def fit(self, X, y):
"""
Fits TOPIC DEFINITIONS and topic labels to MSC Model.
Parameters
----------
X : The training input samples (Topic definitions) list of strings or wiki pages.
y : The target values. An array of int.s
Returns
-------
self : object
Returns self.
"""
# Adequating corpus for inference later
X = list(
doc_utils.prepare_corpus(X,
train_data=True,
preprocess=self.preprocess,
dataset_type=self.dataset_type))
self.model.build_vocab(X)
self.model.train(X,
total_examples=self.model.corpus_count,
epochs=self.model.epochs)
self.classes_ = unique_labels(y)
self.X_ = X
self.y_ = y
return self
def predict(self, X):
"""
Prediction function for max. similarity classifier.
Parameters
----------
X : list
The input dataset: Wiki or Arxiv dataset to classify
Returns
-------
y : ndarray, shape (n_samples,)
Prediction outputs .
"""
# Check is fit had been called
check_is_fitted(self, ['X_', 'y_'])
# Input validation
input_articles = list(
doc_utils.prepare_corpus(X,
train_data=False,
preprocess=self.preprocess,
dataset_type=self.dataset_type))
outputs = list()
for doc in input_articles:
inferred_vector = self.model.infer_vector(doc)
sims = self.model.docvecs.most_similar([inferred_vector],
topn=len(
self.model.docvecs))
most_similar_label = sims[0][0] # index 0 === most similar
outputs.append(most_similar_label)
pred_labels = np.array(outputs)
return pred_labels
def score(self, X, y, eval="weighted"):
"""
Score function that returns depending on "eval" parameter:
- top1: top-1 accuracy
- top2: top-2 accuracy
- weighted: weighted accuracy
of the predictions for articles "X" given true labels "y".
"""
accuracy_list = list()
outputs = list()
check_is_fitted(self, ['X_', 'y_'])
input_articles = list(
doc_utils.prepare_corpus(X,
train_data=False,
preprocess=self.preprocess,
dataset_type=self.dataset_type))
for i, doc in enumerate(input_articles):
inferred_vector = self.model.infer_vector(doc)
sims = self.model.docvecs.most_similar([inferred_vector],
topn=len(
self.model.docvecs))
most_similar_label = sims[0][0] # index 0 === most similar
outputs.append(most_similar_label)
second_most_similar_label = sims[1][0]
if most_similar_label == y[i]:
accuracy_list.append(1)
elif (second_most_similar_label == y[i] and "weighted" in eval):
accuracy_list.append(0.5)
elif (second_most_similar_label == y[i] and "top2" in eval):
accuracy_list.append(1)
else:
accuracy_list.append(0)
accuracy_list = np.array(accuracy_list)
# print("Model {} accuracy over {} test documents: {}%.".format(eval, len(y), np.mean(accuracy_list) * 100))
return np.mean(accuracy_list)
def wiki_pseudo_label(self,
x_train,
dataset,
result="extended",
top_n=2,
debug=False):
input_articles = list(
doc_utils.prepare_corpus(dataset,
train_data=False,
preprocess=self.preprocess,
dataset_type=self.dataset_type))
doc_topics_sims = [[], [], [], [], [], [], [], [], []]
article_list = list()
for topic in dataset:
for article in topic[0]:
article_list.append(article)
for doc_id, doc in enumerate(input_articles):
inferred_vector = self.model.infer_vector(doc)
sims = self.model.docvecs.most_similar([inferred_vector],
topn=len(
self.model.docvecs))
top_n_sims = sims[:top_n]
for i in range(top_n):
topic_id = top_n_sims[i][0] #predicted topic id of paper
topic_sim = top_n_sims[i][1] #similarity
doc_topics_sims[topic_id].append((topic_sim, doc_id))
best_articles_per_topic = [-1, -1, -1, -1, -1, -1, -1, -1, -1]
for i, topic in enumerate(doc_topics_sims):
article_id = (max(topic, key=lambda i: i[0])[1])
if debug:
print("For topic {}, best article id: {}".format(
i, article_id))
best_articles_per_topic[
i] = article_id #TODO: add topN papers and not just best one
x_train_ext = ["", "", "", "", "", "", "", "", ""]
x_train_articles = ["", "", "", "", "", "", "", "", ""]
for topic_id, train_sample in enumerate(x_train):
best_article_id = best_articles_per_topic[topic_id]
if debug:
print(
"For topic id {}, Adding article id {} to dataset".format(
topic_id, best_article_id))
print(article_list[best_article_id])
#Creating extended train data
if result in "extended":
x_train_ext[topic_id] = " . ".join(
[article_list[best_article_id], train_sample])
elif result in "bestpapers": #train samples are the best articles
x_train_articles[topic_id] = article_list[best_article_id]
else:
print(
"[ERROR] Result argument can be only 'extended' or 'bestpapers'."
)
return -1
if result in "extended":
new_x_train = x_train_ext
else: #best articles
new_x_train = x_train_articles
return new_x_train
def arxiv_pseudo_label(self,
x_train,
dataset,
paperslist,
result="extended",
top_n=1,
debug=False):
if paperslist is None:
print("ERROR: paperlist from Arxivparser needed")
return -1
input_articles = list(
doc_utils.prepare_corpus(dataset,
train_data=False,
preprocess=self.preprocess,
dataset_type=self.dataset_type))
doc_topics_sims = [[], [], [], [], [], [], [], []]
for doc_id, doc in enumerate(input_articles):
inferred_vector = self.model.infer_vector(doc)
sims = self.model.docvecs.most_similar([inferred_vector],
topn=len(
self.model.docvecs))
top_n_sims = sims[:top_n]
for i in range(top_n):
topic_id = top_n_sims[i][0] #predicted topic id of paper
topic_sim = top_n_sims[i][1] #similarity
doc_topics_sims[topic_id].append((topic_sim, doc_id))
best_papers_per_topic = [-1, -1, -1, -1, -1, -1, -1, -1]
n_papers_per_topic = len(paperslist) // len(
doc_utils.ARXIV_WIKI_TOPICS)
for i, topic in enumerate(doc_topics_sims):
paper_id = (max(topic, key=lambda i: i[0])[1])
best_papers_per_topic[
i] = paper_id #TODO: add topN papers and not just best one
if debug:
true_label = paper_id // n_papers_per_topic
print("Topic {} ({}) best matching paper: id #{}".format(
i, doc_utils.ARXIV_WIKI_TOPICS[i], paper_id))
print("\t--->True label:[", str(true_label), "](",
doc_utils.ARXIV_WIKI_TOPICS[true_label],
") \t\tPaper title:", paperslist[paper_id]['title'])
x_train_ext = ["", "", "", "", "", "", "", ""]
x_train_papers = ["", "", "", "", "", "", "", ""]
for topic_id, train_sample in enumerate(x_train):
best_paper_id = best_papers_per_topic[topic_id]
#Creating extended train data
if result in "extended":
x_train_ext[topic_id] = " . ".join([
train_sample, paperslist[best_paper_id]["title"],
paperslist[best_paper_id]["abstract"]
])
elif result in "bestpapers": #train samples are the best papers
x_train_papers[topic_id] = paperslist[best_paper_id][
"title"] + " : " + paperslist[best_paper_id]["abstract"]
else:
print(
"[ERROR] Result argument can be only 'extended' or 'bestpapers'."
)
return -1
if result in "extended":
new_x_train = x_train_ext
else: #best papers
new_x_train = x_train_papers
return new_x_train