def getInternalSetWithReducing(arr_words, target_word, model):
"""
Get IntS (internal, kernel) words for words in the array arr_words.
If |IntS (arr_words)| == 0 then try reduce arr_words,
(1) until |IntS (reduced arr_words)| > 0
(2) and set (reduced arr_words) contains target_word
Parameters
----------
arr_words : array of Strings
Array of source words, for example a synonym set or a sentence' words.
target_word : String
(1) arr_words contain target_word,
(2) the result IntS should contain target_word too.
model : object
Word2Vec model.
Returns
-------
array of Strings
Internal subset for the source list of words (arr_words) or for subset of arr_words.
Empty array if there are no such words.
"""
#result_int_s = []
arr_words = filter_vocab_words.filterVocabWords(arr_words, model.vocab)
#print string_util.joinUtf8( ",", arr_words ) # after filter, now there are only words with vectors
#if len(arr_words) < 3:
# return []
while len(
arr_words
) >= 3: # it is possible calculate IntS only when there are >= 3 words
int_s = getInternalSet(arr_words, model)
if len(int_s) > 0:
return int_s
# then now: len (int_s) == 0
# let's find word_remote (1) within arr_words, (2) the most distant word to the target word
target_vector = model[target_word]
word_remote = ""
#arr_new = []
sim_min = 1.0
for word in arr_words:
if word == target_word:
continue # let's skip and do not delete target word itself
vector = model[word]
sim = 1 - spatial.distance.cosine(target_vector, vector)
#print u"sim({}, {}) = {}".format( target_word, word, sim )
if sim < sim_min:
#print u"UPDATE: new sim {} < sim_min {}, word_remote old = {}, new = {}".format( sim, sim_min, word_remote, word )
sim_min = sim
word_remote = word
#print
if len(word_remote
) == 0: # it is very strange that we did not find any word!
return []
arr_words.remove(word_remote)
print string_util.joinUtf8(",", arr_words)
return []
def getInternalSet( arr_words, model):
"""
Get IntS (internal, kernel) words for words in the array arr_words.
Word w belong to IntS iff
(1) set S = S1 and S2 (subsets of S),
(2) subsets S1 and S2 are not empty,
(3) dist(S1, S2) > dist(S1 + w, S2),
(4) dist(S1, S2) > dist(S1 , S2 + w),
(where "dist" is Cosine similarity between average vectors of S1 and S2 - subsets of vectors),
i.e. the word w makes close any subsets S1 and S2 after adding.
Parameters
----------
arr_words : array of Strings
Array of source words, for example a ynonym set or a sentence' words.
model : object
Word2Vec model.
Returns
-------
array of Strings
Internal subset for the source list of words.
Empty array if there are no such words.
"""
result_int_s = []
DEBUG_PRINT = False #True
arr_words = filter_vocab_words.filterVocabWords( arr_words, model.vocab )
#print string_util.joinUtf8( ",", arr_words ) # after filter, now there are only words with vectors
len_words = len(arr_words)
#print "len_words = {}".format( len_words )
if len_words < 3:
return [] # it is possible calculate IntS only when there are >= 3 words
# current_synset = lib.synset.Synset()
# current_synset.headword = arr_words[0] # let's first element in synset is a headword (? target word)
# current_synset.line = line
# syn_rank = dict() # integer
# syn_centrality = dict() # float
# syn_internal = dict() # boolean (true for IntS, for synonyms, which always make subsets more nearer)
# let's take all subsets for every 'out' element
i=0
while (i < len_words):
gr = arr_words[:]
# extract the element 'out' which is under consideration
test_word = gr.pop(i)
#test_word_counter_int = 0
#test_word_counter_float = 0
sim12_greater_sim0_always = True
for j in range(0, len(gr)):
for l in range(j, len(gr)-1):
gr1 = gr[j:l+1]
gr2 = gr[0:j]+gr[l+1:len(gr)]
if DEBUG_PRINT:
print u"{} | gr1={} | gr2={}".format( test_word, string_util.joinUtf8( ",", gr1 ),
string_util.joinUtf8( ",", gr2 ) )
gr1_and_test_word = gr1[:]
gr1_and_test_word.append( test_word )
gr2_and_test_word = gr2[:]
gr2_and_test_word.append( test_word )
sim0 = model.n_similarity(gr1, gr2)
sim1 = model.n_similarity(gr1_and_test_word, gr2)
sim2 = model.n_similarity(gr1, gr2_and_test_word)
if DEBUG_PRINT:
print "sim0 = {:5.3f}".format( sim0 )
print "sim1 = {:5.3f}".format( sim1 )
print "sim2 = {:5.3f}".format( sim2 )
if sim0 > sim1 or sim0 > sim2:
sim12_greater_sim0_always = False
if DEBUG_PRINT:
a = 1 if sim1 > sim0 else -1
b = 1 if sim2 > sim0 else -1
#test_word_counter_int += (a + b)/2
#test_word_counter_float += (sim1 - sim0) + (sim2 - sim0)
#print "test_word_counter_int = {}".format( test_word_counter_int )
#print "test_word_counter_float = {}".format( test_word_counter_float )
if DEBUG_PRINT:
print ("---")
#syn_rank [test_word] = test_word_counter_int;
#syn_centrality[test_word] = test_word_counter_float;
#syn_internal [test_word] = sim12_greater_sim0_always;
if sim12_greater_sim0_always:
result_int_s.append( test_word )
if DEBUG_PRINT:
print ("+++++++")
print
i += 1
return result_int_s
def getInternalSet(arr_words, model):
"""
Get IntS (internal, kernel) words for words in the array arr_words.
Word w belong to IntS iff
(1) set S = S1 and S2 (subsets of S),
(2) subsets S1 and S2 are not empty,
(3) dist(S1, S2) > dist(S1 + w, S2),
(4) dist(S1, S2) > dist(S1 , S2 + w),
(where "dist" is Cosine similarity between average vectors of S1 and S2 - subsets of vectors),
i.e. the word w makes close any subsets S1 and S2 after adding.
Parameters
----------
arr_words : array of Strings
Array of source words, for example a ynonym set or a sentence' words.
model : object
Word2Vec model.
Returns
-------
array of Strings
Internal subset for the source list of words.
Empty array if there are no such words.
"""
result_int_s = []
DEBUG_PRINT = False #True
arr_words = filter_vocab_words.filterVocabWords(arr_words, model.vocab)
#print string_util.joinUtf8( ",", arr_words ) # after filter, now there are only words with vectors
len_words = len(arr_words)
#print "len_words = {}".format( len_words )
if len_words < 3:
return [
] # it is possible calculate IntS only when there are >= 3 words
# current_synset = lib.synset.Synset()
# current_synset.headword = arr_words[0] # let's first element in synset is a headword (? target word)
# current_synset.line = line
# syn_rank = dict() # integer
# syn_centrality = dict() # float
# syn_internal = dict() # boolean (true for IntS, for synonyms, which always make subsets more nearer)
# let's take all subsets for every 'out' element
i = 0
while (i < len_words):
gr = arr_words[:]
# extract the element 'out' which is under consideration
test_word = gr.pop(i)
#test_word_counter_int = 0
#test_word_counter_float = 0
sim12_greater_sim0_always = True
for j in range(0, len(gr)):
for l in range(j, len(gr) - 1):
gr1 = gr[j:l + 1]
gr2 = gr[0:j] + gr[l + 1:len(gr)]
if DEBUG_PRINT:
print u"{} | gr1={} | gr2={}".format(
test_word, string_util.joinUtf8(",", gr1),
string_util.joinUtf8(",", gr2))
gr1_and_test_word = gr1[:]
gr1_and_test_word.append(test_word)
gr2_and_test_word = gr2[:]
gr2_and_test_word.append(test_word)
sim0 = model.n_similarity(gr1, gr2)
sim1 = model.n_similarity(gr1_and_test_word, gr2)
sim2 = model.n_similarity(gr1, gr2_and_test_word)
if DEBUG_PRINT:
print "sim0 = {:5.3f}".format(sim0)
print "sim1 = {:5.3f}".format(sim1)
print "sim2 = {:5.3f}".format(sim2)
if sim0 > sim1 or sim0 > sim2:
sim12_greater_sim0_always = False
if DEBUG_PRINT:
a = 1 if sim1 > sim0 else -1
b = 1 if sim2 > sim0 else -1
#test_word_counter_int += (a + b)/2
#test_word_counter_float += (sim1 - sim0) + (sim2 - sim0)
#print "test_word_counter_int = {}".format( test_word_counter_int )
#print "test_word_counter_float = {}".format( test_word_counter_float )
if DEBUG_PRINT:
print("---")
#syn_rank [test_word] = test_word_counter_int;
#syn_centrality[test_word] = test_word_counter_float;
#syn_internal [test_word] = sim12_greater_sim0_always;
if sim12_greater_sim0_always:
result_int_s.append(test_word)
if DEBUG_PRINT:
print("+++++++")
print
i += 1
return result_int_s
def getInternalSetWithReducing( arr_words, target_word, model):
"""
Get IntS (internal, kernel) words for words in the array arr_words.
If |IntS (arr_words)| == 0 then try reduce arr_words,
(1) until |IntS (reduced arr_words)| > 0
(2) and set (reduced arr_words) contains target_word
Parameters
----------
arr_words : array of Strings
Array of source words, for example a synonym set or a sentence' words.
target_word : String
(1) arr_words contain target_word,
(2) the result IntS should contain target_word too.
model : object
Word2Vec model.
Returns
-------
array of Strings
Internal subset for the source list of words (arr_words) or for subset of arr_words.
Empty array if there are no such words.
"""
#result_int_s = []
arr_words = filter_vocab_words.filterVocabWords( arr_words, model.vocab )
#print string_util.joinUtf8( ",", arr_words ) # after filter, now there are only words with vectors
#if len(arr_words) < 3:
# return []
while len(arr_words) >= 3: # it is possible calculate IntS only when there are >= 3 words
int_s = getInternalSet (arr_words, model)
if len( int_s ) > 0:
return int_s
# then now: len (int_s) == 0
# let's find word_remote (1) within arr_words, (2) the most distant word to the target word
target_vector = model [ target_word ]
word_remote = ""
#arr_new = []
sim_min = 1.0
for word in arr_words:
if word == target_word:
continue # let's skip and do not delete target word itself
vector = model [ word ]
sim = 1 - spatial.distance.cosine( target_vector, vector )
#print u"sim({}, {}) = {}".format( target_word, word, sim )
if sim < sim_min:
#print u"UPDATE: new sim {} < sim_min {}, word_remote old = {}, new = {}".format( sim, sim_min, word_remote, word )
sim_min = sim
word_remote = word
#print
if len( word_remote ) == 0: # it is very strange that we did not find any word!
return []
arr_words.remove( word_remote )
print string_util.joinUtf8( ",", arr_words )
return []
