topicModelingTickets/topicModeling.py

473 lines
14 KiB
Python

# -*- coding: utf-8 -*-
from datetime import datetime
import time
import numpy as np
import csv
import sys
import json
import os.path
import subprocess
from textacy import Vectorizer, viz
from miscellaneous import *
import textacy
from scipy import *
import os
csv.field_size_limit(sys.maxsize)
FILEPATH = os.path.dirname(os.path.realpath(__file__)) + "/"
# load config
config_ini = FILEPATH + "config.ini"
config = ConfigParser.ConfigParser()
with open(config_ini) as f:
config.read_file(f)
def label2ID(label, labeldict):
return labeldict.get(label, len(labeldict))
def generate_lablelID_lines(textacyCorpus, labeldict):
for doc in textacyCorpus:
# generate [topic1, topic2....] tok1 tok2 tok3 out of corpi
yield "[" + str(label2ID(doc.metadata["categoryName"], labeldict)) + "] " + doc.text
"""
def printvecotorization(de_corpus, ngrams=1, min_df=1, max_df=1.0, weighting='tf', named_entities=True):
logprint(str("ngrams: {0}".format(ngrams)))
logprint(str("min_df: {0}".format(min_df)))
logprint(str("max_df: {0}".format(max_df)))
logprint(str("named_entities: {0}".format(named_entities)))
# printlog("vectorize corpi...")
vectorizer = Vectorizer(weighting=weighting, min_df=min_df, max_df=max_df)
terms_list = (doc.to_terms_list(ngrams=ngrams, named_entities=named_entities, as_strings=True) for doc in de_corpus)
doc_term_matrix = vectorizer.fit_transform(terms_list)
id2term = vectorizer.__getattribute__("id_to_term")
for t in terms_list:
print(t)
logprint("doc_term_matrix: {0}".format(doc_term_matrix))
logprint("id2term: {0}".format(id2term))
"""
def textacyTopicModeling(corpus,
n_topics = 15, top_topic_words = 7, top_document_labels_per_topic = 5,
ngrams = 1, min_df=1, max_df=1.0,
topicModel='lda'):
n_terms = int(n_topics * top_topic_words)
sort_terms_by = 'seriation' # 'seriation', 'weight', 'index', 'alphabetical'
rank_terms_by = 'corpus' # 'corpus', 'topic'
logprint(
"############### Topic Modeling {0} ###########################".format(
topicModel))
logprint(str("ngrams: {0}".format(ngrams)))
logprint(str("min_df: {0}".format(min_df)))
logprint(str("max_df: {0}".format(max_df)))
logprint(str("n_topics: {0}".format(n_topics)))
logprint("\n")
start = time.time()
# http://textacy.readthedocs.io/en/latest/api_reference.html#textacy.tm.topic_model.TopicModel.get_doc_topic_matrix
weighting = ('tf' if topicModel == 'lda' else 'tfidf')
#################### vectorize corpi ####################
vectorizer = Vectorizer(weighting=weighting, min_df=min_df, max_df=max_df)
terms_list = (doc.to_terms_list(ngrams=ngrams, named_entities=False, as_strings=True) for doc in corpus)
doc_term_matrix = vectorizer.fit_transform(terms_list)
id2term = vectorizer.__getattribute__("id_to_term")
# printlog("terms_list: {0}".format(list(terms_list)))
# printlog("doc_term_matrix: {0}".format(doc_term_matrix))
##################### Initialize and train a topic model ##############################################
model = textacy.tm.TopicModel(topicModel, n_topics=n_topics)
model.fit(doc_term_matrix)
doc_topic_matrix = model.transform(doc_term_matrix)
for topic_idx, top_terms in model.top_topic_terms(vectorizer.id_to_term, top_n=top_topic_words):
logprint('topic {0}: {1}'.format(topic_idx, " ".join(top_terms)))
for topic_idx, top_docs in model.top_topic_docs(doc_topic_matrix, top_n=top_document_labels_per_topic):
logprint(topic_idx)
for j in top_docs:
logprint(corpus[j].metadata['categoryName'])
####################### termite plot ###################################################################
grams_label = "uni" if ngrams == 1 else "bi"
model.termite_plot(doc_term_matrix, id2term,
n_terms=n_terms,
sort_terms_by=sort_terms_by,
rank_terms_by=rank_terms_by+'_weight',
save= FILEPATH + "results/{}_{}_{}_{}_{}_{}.png".format(grams_label,topicModel,n_topics,n_terms,sort_terms_by,rank_terms_by))
end = time.time()
logprint("\n\n\nTime Elapsed Topic Modeling with {1}:{0} min\n\n".format((end - start) / 60, topicModel))
def jgibbsLLDA(corpus, path2save_results, top_topic_words=7):
start = time.time()
jgibbsLLDA_root = FILEPATH + "java_LabledLDA/"
LLDA_filepath = "{0}models/tickets/tickets.gz".format(jgibbsLLDA_root)
# build dictionary of ticketcategories
labelist = []
for texdoc in corpus.get(lambda texdoc: texdoc.metadata["categoryName"] not in labelist):
labelist.append(texdoc.metadata["categoryName"])
labeldict = {k: v for v, k in enumerate(labelist)}
reverse_labeldict = {v: k for k, v in labeldict.items()}
#and save
labeldict_path = FILEPATH + "results/labeldict.txt"
with open(labeldict_path, 'w') as file:
file.write(json.dumps(labeldict))
n_topics = len(labeldict) #+1 #default-topic
# create file with label_IDs (input for llda)
textacy.fileio.write_file_lines(generate_lablelID_lines(corpus, labeldict), filepath=LLDA_filepath)
# wait for file to exist
while not os.path.exists(LLDA_filepath):
time.sleep(1)
logprint("")
logprint("start LLDA:")
# run JGibbsLLDA file
FNULL = open(os.devnull, 'w') # supress output
cmd_jgibbs_java = ["java", "-cp",
"{0}lib/trove-3.0.3.jar:{0}lib/args4j-2.0.6.jar:{0}out/production/LabledLDA/".format(
jgibbsLLDA_root),
"jgibblda.LDA", "-est", "-dir", "{0}models/tickets".format(jgibbsLLDA_root), "-dfile",
"tickets.gz",
"-twords", str(top_topic_words), "-ntopics", str(n_topics)]
subprocess.call(cmd_jgibbs_java, stdout=FNULL)
# ANMERKUNG: Dateien sind versteckt. zu finden in models/
cmd_gzip = ["gzip", "-dc", "{0}/models/tickets/.twords.gz".format(jgibbsLLDA_root)]
output = subprocess.check_output(cmd_gzip).decode("utf-8")
topic_regex = re.compile(r'Topic [0-9]*')
#####################################
# todo save results in file aufgrund von results
result = []
for line in output.splitlines():
findall = topic_regex.findall(line)
if len(findall) != 0:
try:
index = int(findall[0].split()[1])
result.append("Topic {} {}:".format(index, reverse_labeldict[index]))
except:
result.append(line)
else:
result.append(line)
textacy.fileio.write_file_lines(result, path2save_results+".txt")
#####################################
results = []
res_dict = {}
count =0
for line in output.splitlines():
findall = topic_regex.findall(line)
if len(findall) != 0:
if len(res_dict) != 0:
results.append(res_dict) #vorheriges an die liste ran (ist ja dann fertig)
index = int(findall[0].split()[1])
res_dict = {index : str(reverse_labeldict[index]) }
else:
splitted = line.split()
res_dict[splitted[0]] = float(splitted[1])
### print terms that are topics
for s in list(res_dict.values()):
if isinstance(s,str) and splitted[0] in s:
vals = list(res_dict.values())
keys = list(res_dict.keys())
for v in vals:
if not isinstance(v,float):
print("{}".format(v))
print("{}".format(splitted[0]))
count +=1
print()
###
if len(res_dict) != 0:
results.append(res_dict) # letzes an die liste ran
print(count)
print(float(count)/float(len(labelist)))
# {0: 'betrieb', 'service': 0.24162679425837305, 'support': 0.24162679425837305, 'browser': 0.24162679425837305, 'unicard': 0.24162679425837305, 'telefon': 0.0023923444976076593}
# every term in the resulsts to a list
terms=[]
for res in results:
for key,value in res.items():
if not isinstance(key, int) and not key in terms:
terms.append(key)
term2id = {t:i for i,t in enumerate(terms)} #and to dict
################# termite plot #####################################################################
#term_topic_weights.shape = (len(term_ids),len(topic_ids)
#topic_labels = tuple(labelist)
topic_labels = list(range(len(labelist)))
term_labels = list(range(len(term2id))) #tuple([key for key in term2id.keys()])
term_topic_weights = np.zeros((len(term2id),len(topic_labels)))
for i,res in enumerate(results):
for key,value in res.items():
if not isinstance(key, int):
term_topic_weights[term2id[key]][i] = value
term_labels[term2id[key]] = key
else:
topic_labels[i] = reverse_labeldict[key]
viz.draw_termite_plot(
term_topic_weights, topic_labels, term_labels, save=path2save_results+".png")
end = time.time()
logprint("Time Elapsed Topic Modeling JGibbsLLDA:{0} min\n".format((end - start) / 60))
def main(use_cleaned=False, algorithm="llda"):
# idee http://bigartm.org/
# idee http://wiki.languagetool.org/tips-and-tricks
# idee https://en.wikipedia.org/wiki/Noisy_text_analytics
# idee https://gate.ac.uk/family/
logprint("Topic Modeling: {0}".format(datetime.now()))
corpus_de_path = FILEPATH + config.get("de_corpus", "path")
corpus_en_path = FILEPATH + config.get("en_corpus", "path")
if use_cleaned:
preCorpus_name = "de" + "_clean_ticket"
resultspath = FILEPATH + "results/clean"
else:
preCorpus_name = "de" + "_pre_ticket"
resultspath = FILEPATH + "results/pre"
# load cleand corpus
de_corpus, parser = load_corpus(corpus_name=preCorpus_name, corpus_path=corpus_de_path)
logprint("Corpus loaded: {0}".format(de_corpus.lang))
# todo llda topics zusammenfassen
# idee lda so trainieren, dass zuordnung term <-> topic nicht zu schwach wird, aber möglichst viele topics
# frage wieviele tickets pro topic?
"""
ngrams = 1
min_df = 1
max_df = 1.0
weighting = 'tf'
# weighting ='tfidf'
named_entities = False
printvecotorization(ngrams=1, min_df=1, max_df=1.0, weighting=weighting)
printvecotorization(ngrams=1, min_df=1, max_df=0.5, weighting=weighting)
printvecotorization(ngrams=1, min_df=1, max_df=0.8, weighting=weighting)
printvecotorization(ngrams=(1, 2), min_df=1, max_df=1.0, weighting=weighting)
printvecotorization(ngrams=(1, 2), min_df=1, max_df=0.5, weighting=weighting)
printvecotorization(ngrams=(1, 2), min_df=1, max_df=0.8, weighting=weighting)
"""
if algorithm == "llda":
top_topic_words = 5
path2save_results = resultspath + "_{}_{}".format(algorithm,"top"+str(top_topic_words))
jgibbsLLDA(de_corpus, path2save_results=path2save_results, top_topic_words=top_topic_words)
"""
top_topic_words = 10
path2save_results = resultspath + "_{}_{}".format(algorithm,"top"+str(top_topic_words))
jgibbsLLDA(de_corpus, path2save_results=path2save_results, top_topic_words=top_topic_words)
top_topic_words = 15
path2save_results = resultspath + "_{}_{}".format(algorithm, "top" + str(top_topic_words))
jgibbsLLDA(de_corpus, path2save_results=path2save_results, top_topic_words=top_topic_words)
top_topic_words = 20
path2save_results = resultspath + "_{}_{}".format(algorithm, "top" + str(top_topic_words))
jgibbsLLDA(de_corpus, path2save_results=path2save_results, top_topic_words=top_topic_words)
"""
else:
textacyTopicModeling(ngrams = 1,
min_df = 1,
max_df = 0.9,
topicModel = algorithm,
n_topics =15,
corpus=de_corpus)
"""
textacyTopicModeling(ngrams=1,
min_df=1,
max_df=0.9,
topicModel=algorithm,
n_topics=20,
corpus=de_corpus)
textacyTopicModeling(ngrams=1,
min_df=1,
max_df=0.9,
topicModel=algorithm,
n_topics=25,
corpus=de_corpus)
textacyTopicModeling(ngrams=1,
min_df=1,
max_df=0.9,
topicModel=algorithm,
n_topics=30,
corpus=de_corpus)
"""
textacyTopicModeling(ngrams=(1, 2),
min_df=1,
max_df=0.9,
topicModel=algorithm,
n_topics=15,
corpus=de_corpus)
"""
textacyTopicModeling(ngrams = (1,2),
min_df = 1,
max_df = 0.9,
topicModel = algorithm,
n_topics =20,
corpus=de_corpus)
textacyTopicModeling(ngrams = (1,2),
min_df = 1,
max_df = 0.9,
topicModel = algorithm,
n_topics =25,
corpus=de_corpus)
textacyTopicModeling(ngrams = (1,2),
min_df = 1,
max_df = 0.9,
topicModel = algorithm,
n_topics =30,
corpus=de_corpus)
"""
if __name__ == "__main__":
main()