def triple_perplexity(documents: List[np.ndarray],
author_topic, author_topic_c,
category_topic, category_topic_c,
topic_word, topic_word_c) -> float:
"""
Calculates the perplexity based on the documents given
:param documents: a list of documents with word ids
:return: the perplexity of the documents given
"""
n = 0
ll = 0.0
doc2author = prepro_file_load("doc2author")
doc2category = prepro_file_load("doc2category")
for d, doc in documents:
author = doc2author[d]
category = doc2category[d]
for w in doc:
ll += np.log(((topic_word[:, w] / topic_word_c) *
(category_topic[category, :] / category_topic_c[category]) *
(author_topic[author, :] / author_topic_c[author])).sum())
n += 1
return np.exp(ll / (-n))
def compute_metrics_on_saved_model(save_model: str or Model, test_documents, multi: bool = False):
"""
Compute the three metrics (perplexity, coherence and topic diff) on a saved model
:param save_model: the name of the model
:param test_documents: the test documents we want to test on
:return: a dict containing the results
"""
if type(save_model) == str and not multi:
loaded_model = load_model(save_model)
elif type(save_model) == str and multi:
loaded_model = load_model(save_model, True)
else:
loaded_model = save_model
doc2bow, dictionary, texts = prepro_file_load('doc2bow'), prepro_file_load('corpora'), list(
prepro_file_load('doc2pre_text').values())
if not multi:
model_perplexity = perplexity(test_documents,
loaded_model.doc_topic,
loaded_model.doc_topic_count,
loaded_model.topic_word,
loaded_model.topic_word_count)
model_coherence = get_coherence(doc2bow, dictionary, texts, loaded_model.num_topics, loaded_model.topic_word)
model_topic_difference = mean_topic_diff(loaded_model.topic_word)
else:
model_perplexity = triple_perplexity(test_documents,
loaded_model.feature_topic,
loaded_model.feature_topic_count,
loaded_model.feature2_topic,
loaded_model.feature2_topic_count,
loaded_model.topic_word,
loaded_model.topic_word_count)
model_coherence = get_coherence(doc2bow, dictionary, texts, loaded_model.num_topics, loaded_model.topic_word)
model_topic_difference = mean_topic_diff(loaded_model.topic_word)
return {"perplexity": model_perplexity,
"coherence": model_coherence,
"topic_diff": model_topic_difference}
def print_metadata_documents(metadata_type: str,
metadata_name: str,
sample_size: int = 10,
print_top_topics=False):
# load necessary data
doc2pre = pre.prepro_file_load('doc2pre_text', folder_name='full')
doc2raw = pre.prepro_file_load('doc2raw_text', folder_name='full')
id2doc = pre.prepro_file_load('id2doc', folder_name='full')
doc2author = pre.prepro_file_load('doc2author', folder_name='full')
doc2category = pre.prepro_file_load('doc2category', folder_name='full')
doc2taxonomy = pre.prepro_file_load('doc2taxonomy', folder_name='full')
id2category = pre.prepro_file_load('id2category', folder_name='full')
id2author = pre.prepro_file_load('id2author', folder_name='full')
id2taxonomy = pre.prepro_file_load('id2taxonomy', folder_name='full')
if metadata_type == "category":
doc2meta = pre.prepro_file_load('doc2category', folder_name='full')
id2meta = id2category
elif metadata_type == "author":
doc2meta = pre.prepro_file_load('doc2author', folder_name='full')
id2meta = id2author
elif metadata_type == "taxonomy":
doc2meta = pre.prepro_file_load('doc2taxonomy', folder_name='full')
id2meta = id2taxonomy
else:
print(f"'{metadata_type}' not found!")
exit()
# get metadata ID from name (examples: '26. Frederik', 'System Administrator', 'EMNER')
metaID = find_id_from_value(id2meta, metadata_name)
# get document IDs for documents with the given metadata
documentIDs = get_metadata_document_ids(doc2meta, metaID)
documents = {}
documentsRaw = {}
docAuthors = {}
docCategories = {}
docTaxonomies = {}
docFileNames = {}
# get data based on document IDs
for docID in documentIDs:
documents[docID] = doc2pre[docID]
documentsRaw[docID] = doc2raw[docID]
docAuthors[docID] = doc2author[docID]
docCategories[docID] = doc2category[docID]
docTaxonomies[docID] = doc2taxonomy[docID]
docFileNames[docID] = id2doc[docID]
# document set information
print(f"{len(documents)} documents found\n")
print_meta_document_set_info(docAuthors, id2author, "authors")
print_meta_document_set_info(docCategories, id2category, "categories")
print_meta_document_set_info(docTaxonomies, id2taxonomy, "taxonomies")
# random examples of documents with metadata information
if print_top_topics:
corpora = pre.prepro_file_load("corpora", "full")
model_path = "../model/models/90_0.01_0.1_category"
model = load_model(model_path)
num_topics = model.num_topics
topic_word_dist = model.topic_word
topic_top_words = get_topics(corpora, num_topics, topic_word_dist)
print("Random documents:")
sampleIDs = random.sample(documentIDs, len(documentIDs))
for count in range(sample_size):
if count == len(sampleIDs):
break
id = sampleIDs[count]
print(f"ID: {id}")
print(f"Author: {id2author[docAuthors[id]]}")
print(f"Category: {id2category[docCategories[id]]}")
print(
f"Taxonomy: {[id2taxonomy[taxID] for taxID in docTaxonomies[id]]}")
print(f"File name: {docFileNames[id]}")
print(documents[id])
print(documentsRaw[id] + "\n")
if print_top_topics:
item_top_topics = sample_and_sort_items(model,
item_id=docCategories[id])
print(f"Top words in category top topics:")
for item in item_top_topics.items():
for topic in item[1]:
print(
f"Topic ID/probability: {topic[0]}/{'{:.2f}'.format(topic[1])} {topic_top_words[topic[0]]}"
)
print()
def hellinger_dis(p, q):
return norm(np.sqrt(p) - np.sqrt(q)) / _SQRT2
def js_sim(p, q):
return 1 - jensenshannon(p, q)
def hellinger_sim(p, q):
return 1 - hellinger_dis(p, q)
if __name__ == '__main__':
folder = "nice"
doc2auth = prepro_file_load("doc2author", folder_name=folder)
id2auth = prepro_file_load("id2author", folder_name=folder)
auth2doc = {}
for doc, auth in doc2auth.items():
auth2doc[auth] = auth2doc.get(auth, []) + [doc]
in_folder = "nice"
path = f"../model/generated_files/{in_folder}/"
with open(path + "wta.pickle", "rb") as file:
wta = pickle.load(file)
with open(path + "middle.pickle", "rb") as file:
middle = pickle.load(file)
with open(path + "topic_word.pickle", "rb") as file:
topic_word = pickle.load(file)