Experiments on the English Wikipedia¶
To test gensim performance, we run it against the English version of Wikipedia.
This page describes the process of obtaining and processing Wikipedia, so that anyone can reproduce the results. It is assumed you have gensim properly installed.
Preparing the corpus¶
First, download the dump of all Wikipedia articles from http://download.wikimedia.org/enwiki/ (you want the file enwiki-latest-pages-articles.xml.bz2, or enwiki-YYYYMMDD-pages-articles.xml.bz2 for date-specific dumps). This file is about 8GB in size and contains (a compressed version of) all articles from the English Wikipedia.
Convert the articles to plain text (process Wiki markup) and store the result as sparse TF-IDF vectors. In Python, this is easy to do on-the-fly and we don’t even need to uncompress the whole archive to disk. There is a script included in gensim that does just that, run:
$ python -m gensim.scripts.make_wiki
Note
This pre-processing step makes two passes over the 8.2GB compressed wiki dump (one to extract the dictionary, one to create and store the sparse vectors) and takes about 9 hours on my laptop, so you may want to go have a coffee or two.
Also, you will need about 35GB of free disk space to store the sparse output vectors. I recommend compressing these files immediately, e.g. with bzip2 (down to ~13GB). Gensim can work with compressed files directly, so this lets you save disk space.
Latent Semantic Analysis¶
First let’s load the corpus iterator and dictionary, created in the second step above
>>> import logging
>>> import gensim
>>>
>>> logging.basicConfig(format='%(asctime)s : %(levelname)s : %(message)s', level=logging.INFO)
>>>
>>> # load id->word mapping (the dictionary), one of the results of step 2 above
>>> id2word = gensim.corpora.Dictionary.load_from_text('wiki_en_wordids.txt')
>>> # load corpus iterator
>>> mm = gensim.corpora.MmCorpus('wiki_en_tfidf.mm')
>>> # mm = gensim.corpora.MmCorpus('wiki_en_tfidf.mm.bz2') # use this if you compressed the TFIDF output (recommended)
>>>
>>> print(mm)
MmCorpus(3931787 documents, 100000 features, 756379027 non-zero entries)
We see that our corpus contains 3.9M documents, 100K features (distinct tokens) and 0.76G non-zero entries in the sparse TF-IDF matrix. The Wikipedia corpus contains about 2.24 billion tokens in total.
Now we’re ready to compute LSA of the English Wikipedia:
>>> # extract 400 LSI topics; use the default one-pass algorithm
>>> lsi = gensim.models.lsimodel.LsiModel(corpus=mm, id2word=id2word, num_topics=400)
>>>
>>> # print the most contributing words (both positively and negatively) for each of the first ten topics
>>> lsi.print_topics(10)
topic #0(332.762): 0.425*"utc" + 0.299*"talk" + 0.293*"page" + 0.226*"article" + 0.224*"delete" + 0.216*"discussion" + 0.205*"deletion" + 0.198*"should" + 0.146*"debate" + 0.132*"be"
topic #1(201.852): 0.282*"link" + 0.209*"he" + 0.145*"com" + 0.139*"his" + -0.137*"page" + -0.118*"delete" + 0.114*"blacklist" + -0.108*"deletion" + -0.105*"discussion" + 0.100*"diff"
topic #2(191.991): -0.565*"link" + -0.241*"com" + -0.238*"blacklist" + -0.202*"diff" + -0.193*"additions" + -0.182*"users" + -0.158*"coibot" + -0.136*"user" + 0.133*"he" + -0.130*"resolves"
topic #3(141.284): -0.476*"image" + -0.255*"copyright" + -0.245*"fair" + -0.225*"use" + -0.173*"album" + -0.163*"cover" + -0.155*"resolution" + -0.141*"licensing" + 0.137*"he" + -0.121*"copies"
topic #4(130.909): 0.264*"population" + 0.246*"age" + 0.243*"median" + 0.213*"income" + 0.195*"census" + -0.189*"he" + 0.184*"households" + 0.175*"were" + 0.167*"females" + 0.166*"males"
topic #5(120.397): 0.304*"diff" + 0.278*"utc" + 0.213*"you" + -0.171*"additions" + 0.165*"talk" + -0.159*"image" + 0.159*"undo" + 0.155*"www" + -0.152*"page" + 0.148*"contribs"
topic #6(115.414): -0.362*"diff" + -0.203*"www" + 0.197*"you" + -0.180*"undo" + -0.180*"kategori" + 0.164*"users" + 0.157*"additions" + -0.150*"contribs" + -0.139*"he" + -0.136*"image"
topic #7(111.440): 0.429*"kategori" + 0.276*"categoria" + 0.251*"category" + 0.207*"kategorija" + 0.198*"kategorie" + -0.188*"diff" + 0.163*"категория" + 0.153*"categoría" + 0.139*"kategoria" + 0.133*"categorie"
topic #8(109.907): 0.385*"album" + 0.224*"song" + 0.209*"chart" + 0.204*"band" + 0.169*"released" + 0.151*"music" + 0.142*"diff" + 0.141*"vocals" + 0.138*"she" + 0.132*"guitar"
topic #9(102.599): -0.237*"league" + -0.214*"he" + -0.180*"season" + -0.174*"football" + -0.166*"team" + 0.159*"station" + -0.137*"played" + -0.131*"cup" + 0.131*"she" + -0.128*"utc"
Creating the LSI model of Wikipedia takes about 4 hours and 9 minutes on my laptop 1. That’s about 16,000 documents per minute, including all I/O.
Note
If you need your results even faster, see the tutorial on Distributed Computing. Note that the BLAS libraries inside gensim make use of multiple cores transparently, so the same data will be processed faster on a multicore machine “for free”, without any distributed setup.
We see that the total processing time is dominated by the preprocessing step of preparing the TF-IDF corpus from a raw Wikipedia XML dump, which took 9h. 2
The algorithm used in gensim only needs to see each input document once, so it is suitable for environments where the documents come as a non-repeatable stream, or where the cost of storing/iterating over the corpus multiple times is too high.
Latent Dirichlet Allocation¶
As with Latent Semantic Analysis above, first load the corpus iterator and dictionary
>>> import logging
>>> import gensim
>>>
>>> logging.basicConfig(format='%(asctime)s : %(levelname)s : %(message)s', level=logging.INFO)
>>>
>>> # load id->word mapping (the dictionary), one of the results of step 2 above
>>> id2word = gensim.corpora.Dictionary.load_from_text('wiki_en_wordids.txt')
>>> # load corpus iterator
>>> mm = gensim.corpora.MmCorpus('wiki_en_tfidf.mm')
>>> # mm = gensim.corpora.MmCorpus('wiki_en_tfidf.mm.bz2') # use this if you compressed the TFIDF output
>>>
>>> print(mm)
MmCorpus(3931787 documents, 100000 features, 756379027 non-zero entries)
We will run online LDA (see Hoffman et al. 3), which is an algorithm that takes a chunk of documents, updates the LDA model, takes another chunk, updates the model etc. Online LDA can be contrasted with batch LDA, which processes the whole corpus (one full pass), then updates the model, then another pass, another update… The difference is that given a reasonably stationary document stream (not much topic drift), the online updates over the smaller chunks (subcorpora) are pretty good in themselves, so that the model estimation converges faster. As a result, we will perhaps only need a single full pass over the corpus: if the corpus has 3 million articles, and we update once after every 10,000 articles, this means we will have done 300 updates in one pass, quite likely enough to have a very accurate topics estimate
>>> # extract 100 LDA topics, using 1 pass and updating once every 1 chunk (10,000 documents)
>>> lda = gensim.models.ldamodel.LdaModel(corpus=mm, id2word=id2word, num_topics=100, update_every=1, passes=1)
using serial LDA version on this node
running online LDA training, 100 topics, 1 passes over the supplied corpus of 3931787 documents, updating model once every 10000 documents
...
Unlike LSA, the topics coming from LDA are easier to interpret
>>> # print the most contributing words for 20 randomly selected topics
>>> lda.print_topics(20)
topic #0: 0.009*river + 0.008*lake + 0.006*island + 0.005*mountain + 0.004*area + 0.004*park + 0.004*antarctic + 0.004*south + 0.004*mountains + 0.004*dam
topic #1: 0.026*relay + 0.026*athletics + 0.025*metres + 0.023*freestyle + 0.022*hurdles + 0.020*ret + 0.017*divisão + 0.017*athletes + 0.016*bundesliga + 0.014*medals
topic #2: 0.002*were + 0.002*he + 0.002*court + 0.002*his + 0.002*had + 0.002*law + 0.002*government + 0.002*police + 0.002*patrolling + 0.002*their
topic #3: 0.040*courcelles + 0.035*centimeters + 0.023*mattythewhite + 0.021*wine + 0.019*stamps + 0.018*oko + 0.017*perennial + 0.014*stubs + 0.012*ovate + 0.011*greyish
topic #4: 0.039*al + 0.029*sysop + 0.019*iran + 0.015*pakistan + 0.014*ali + 0.013*arab + 0.010*islamic + 0.010*arabic + 0.010*saudi + 0.010*muhammad
topic #5: 0.020*copyrighted + 0.020*northamerica + 0.014*uncopyrighted + 0.007*rihanna + 0.005*cloudz + 0.005*knowles + 0.004*gaga + 0.004*zombie + 0.004*wigan + 0.003*maccabi
topic #6: 0.061*israel + 0.056*israeli + 0.030*sockpuppet + 0.025*jerusalem + 0.025*tel + 0.023*aviv + 0.022*palestinian + 0.019*ifk + 0.016*palestine + 0.014*hebrew
topic #7: 0.015*melbourne + 0.014*rovers + 0.013*vfl + 0.012*australian + 0.012*wanderers + 0.011*afl + 0.008*dinamo + 0.008*queensland + 0.008*tracklist + 0.008*brisbane
topic #8: 0.011*film + 0.007*her + 0.007*she + 0.004*he + 0.004*series + 0.004*his + 0.004*episode + 0.003*films + 0.003*television + 0.003*best
topic #9: 0.019*wrestling + 0.013*château + 0.013*ligue + 0.012*discus + 0.012*estonian + 0.009*uci + 0.008*hockeyarchives + 0.008*wwe + 0.008*estonia + 0.007*reign
topic #10: 0.078*edits + 0.059*notability + 0.035*archived + 0.025*clearer + 0.022*speedy + 0.021*deleted + 0.016*hook + 0.015*checkuser + 0.014*ron + 0.011*nominator
topic #11: 0.013*admins + 0.009*acid + 0.009*molniya + 0.009*chemical + 0.007*ch + 0.007*chemistry + 0.007*compound + 0.007*anemone + 0.006*mg + 0.006*reaction
topic #12: 0.018*india + 0.013*indian + 0.010*tamil + 0.009*singh + 0.008*film + 0.008*temple + 0.006*kumar + 0.006*hindi + 0.006*delhi + 0.005*bengal
topic #13: 0.047*bwebs + 0.024*malta + 0.020*hobart + 0.019*basa + 0.019*columella + 0.019*huon + 0.018*tasmania + 0.016*popups + 0.014*tasmanian + 0.014*modèle
topic #14: 0.014*jewish + 0.011*rabbi + 0.008*bgwhite + 0.008*lebanese + 0.007*lebanon + 0.006*homs + 0.005*beirut + 0.004*jews + 0.004*hebrew + 0.004*caligari
topic #15: 0.025*german + 0.020*der + 0.017*von + 0.015*und + 0.014*berlin + 0.012*germany + 0.012*die + 0.010*des + 0.008*kategorie + 0.007*cross
topic #16: 0.003*can + 0.003*system + 0.003*power + 0.003*are + 0.003*energy + 0.002*data + 0.002*be + 0.002*used + 0.002*or + 0.002*using
topic #17: 0.049*indonesia + 0.042*indonesian + 0.031*malaysia + 0.024*singapore + 0.022*greek + 0.021*jakarta + 0.016*greece + 0.015*dord + 0.014*athens + 0.011*malaysian
topic #18: 0.031*stakes + 0.029*webs + 0.018*futsal + 0.014*whitish + 0.013*hyun + 0.012*thoroughbred + 0.012*dnf + 0.012*jockey + 0.011*medalists + 0.011*racehorse
topic #19: 0.119*oblast + 0.034*uploaded + 0.034*uploads + 0.033*nordland + 0.025*selsoviet + 0.023*raion + 0.022*krai + 0.018*okrug + 0.015*hålogaland + 0.015*russiae + 0.020*manga + 0.017*dragon + 0.012*theme + 0.011*dvd + 0.011*super + 0.011*hunter + 0.009*ash + 0.009*dream + 0.009*angel
>>>
>>> import pickle # noqa: E402
>>>
>>> # Get an article and its topic distribution
>>> with open("wiki_en_bow.mm.metadata.cpickle", 'rb') as meta_file:
... docno2metadata = pickle.load(meta_file)
>>>
>>> doc_num = 0
>>> print("Title: {}".format(docno2metadata[doc_num][1])) # take the first article as an example
Title: Anarchism
>>>
>>> vec = mm[doc_num] # get tf-idf vector
>>> lda.get_document_topics(vec)
[(1, 0.028828567), (10, 0.32766217), (36, 0.021675354), (55, 0.2521854), (57, 0.27154338)]
Creating this LDA model of Wikipedia takes about 6 hours and 20 minutes on my laptop 1. If you need your results faster, consider running Distributed Latent Dirichlet Allocation on a cluster of computers.
Note two differences between the LDA and LSA runs: we asked LSA to extract 400 topics, LDA only 100 topics (so the difference in speed is in fact even greater). Secondly, the LSA implementation in gensim is truly online: if the nature of the input stream changes in time, LSA will re-orient itself to reflect these changes, in a reasonably small amount of updates. In contrast, LDA is not truly online (the name of the 3 article notwithstanding), as the impact of later updates on the model gradually diminishes. If there is topic drift in the input document stream, LDA will get confused and be increasingly slower at adjusting itself to the new state of affairs.
In short, be careful if using LDA to incrementally add new documents to the model over time. Batch usage of LDA, where the entire training corpus is either known beforehand or does not exhibit topic drift, is ok and not affected.
To run batch LDA (not online), train LdaModel with:
>>> # extract 100 LDA topics, using 20 full passes, no online updates
>>> lda = gensim.models.ldamodel.LdaModel(corpus=mm, id2word=id2word, num_topics=100, update_every=0, passes=20)
As usual, a trained model can used be to transform new, unseen documents (plain bag-of-words count vectors) into LDA topic distributions:
>>> doc_lda = lda[doc_bow]
- 1(1,2)
My laptop = MacBook Pro, Intel Core i7 2.3GHz, 16GB DDR3 RAM, OS X with libVec.
- 2
Here we’re mostly interested in performance, but it is interesting to look at the retrieved LSA concepts, too. I am no Wikipedia expert and don’t see into Wiki’s bowels, but Brian Mingus had this to say about the result:
There appears to be a lot of noise in your dataset. The first three topics in your list appear to be meta topics, concerning the administration and cleanup of Wikipedia. These show up because you didn't exclude templates such as these, some of which are included in most articles for quality control: http://en.wikipedia.org/wiki/Wikipedia:Template_messages/Cleanup The fourth and fifth topics clearly shows the influence of bots that import massive databases of cities, countries, etc. and their statistics such as population, capita, etc. The sixth shows the influence of sports bots, and the seventh of music bots.
So the top ten concepts are apparently dominated by Wikipedia robots and expanded templates; this is a good reminder that LSA is a powerful tool for data analysis, but no silver bullet. As always, it’s garbage in, garbage out… By the way, improvements to the Wiki markup parsing code are welcome :-)
- 3(1,2)
Hoffman, Blei, Bach. 2010. Online learning for Latent Dirichlet Allocation [pdf] [code]