def test_similarity_two_texts(self):
# First preprocess documents
doc1 = "This is test. Aeroplane flight"
doc2 = "This is test"
doc1 = preprocess(doc1)
doc2 = preprocess(doc2)
similarity = self.similarity_model.documents_similarity(doc1, doc2)
assert isinstance(similarity, float)
def assign_cluster_to_doc(doc_id):
doc = ClassifiedDocument.objects.get(id=doc_id)
grp_id = doc.group_id
cluster_model = ClusteringModel.objects.get(group_id=grp_id)
model = cluster_model.model # instance of KMeansDocs class
processed = preprocess(doc.text)
X = model.vectorizer.transform([processed]).toarray()[0]
label = int(model.model.predict([X])[0])
docs_labels = [(doc_id, label)]
feature = compress_sparse_vector(list(map(lambda x: float(x), X)))
features = {doc_id: feature}
# update labels data
write_cluster_labels_data(cluster_model,
docs_labels,
features,
update=True)
# calculate new silhouette score
silhouette_score = cluster_model.calculate_silhouette_score()
cluster_model.silhouette_score = silhouette_score
cluster_model.save()
# update_size vs silhouette scores
update_cluster_score_vs_size(cluster_model, 1) # increased size is 1
if silhouette_score < SILHOUETTE_THRESHOLD:
logger.warning(
"cluster score for cluster {} reached below threshold {}. Re-clustering."
.format( # noqa
cluster_model.id, SILHOUETTE_THRESHOLD))
update_cluster(cluster_model.id)
return True
def update_cluster(cluster_id):
try:
cluster_model = ClusteringModel.objects.get(id=cluster_id)
except ClusteringModel.DoesNotExist:
logger.warn(
"Clustering Model with id {} does not exist".format(cluster_id))
if cluster_model.all_clustered:
# if all clustered, no need to recluster
return
docs = get_unclustered_docs(cluster_model)
increased_size = docs.count()
texts = list(
map(lambda x: preprocess(x['text'], ignore_numbers=True), docs))
docids = list(map(lambda x: x['id'], docs))
kmeans_model = cluster_model.model
CLUSTER_CLASS = type(kmeans_model)
# TODO: add update criteria for doc2vec
# update status
# cluster_model.ready = False
cluster_model.last_clustering_started = timezone.now()
cluster_model.save()
kmeans_model.update_cluster(texts)
docs_labels = list(
zip(
docids, # convert from np.int64 to int
list(map(lambda x: int(x), kmeans_model.model.labels_))))
if CLUSTER_CLASS == KMeansDocs:
features = []
vectorizer = kmeans_model.vectorizer
for txt in texts:
arr = vectorizer.transform([txt]).toarray()[0]
compressed = compress_sparse_vector(arr)
features.append(compressed)
# write/update to file
docids_features = dict(zip(docids, features))
write_clustered_data_to_files(cluster_model,
docs_labels,
kmeans_model.model.cluster_centers_,
docids_features,
update=True)
# relevant terms can be calculated only after writing other data
relevant_terms = cluster_model.compute_relevant_terms()
write_relevant_terms_data(cluster_model, relevant_terms)
# update status
cluster_model.last_clustered_on = timezone.now()
cluster_model.silhouette_score = cluster_model.calculate_silhouette_score()
cluster_model.ready = True
cluster_model.save()
# Update size vs silhouette scores
update_cluster_score_vs_size(cluster_model, increased_size)
def text_rank(sentences):
"""
@sentences: list of sentences
"""
# first calculate the transition matrix using sentence similarity
size = len(sentences)
processed = [preprocess(x) for x in sentences]
matrix = [[1.0 for _ in range(size)] for _ in range(size)] # later updated
for i, s1 in enumerate(sentences):
for j in range(i+1, size):
# no need to check for i==j case as it is already set 1.0
similarity = sentences_similarity(processed[i], processed[j])
matrix[i][j] = similarity
matrix[j][i] = similarity
return page_rank(np.asarray(matrix))
def get_text_vector(self, text):
processed = preprocess(text, ignore_numbers=True)
text_terms = {}
# fill in text_terms first
for x in processed.split():
text_terms[x] = text_terms.get(x, 0.0) + 1
vector = [0.0] * self.terms_len
for k, v in text_terms.items():
termid = self.terms_indices.get(k)
# only if the term is already in our index
if termid:
inv_freq = self.inverse_freqs.get(termid, 0)
idf = 0 if not inv_freq else\
math.log(self.total_docs/float(inv_freq))
vector[self.terms_indices[k]] = v * idf
return vector
def preprocess(inp):
return preprocess(inp)
def create_new_clusters(name,
group_id,
n_clusters,
CLUSTER_CLASS=KMeansDocs,
doc2vec_group_id=None):
"""If already exists, override it"""
try:
cluster_model = ClusteringModel.objects.get(group_id=group_id)
except ClusteringModel.DoesNotExist:
cluster_model = ClusteringModel.objects.create(name=name,
group_id=group_id,
n_clusters=n_clusters)
options = ClusteringOptions(n_clusters=n_clusters)
if CLUSTER_CLASS == KMeansDocs:
docs = ClassifiedDocument.objects.filter(group_id=group_id).\
values('id', 'text')
if not docs or docs.count() < n_clusters:
logger.warn(
"Too less documents for clustering for group_id {}".format(
group_id))
raise Exception("Too less documents for given number of clusters")
texts = list(
map(lambda x: preprocess(x['text'], ignore_numbers=True), docs))
docids = list(map(lambda x: x['id'], docs))
cluster_params = texts
elif CLUSTER_CLASS == KMeansDoc2Vec:
# get Doc2VecModel
doc2vecmodel = Doc2VecModel.objects.get(group_id=doc2vec_group_id)
cluster_params = [x for x in doc2vecmodel.model.docvecs]
docids = doc2vecmodel.model.docvecs.doctags.keys()
features = cluster_params
else:
raise Exception("Invalid class")
k_means = CLUSTER_CLASS(options)
logger.info("Creating clustering model for group_id {}".format(group_id))
kmeans_model = k_means.perform_cluster(cluster_params)
docs_labels = zip(
docids, # convert from np.int64 to int
list(map(lambda x: int(x), kmeans_model.model.labels_)))
# Save to database
cluster_model.model = kmeans_model
cluster_model.name = name
cluster_model.group_id = group_id
cluster_model.n_clusters = n_clusters
logger.info("Saving model to database. Group_id".format(group_id))
cluster_model.silhouette_score = kmeans_model.get_silhouette_score()
cluster_model.save()
# create features for KMeansDocs
if CLUSTER_CLASS == KMeansDocs:
features = []
vectorizer = kmeans_model.vectorizer
for txt in texts:
arr = vectorizer.fit_transform([txt]).toarray()[0]
compressed = compress_sparse_vector(arr)
features.append(compressed)
relevant_terms = get_relevant_terms(list(map(tokenize, texts)))
docids_features = dict(zip(docids, features))
# Now write to files
logger.info(
"Writing clustering results to files. Group id: {}".format(group_id))
write_clustured_data_to_files(cluster_model, docs_labels,
kmeans_model.model.cluster_centers_,
docids_features, relevant_terms)
# mark clustering complete as true
cluster_model.ready = True
cluster_model.save()
return cluster_model
def perform_clustering(cluster_model,
CLUSTER_CLASS=KMeansDocs,
doc2vec_group_id=None):
n_clusters = cluster_model.n_clusters
group_id = cluster_model.group_id
name = cluster_model.name
cluster_model.last_clustering_started = timezone.now()
cluster_model.save()
options = ClusteringOptions(n_clusters=n_clusters)
if CLUSTER_CLASS == KMeansDocs:
docs = ClassifiedDocument.objects.filter(group_id=group_id).\
values('id', 'text')
if not docs or docs.count() < n_clusters:
logger.warn(
"Too less documents for clustering for group_id {}".format(
group_id))
raise Exception("Too less documents for given number of clusters")
texts = list(
map(lambda x: preprocess(x['text'], ignore_numbers=True), docs))
docids = list(map(lambda x: x['id'], docs))
cluster_params = texts
elif CLUSTER_CLASS == KMeansDoc2Vec:
# get Doc2VecModel
doc2vecmodel = Doc2VecModel.objects.get(group_id=doc2vec_group_id)
cluster_params = [x for x in doc2vecmodel.model.docvecs]
docids = doc2vecmodel.model.docvecs.doctags.keys()
features = cluster_params
else:
raise Exception("Invalid class")
k_means = CLUSTER_CLASS(options)
logger.info("Creating clustering model for group_id {}".format(group_id))
kmeans_model = k_means.perform_cluster(cluster_params)
docs_labels = zip(
docids, # convert from np.int64 to int
list(map(lambda x: int(x), kmeans_model.model.labels_)))
# Save to database
cluster_model.model = kmeans_model
cluster_model.name = name
cluster_model.group_id = group_id
cluster_model.n_clusters = n_clusters
logger.info("Saving model to database. Group_id".format(group_id))
cluster_model.save()
# create features for KMeansDocs
if CLUSTER_CLASS == KMeansDocs:
features = []
vectorizer = kmeans_model.vectorizer
for txt in texts:
arr = vectorizer.transform([txt]).toarray()[0]
compressed = compress_sparse_vector(arr)
features.append(compressed)
docids_features = dict(zip(docids, features))
# Now write to files
logger.info(
"Writing clustering results to files. Group id: {}".format(group_id))
write_clustered_data_to_files(cluster_model, docs_labels,
kmeans_model.model.cluster_centers_,
docids_features)
# relevant terms can be calculated only after writing other data
relevant_terms = cluster_model.compute_relevant_terms()
write_relevant_terms_data(cluster_model, relevant_terms)
# mark clustering complete as true, and update clustered date
cluster_model.ready = True
cluster_model.silhouette_score = cluster_model.calculate_silhouette_score()
cluster_model.last_clustered_on = timezone.now()
cluster_model.save()
return cluster_model
def classify_text(classifier, text):
text = preprocess(text)
classified = classifier.classify_text(text)
classified.sort(key=lambda x: x[1], reverse=True)
return classified
