia.search(search)
# If search results are found.
if len(search_results) != 0:
# Get a summary of all the results found.
for result in search_results:
try:
wiki_results.append([result, wikipedia.summary(result, sentences=2)])
except wikipedia.exceptions.DisambiguationError as e:
for result_e in e:
wiki_results.append([result_e, wikipedia.summary(result, sentences=2)])
except wikipedia.exceptions.PageError:
pass
# Cleanup the search results, so a synset can be created
for result in wiki_results:
result_words = result[0].split(" ")
if len(result_words) >= 1:
# Cleanup the search results
result_clean = "_".join(result_words)
# Lookup the synset of the search result, using the summary of the search word
ss = lesk(result[1], result_clean, "n")
try:
if ss == None:
result.append("-")
else:
result.append(str(ss))
result_syns.append(str(ss))
except AttributeError:
result.append("-")
result_syns.append("-")
else:
result.append("-")
# Create a url for all the search results
page = wikipedia.page(result[0])
result.append(page.url)
url_list.append(page.url)
print(search, search_results, url_list)
# If a synset was found, compare the found synset with the synset of the search.
if search_syn != None:
if search_syn in result_syns:
for result in wiki_results:
if result[2] == search_syn:
to_return = [result[3], "-", "-"]
# Else return the first link
else:
to_return = [url_list[0], "-", "-"]
# If the tag is in the list with tags that return one link, return one link
elif tag in tagcheck:
to_return = [url_list[0], "-", "-"]
# Else return up to 3 links, if possible
else:
if len(url_list) >= 3:
to_return = [url_list[0], url_list[1], url_list[2]]
elif len(url_list) == 2:
to_return = [url_list[0], url_list[1], "-"]
else:
to_return = [url_list[0], "-", "-"]
else:
to_return = ["-", "-", "-"]
else:
to_return = ["-", "-", "-"]
return to_return
def RecursiveGlossOverlap_Classify(text):
definitiongraphedges=defaultdict(list)
definitiongraphedgelabels=defaultdict(list)
#---------------------------------------------------------------------------------
#2.Compute intrinsic merit (either using linear or quadratic overlap)
#---------------------------------------------------------------------------------
tokenized = nltk.word_tokenize(text)
fdist1 = FreqDist(tokenized)
stopwords = nltk.corpus.stopwords.words('english')
stopwords = stopwords + [u' ',u'or',u'and',u'who',u'he',u'she',u'whom',u'well',u'is',u'was',u'were',u'are',u'there',u'where',u'when',u'may', u'The', u'the', u'In',u'in',u'A',u'B',u'C',u'D',u'E',u'F',u'G',u'H',u'I',u'J',u'K',u'L',u'M',u'N',u'O',u'P',u'Q',u'R',u'S',u'T',u'U',u'V',u'W',u'X',u'Y',u'Z']
puncts = [u' ',u'.', u'"', u',', u'{', u'}', u'+', u'-', u'*', u'/', u'%', u'&', u'(', ')', u'[', u']', u'=', u'@', u'#', u':', u'|', u';',u'\'s']
#at present tfidf filter is not applied
#freqterms1 = [w for w in fdist1.keys() if w not in stopwords and w not in puncts and (fdist1.freq(w) * compute_idf(corpus, w))]
freqterms1 = [w.decode("utf-8") for w in fdist1.keys() if w not in stopwords and w not in puncts]
current_level = 1
nodewithmaxparents = ''
noofparents = 0
maxparents = 0
relatedness = 0
first_convergence_level = 1
tokensofthislevel = []
convergingterms = []
convergingparents = []
tokensofprevlevel = []
prevlevelsynsets = []
commontokens = []
vertices = 0
edges = 0
overlap = 0
iter = 0
from nltk.corpus import wordnet as wn
#recurse down to required depth and update intrinsic merit score
#relatedness is either sum(overlaps) or sum((overlapping_parents)*(overlaps)^2) also called convergence factor
while current_level < 3:
#crucial - gather nodes which converge/overlap (have more than 1 parent)
if current_level > 1:
print current_level
for x in freqterms1:
for y in parents(x,prevlevelsynsets):
ylemmanames=y.lemma_names()
#for yl in ylemmanames:
# definitiongraphedges[x].append(yl)
definitiongraphedges[x].append(ylemmanames[0])
definitiongraphedgelabels[x + " - " + ylemmanames[0]].append(" is a subinstance of ")
definitiongraphedgelabels[ylemmanames[0] + " - " + x].append(" is a superinstance of ")
convergingterms = [w for w in freqterms1 if len(parents(w,prevlevelsynsets)) > 1]
for kw in freqterms1:
convergingparents = convergingparents + ([w for w in parents(kw, prevlevelsynsets) if len(parents(kw, prevlevelsynsets)) > 1])
for kw in freqterms1:
noofparents = len(parents(kw, prevlevelsynsets))
if noofparents > maxparents:
maxparents = noofparents
nodewithmaxparents = kw
for keyword in freqterms1:
#WSD - invokes Lesk's algorithm adapted to recursive gloss overlap- best_matching_synset()
#disamb_synset = best_matching_synset(set(doc1), wn.synsets(keyword))
if use_pywsd_lesk:
disamb_synset = simple_lesk(" ".join(freqterms1), keyword)
if use_nltk_lesk:
disamb_synset = lesk(freqterms1, keyword)
else:
disamb_synset = best_matching_synset(freqterms1, wn.synsets(keyword))
prevlevelsynsets = prevlevelsynsets + [disamb_synset]
if len(wn.synsets(keyword)) != 0:
disamb_synset_def = disamb_synset.definition()
tokens = nltk.word_tokenize(disamb_synset_def)
fdist_tokens = FreqDist(tokens)
#at present frequency filter is not applied
#if keyword in convergingterms:
tokensofthislevel = tokensofthislevel + ([w for w in fdist_tokens.keys() if w not in stopwords and w not in puncts and fdist_tokens.freq(w)])
listcount = len(tokensofthislevel)
setcount = len(set(tokensofthislevel))
overlap = listcount-setcount
if overlap > 0 and iter == 0 :
first_convergence_level = current_level
iter = 1
#choose between two relatedness/convergence criteria :-
#1) simple linear overlap or 2) zipf distributed quadratic overlap
#relatedness = relatedness + len(convergingparents)*overlap
relatedness = relatedness + overlap + len(convergingparents)
#relatedness = relatedness + ((len(convergingparents)*overlap*overlap) + 1)
#find out common tokens of this and previous level so that same token does not get grasped again -
#relatedness must be increased since repetition of keywords in two successive levels is a sign of
#interrelatedness(a backedge from child-of-one-of-siblings to one-of-siblings). Remove vertices and edges #corresponding to common tokens
commontokens = set(tokensofthislevel).intersection(set(tokensofprevlevel))
tokensofthislevel = set(tokensofthislevel).difference(commontokens)
relatedness = relatedness + len(commontokens)
#decrease the vertices count to address common tokens removed above - edges should remain same since they
#would just point elsewhere
vertices = vertices + setcount - len(commontokens)
edges = edges + listcount
current_level = current_level + 1
freqterms1 = set(tokensofthislevel)
tokensofprevlevel = tokensofthislevel
tokensofthislevel = []
intrinsic_merit = vertices*edges*relatedness / first_convergence_level
print definitiongraphedges
nxg=nx.DiGraph()
pos=nx.spring_layout(nxg)
#pos=nx.shell_layout(nxg)
#pos=nx.random_layout(nxg)
#pos=nx.spectral_layout(nxg)
#nx.draw_graphviz(nxg,prog="neato")
for k,v in definitiongraphedges.iteritems():
for l in v:
nxg.add_edge(k,l)
nxg.add_edge(l,k)
#nx.draw_networkx(nxg)
#plt.show()
nxg.remove_edges_from(nxg.selfloop_edges())
#print "Core number =",nx.core_number(nxg)
sorted_core_nxg=sorted(nx.core_number(nxg).items(),key=operator.itemgetter(1), reverse=True)
print "Core number (sorted) :",sorted_core_nxg
print "============================================================================================================="
print "Unsupervised Classification based on top percentile Core numbers of the definition graph(subgraph of WordNet)"
print "============================================================================================================="
no_of_classes=len(nx.core_number(nxg))
top_percentile=0
max_core_number=0
max_core_number_class=""
for n in sorted_core_nxg:
print "This document belongs to class:",n[0],",core number=",n[1]
if top_percentile < no_of_classes*0.50:
top_percentile+=1
else:
break
if n[1] > max_core_number:
max_core_number=n[1]
max_core_number_class=n[0]
print " max_core_number",max_core_number
print "==================================================================="
print "Betweenness Centrality of Recursive Gloss Overlap graph vertices"
print "==================================================================="
bc=nx.betweenness_centrality(nxg)
sorted_bc=sorted(bc.items(),key=operator.itemgetter(1),reverse=True)
print sorted_bc
print "==================================================================="
print "Closeness Centrality of Recursive Gloss Overlap graph vertices"
print "==================================================================="
cc=nx.closeness_centrality(nxg)
sorted_cc=sorted(cc.items(),key=operator.itemgetter(1),reverse=True)
print sorted_cc
print "==================================================================="
print "Degree Centrality of Recursive Gloss Overlap graph vertices"
print "==================================================================="
dc=nx.degree_centrality(nxg)
sorted_dc=sorted(dc.items(),key=operator.itemgetter(1),reverse=True)
print sorted_dc
print "==================================================================="
print "Page Rank of the vertices of RGO Definition Graph (a form of Eigenvector Centrality)"
print "==================================================================="
sorted_pagerank_nxg=sorted(nx.pagerank(nxg).items(),key=operator.itemgetter(1),reverse=True)
print sorted_pagerank_nxg
return (sorted_core_nxg, sorted_pagerank_nxg)
