To illustrate how to use pyLDAvis
's gensim helper funtions we will create a model from the 20 Newsgroup corpus. Minimal preprocessing is done and so the model is not the best. However, the goal of this notebook is to demonstrate the helper functions.
%%bash
mkdir -p data
pushd data
if [ -d "20news-bydate-train" ]
then
echo "The data has already been downloaded..."
else
wget http://qwone.com/%7Ejason/20Newsgroups/20news-bydate.tar.gz
tar xfv 20news-bydate.tar.gz
rm 20news-bydate.tar.gz
fi
echo "Lets take a look at the groups..."
!ls 20news-bydate-train/
popd
Each group dir has a set of files:
!ls -lah data/20news-bydate-train/sci.space | tail -n 5
Lets take a peak at one email:
!head data/20news-bydate-train/sci.space/61422
from glob import glob
import re
import string
import funcy as fp
from gensim import models
from gensim.corpora import Dictionary, MmCorpus
import nltk
import pandas as pd
# quick and dirty....
EMAIL_REGEX = re.compile(r"[a-z0-9\.\+_-]+@[a-z0-9\._-]+\.[a-z]*")
FILTER_REGEX = re.compile(r"[^a-z '#]")
TOKEN_MAPPINGS = [(EMAIL_REGEX, "#email"), (FILTER_REGEX, ' ')]
def tokenize_line(line):
res = line.lower()
for regexp, replacement in TOKEN_MAPPINGS:
res = regexp.sub(replacement, res)
return res.split()
def tokenize(lines, token_size_filter=2):
tokens = fp.mapcat(tokenize_line, lines)
return [t for t in tokens if len(t) > token_size_filter]
def load_doc(filename):
group, doc_id = filename.split('/')[-2:]
with open(filename, errors='ignore') as f:
doc = f.readlines()
return {'group': group,
'doc': doc,
'tokens': tokenize(doc),
'id': doc_id}
docs = pd.DataFrame(list(map(load_doc, glob('data/20news-bydate-train/*/*')))).set_index(['group','id'])
docs.head()
def nltk_stopwords():
return set(nltk.corpus.stopwords.words('english'))
def prep_corpus(docs, additional_stopwords=set(), no_below=5, no_above=0.5):
print('Building dictionary...')
dictionary = Dictionary(docs)
stopwords = nltk_stopwords().union(additional_stopwords)
stopword_ids = map(dictionary.token2id.get, stopwords)
dictionary.filter_tokens(stopword_ids)
dictionary.compactify()
dictionary.filter_extremes(no_below=no_below, no_above=no_above, keep_n=None)
dictionary.compactify()
print('Building corpus...')
corpus = [dictionary.doc2bow(doc) for doc in docs]
return dictionary, corpus
dictionary, corpus = prep_corpus(docs['tokens'])
MmCorpus.serialize('newsgroups.mm', corpus)
dictionary.save('newsgroups.dict')
%%time
lda = models.ldamodel.LdaModel(corpus=corpus, id2word=dictionary, num_topics=50, passes=10)
lda.save('newsgroups_50_lda.model')
Okay, the moment we have all been waiting for is finally here! You'll notice in the visualization that we have a few junk topics that would probably disappear after better preprocessing of the corpus. This is left as an exercises to the reader. :)
import pyLDAvis.gensim_models as gensimvis
import pyLDAvis
vis_data = gensimvis.prepare(lda, corpus, dictionary)
pyLDAvis.display(vis_data)
We can both visualize LDA models as well as gensim HDP models with pyLDAvis.
The difference between HDP and LDA is that HDP is a non-parametric method. Which means that we don't need to specify the number of topics. HDP will fit as many topics as it can and find the optimal number of topics by itself.
%%time
# The optional parameter T here indicates that HDP should find no more than 50 topics
# if there exists any.
hdp = models.hdpmodel.HdpModel(corpus, dictionary, T=50)
hdp.save('newsgroups_hdp.model')
As for the LDA model, in order to prepare the visualization you only need to pass it your model, the corpus, and the associated dictionary.
vis_data = gensimvis.prepare(hdp, corpus, dictionary)
pyLDAvis.display(vis_data)