def jensen_shannon_div(P, Q):
"""
Compute the Jensen-Shannon divergence between two probability distributions of equal length.
-----
:param P: Probability distributions that sum to 1
:param Q: Probability distributions that sum to 1
:return: float
"""
M = 0.5 * (P + Q)
# return 0.5 * (_kldiv(P, M) +_kldiv(Q, M))
return 0.5 * (kullback_leibler(P, M) + kullback_leibler(Q, M))
def rank_documents(model, model_name, type, query):
sims_list = []
processed_query = read_ap.process_text(query)
print(processed_query)
if model_name == "LSI":
if type == "bow":
# calculating cosine similarity for LSI (BoW)
index = gensim.similarities.MatrixSimilarity(model[corpus])
#make a bow representation of the query, and split the words
vec_bow = dictionary.doc2bow(processed_query)
vec_lsi = model[vec_bow] # convert the query to LSI space
sims = index[vec_lsi] # get index
sims = sorted(enumerate(sims), key=lambda item: -item[1])
# store the scores with the associated doc id's for the retrieval evaluation
doc_ids = list(new_docs.keys())
for i, s in sims:
sims_list.append((doc_ids[i], np.float64(s)))
return sims_list
if type == "tfidf":
#calculating cosine similarity for LSI, tf idf using similarities
#use the tfidf corpus -> lsi corpus
corpus_lsi = model[corpus_tfidf]
#transform corpus to LSI space and index it
index = gensim.similarities.MatrixSimilarity(corpus_lsi)
#convert query to lsi space via tf-idf
vec_bow = dictionary.doc2bow(processed_query)
vec_lsi = model[vec_bow]
sims = index[vec_lsi]
#same as with LSI BoW
sims = sorted(enumerate(sims), key=lambda item: -item[1])
doc_ids = list(new_docs.keys())
for i, s in sims:
sims_list.append((doc_ids[i], np.float64(s)))
return sims_list
else:
#calculating the negative Kullback–Leibler divergence scores for LDA
#transform query
vec_bow = dictionary.doc2bow(processed_query)
# transform query to the LDA space
vec_lda_query = model[vec_bow][0]
kl_divergence = []
for text in corpus:
#transform current document text in bow space to lda space
vec_lda_text = model[text][0]
# KL(Q||D) =\sum_w p(w|Q) log p(w|D) as explained in http://times.cs.uiuc.edu/course/410s11/kldir.pdf, using gensim mathutil
kl_divergence.append(kullback_leibler(vec_lda_query, vec_lda_text))
#sims = index[vec_lda]
#sort the kl scores
kl_divergence = sorted(enumerate(kl_divergence),
key=lambda item: -item[1])
doc_ids = list(new_docs.keys())
for i, s in kl_divergence:
sims_list.append((doc_ids[i], np.float64(s)))
return sims_list
def get_most_similar_documents(query, corpus, dictionary, k=10):
distances = []
for c in corpus:
distances.append(
kullback_leibler(query, c, num_features=len(dictionary)))
indices = np.array(distances).argsort()[:k]
return indices
def test_distributions(self):
# checking bag of words as inputs
vec_1 = [(2, 0.1), (3, 0.4), (4, 0.1), (5, 0.1), (1, 0.1), (7, 0.2)]
vec_2 = [(1, 0.1), (3, 0.8), (4, 0.1)]
result = matutils.kullback_leibler(vec_2, vec_1, 8)
expected = 0.55451775
self.assertAlmostEqual(expected, result)
# KL is not symetric; vec1 compared with vec2 will contain log of zeros and return infinity
vec_1 = [(2, 0.1), (3, 0.4), (4, 0.1), (5, 0.1), (1, 0.1), (7, 0.2)]
vec_2 = [(1, 0.1), (3, 0.8), (4, 0.1)]
result = matutils.kullback_leibler(vec_1, vec_2, 8)
self.assertTrue(math.isinf(result))
# checking ndarray, csr_matrix as inputs
vec_1 = numpy.array([[1, 0.3], [0, 0.4], [2, 0.3]])
vec_2 = csr_matrix([[1, 0.4], [0, 0.2], [2, 0.2]])
result = matutils.kullback_leibler(vec_1, vec_2, 3)
expected = 0.0894502
self.assertAlmostEqual(expected, result)
# checking ndarray, list as inputs
vec_1 = numpy.array([0.6, 0.1, 0.1, 0.2])
vec_2 = [0.2, 0.2, 0.1, 0.5]
result = matutils.kullback_leibler(vec_1, vec_2)
expected = 0.40659450877
self.assertAlmostEqual(expected, result)
# testing LDA distribution vectors
numpy.random.seed(0)
model = self.class_(self.corpus,
id2word=dictionary,
num_topics=2,
passes=100)
lda_vec1 = model[[(1, 2), (2, 3)]]
lda_vec2 = model[[(2, 2), (1, 3)]]
result = matutils.kullback_leibler(lda_vec1, lda_vec2)
expected = 4.283407e-12
self.assertAlmostEqual(expected, result)
def test_inputs(self):
# checking empty inputs
vec_1 = []
vec_2 = []
result = matutils.kullback_leibler(vec_1, vec_2)
expected = 0.0
self.assertEqual(expected, result)
# checking np array and list input
vec_1 = np.array([])
vec_2 = []
result = matutils.kullback_leibler(vec_1, vec_2)
expected = 0.0
self.assertEqual(expected, result)
# checking scipy csr matrix and list input
vec_1 = csr_matrix([])
vec_2 = []
result = matutils.kullback_leibler(vec_1, vec_2)
expected = 0.0
self.assertEqual(expected, result)
def test_inputs(self):
# checking empty inputs
vec_1 = []
vec_2 = []
result = matutils.kullback_leibler(vec_1, vec_2)
expected = 0.0
self.assertEqual(expected, result)
# checking numpy array and list input
vec_1 = numpy.array([])
vec_2 = []
result = matutils.kullback_leibler(vec_1, vec_2)
expected = 0.0
self.assertEqual(expected, result)
# checking scipy csr matrix and list input
vec_1 = csr_matrix([])
vec_2 = []
result = matutils.kullback_leibler(vec_1, vec_2)
expected = 0.0
self.assertEqual(expected, result)
def test_distributions(self):
# checking bag of words as inputs
vec_1 = [(2, 0.1), (3, 0.4), (4, 0.1), (5, 0.1), (1, 0.1), (7, 0.2)]
vec_2 = [(1, 0.1), (3, 0.8), (4, 0.1)]
result = matutils.kullback_leibler(vec_2, vec_1, 8)
expected = 0.55451775
self.assertAlmostEqual(expected, result)
# KL is not symetric; vec1 compared with vec2 will contain log of zeros and return infinity
vec_1 = [(2, 0.1), (3, 0.4), (4, 0.1), (5, 0.1), (1, 0.1), (7, 0.2)]
vec_2 = [(1, 0.1), (3, 0.8), (4, 0.1)]
result = matutils.kullback_leibler(vec_1, vec_2, 8)
self.assertTrue(math.isinf(result))
# checking ndarray, csr_matrix as inputs
vec_1 = np.array([[1, 0.3], [0, 0.4], [2, 0.3]])
vec_2 = csr_matrix([[1, 0.4], [0, 0.2], [2, 0.2]])
result = matutils.kullback_leibler(vec_1, vec_2, 3)
expected = 0.0894502
self.assertAlmostEqual(expected, result)
# checking ndarray, list as inputs
vec_1 = np.array([0.6, 0.1, 0.1, 0.2])
vec_2 = [0.2, 0.2, 0.1, 0.5]
result = matutils.kullback_leibler(vec_1, vec_2)
expected = 0.40659450877
self.assertAlmostEqual(expected, result)
# testing LDA distribution vectors
np.random.seed(0)
model = self.class_(self.corpus, id2word=dictionary, num_topics=2, passes= 100)
lda_vec1 = model[[(1, 2), (2, 3)]]
lda_vec2 = model[[(2, 2), (1, 3)]]
result = matutils.kullback_leibler(lda_vec1, lda_vec2)
expected = 4.283407e-12
self.assertAlmostEqual(expected, result)
def search(self, query):
query_repr = read_ap.process_text(query)
vec_query = self.corpus.dictionary.doc2bow(query_repr)
lda_query = sparse2full(self.model[vec_query], self.num_topics)
results = defaultdict(float)
for doc_id, lda_doc_repr in zip(self.corpus.doc_ids,
self.lda_corpus_pers):
results[doc_id] = kullback_leibler(lda_query, lda_doc_repr)
results = {
k: v
for k, v in sorted(
results.items(), key=lambda item: item[1], reverse=True)
}
return list(results.items())
def ranking_LDA(query, model, model_docs, num_topics=10):
scores = []
# Process query to correct KL divergence form
query = read_ap.process_text(query)
query = dictionary.doc2bow(query)
query = model[query]
query = gensim.matutils.sparse2full(query, num_topics)
# Calculate KL divergence for each document in the corpus
for i in range(len(corpus)):
doc = model_docs[i]
neg_kl = float(-1 * kullback_leibler(query, doc))
scores.append((i2str[i], neg_kl))
# Sort on second tuple value
scores = sorted(scores, key=lambda x: x[1], reverse=True)
return scores
def get_sims(model, query, corpus_full, dictionary, n_topics):
''' get ranking for single query '''
# avoid division by 0
eps = 1e-8
# process query
query_processed = read_ap.process_text(query)
query_bow = dictionary.doc2bow(query_processed)
q_lda = sparse2full(model[query_bow], n_topics)
q_lda += eps
sims = []
# loop over all docs
for i, doc in enumerate(corpus_full):
doc += eps
sim = -1 * kullback_leibler(q_lda, doc)
sims.append(sim)
sim_ordered = sorted(enumerate(sims), key=lambda item: -1 * item[1])
return sim_ordered
def rank_docs(query, model, doc_ids, dictionary, corpus_modelspace, tfidf_model=None, index=None):
query_prepro = read_ap.process_text(query)
# transform query to bow vector space
q_cspace = dictionary.doc2bow(query_prepro)
if not tfidf_model == None:
# transform query to tfidf vector space
q_cspace = tfidf_model[q_cspace]
q_modelspace = model[q_cspace]
if isinstance(model, LsiModel):
## LSI
scores = index[q_modelspace]
results = defaultdict(float)
for doc_id, score in zip(doc_ids, scores):
results[doc_id] = score
results = list(results.items())
results.sort(key=lambda _: -_[1])
elif isinstance(model, LdaModel):
## LDA
doc_ids = list(doc_ids)
scores = []
# have to use the for loop, otherwise kullback_leibler has problems
for d in corpus_modelspace:
scores.append(float(-kullback_leibler(q_modelspace, d)))
# have to use torch here to do this more efficiently
order = torch.Tensor(scores).argsort(descending=True).numpy()
ordered_results = [(doc_ids[i], scores[i]) for i in order]
results = dict(ordered_results)
return results
for word in file:
filecontent = filecontent + word + ' '
documents.append(filecontent)
stoplist = set(stopwords.words('english'))
texts = [[
word for word in document.lower().split() if word not in stoplist
] for document in documents]
basetext = []
for list in texts:
for item in list:
basetext.append(item)
bow_1 = lda.id2word.doc2bow(basetext)
lda_1 = lda[bow_1]
print("******************", filename)
print("hellinger", hellinger(lda_1, lda_2))
print("kullback_leibler", kullback_leibler(lda_1, lda_2))
print("jaccard", jaccard(lda_1, lda_2))
file.close()
#dictionary = corpora.Dictionary(texts)
#corpus = [dictionary.doc2bow(text) for text in texts]
#lda1 = gensim.models.ldamodel.LdaModel(corpus=corpus, id2word=dictionary, num_topics=5, update_every=1, chunksize=10000, passes=5)
#print(lda1)
#print(texts)
"""
basetext=[]
for list in texts:
for item in list:
basetext.append(item)
#print(len(basetext))
#print(basetext)
lda_bow_water = model[bow_water]
lda_bow_finance = model[bow_finance]
lda_bow_bank = model[bow_bank]
tfidf_bow_water = tfidf[bow_water]
tfidf_bow_finance = tfidf[bow_finance]
tfidf_bow_bank = tfidf[bow_bank]
from gensim.matutils import kullback_leibler, jaccard, hellinger
hellinger(lda_bow_water, lda_bow_finance)
hellinger(lda_bow_finance, lda_bow_bank)
hellinger(lda_bow_bank, lda_bow_water)
hellinger(lda_bow_finance, lda_bow_water)
kullback_leibler(lda_bow_water, lda_bow_bank)
kullback_leibler(lda_bow_bank, lda_bow_water)
jaccard(bow_water, bow_bank)
jaccard(doc_water, doc_bank)
jaccard(['word'], ['word'])
def make_topics_bow(topic):
# takes the string returned by model.show_topics()
# split on strings to get topics and the probabilities
topic = topic.split('+')
# list to store topic bows
topic_bow = []
for word in topic:
# split probability and word
print('Coherence: {}\n'.format(c_mean))
with open('data/' + model_no + '/evaluation.txt', 'w') as f:
f.write('Coherence: {}\n'.format(c_mean))
#%%
# トピック間のカルバック・ライブラー距離を算出し描画する
# トピックごとの、辞書内の単語を含む文がそのトピックに分類される (事後) 確率
t = model.state.get_lambda()
# トピック同士の確率分布の距離を表示 (離れているほど良い)
ds = []
for i in range(model.num_topics):
for j in range(model.num_topics):
if i != j:
kl = kullback_leibler(t[i], t[j])
# print('{:02}-{:02}: {}'.format(i, j, kl))
ds.append(kl)
# グラフを保存
plt.title('KL-divergence')
plt.hist(ds)
plt.savefig('data/' + model_no + '/kl-divergence.png')
# 平均値をテキスト保存
d_mean = mean([d.astype(float) for d in ds])
print('KL-divergence: {}\n'.format(d_mean))
with open('data/' + model_no + '/evaluation.txt', 'a') as f:
f.write('KL-divergence: {}\n'.format(d_mean))
############################################################################### # Makes sense, right? In the first example, Document 1 and Document 2 are hardly similar, so we get a value of roughly 0.5. # # In the second case, the documents are a lot more similar, semantically. Trained with the model, they give a much less distance value. # ############################################################################### # Kullback–Leibler # ---------------- # # Let's run similar examples down with Kullback Leibler. # from gensim.matutils import kullback_leibler print(kullback_leibler(lda_bow_water, lda_bow_bank)) print(kullback_leibler(lda_bow_finance, lda_bow_bank)) ############################################################################### # .. important:: # KL is not a Distance Metric in the mathematical sense, and hence is not # symmetrical. This means that ``kullback_leibler(lda_bow_finance, # lda_bow_bank)`` is not equal to ``kullback_leibler(lda_bow_bank, # lda_bow_finance)``. # # As you can see, the values are not equal. We'll get more into the details of # this later on in the notebook. # print(kullback_leibler(lda_bow_bank, lda_bow_finance)) ###############################################################################
def simlarity_kullback_leibler(lda_vec1, lda_vec2):
return kullback_leibler(lda_vec1, lda_vec2)
def kldivergence_distance(self, x, y):
replaceZero = 0.000001
""" return KL-divergence between two lists """
return kullback_leibler([replaceZero if e == 0 else e for e in x],
[replaceZero if e == 0 else e for e in y])
