Description
* Overview
This homework is about working with latent Dirichlet allocation (LDA)
in order to infer token–topic (i.e., token–component)
assignments. The purpose of this homework is threefold:
* To understand and implement Gibbs sampling for LDA
* To appreciate the effects and implications of “label switching”.
* To understand and implement an approximate method for computing the
log predictive probability of new data according to LDA.
To complete this homework, you will need to use Python 2.7 [1], NumPy
[2], and SciPy [3]. Although all three are installed on EdLab, I
recommend you install them on your own computer. I also recommend you
install and use IPython [4] instead of the default Python shell.
Before answering the questions below, you should familiarize yourself
with the code in hw7.py. In particular, you should try running
sample(). For example, running ‘sample(array([5, 2, 3]), 10)’
generates 10 samples from the specified distribution as follows:
>>> sample(array([5, 2, 3]), 10)
array([0, 0, 2, 2, 1, 0, 0, 0, 2, 1])
* Questions
** Question 1
Class Corpus (in corpus.py) implements a corpus of (ungrouped)
documents. This class supports slicing as follows:
>>> corpus = Corpus()
>>> corpus.add(‘doc 1’, [‘foo’, ‘bar’])
>>> corpus.add(‘doc 2’, [‘bar’, ‘foo’])
>>> corpus.add(‘doc 3’, [‘bar’, ‘baz’, ‘baz’])
>>> print len(corpus)
3
>>> print len(corpus.vocab)
3
>>> print len(corpus[:2])
2
>>> print len(corpus[:2].vocab)
3
>>> for doc in corpus[:2]:
… print doc.w
…
[0, 1]
[1, 0]
>>> print corpus[2].w
[1, 2, 2]
(Note: although you do not need to understand how this class is
implemented, you may find it helpful to do so.)
CSV file questions.csv contains 5,264 questions about science from the
Madsci Network question-and-answer website [?]. Each question is
represented by a unique ID, and the text of the question itself.
Function preprocess() (in preprocess.py) takes a CSV file, an optional
stopword file, and an optional list of extra stopwords as input and
returns a corpus (i.e., an instance of the Corpus class, as described
above), with any stopwords removed. You can run this code as follows:
>>> corpus = preprocess(‘questions.csv’, ‘stopwordlist.txt’, [‘answer’, ‘dont’, ‘find’, ‘im’, ‘information’, ‘ive’, ‘message’, ‘question’, ‘read’, ‘science’, ‘wondering’])
>>> print ‘V = %s’ % len(corpus.vocab)
V = 21919
The resultant corpus may be split into a corpus of “training”
documents and a corpus of “testing” documents as follows:
>>> train_corpus = corpus[:-100]
>>> print len(train_corpus)
5164
>>> test_corpus = corpus[-100:]
>>> print len(test_corpus)
100
These corpora have the same vocabulary:
>>> train_corpus.vocab == test_corpus.vocab
True
Class LDA (in hw7.py) implements latent Dirichlet allocation.
a) Implement code for performing a single Gibbs sampling iteration,
i.e., drawing a single set of token–topic assignments z_1 through
z_N from P(z_1, …, z_N | w_1, …, w_N, H), by filling in the
missing code in gibbs_iteration(). Hint: there is no need to work
in log space, i.e., you should use sample() rather than
log_sample(). You can run your code as follows:
>>> extra_stopwords = [‘answer’, ‘dont’, ‘find’, ‘im’, ‘information’, ‘ive’, ‘message’,’question’, ‘read’, ‘science’, ‘wondering’]
>>> corpus = preprocess(‘questions.csv’, ‘stopwordlist.txt’, extra_stopwords)
>>> train_corpus = corpus[:-100]
>>> assert train_corpus.vocab == corpus.vocab
>>> test_corpus = corpus[-100:]
>>> assert test_corpus.vocab == corpus.vocab
>>> V = len(corpus.vocab)
>>> T = 100
>>> alpha = 0.1 * T
>>> m = ones(T) / T
>>> beta = 0.01 * V
>>> n = ones(V) / V
>>> lda = LDA(train_corpus, alpha, m, beta, n)
>>> lda.gibbs(num_itns=250)
b) If the random number generator is initialized with a value of 1000
as follows, what is the log evidence for tokens w_1 through w_N
and token–topic (i.e., token–component) assignments z_1 through
z_N given assumptions H at the end of iteration 250?
>>> lda.gibbs(num_itns=250, random_seed=1000)
* Question 2
According to LDA, the mean of the posterior distribution over phi_t
(the categorical distribution over word types corresponding to topic
t) given the observed data w_1 through w_N and assumptions H, i.e.,
the mean of P(phi_t | w_1, …, w_N, H), is given by
E[phi_t] = int_{phi_t} dphi_t phi_t * P(phi_t | w_1, …, w_N, H),
where P(phi_t | w_1, …, w_N, H) = sum_{z_1, …, z_N} P(phi_t | w_1,
…, w_N, z_1, …, z_N, H) * P(z_1, …, z_N | w_1, …, w_N, H)
a) Can this value be computed analytically?
b) Function posterior_mean_phis() approximates the mean of the
posterior distribution over phi_t by the mean of the posterior
distribution over phi_t given a single set of token–topic (i.e.,
token–component) assignments z_1^{(s)} through z_N^{(s)} drawn
from P(z_1, …, z_N | w_1, …, w_N, H), i.e., the mean of
P(phi_t | w_1, …, w_N, z_1^{(s)}, …, z_N^{(s)}). Can you
obtain a better approximation by averaging P(phi_t | w_1, …,
w_N, z_1^{(s)}, …, z_N^{(s)}) over MULTIPLE samples from P(z_1,
…, z_N | w_1, …, w_N, H), obtained from DIFFERENT runs of your
code, or will you encounter the label switching problem?
* Question 3
According to LDA, the predictive probability of a single new token
w_{N + 1} of type v belonging to a NEW document (given observed tokens
w_1 through w_N and assumptions H) may be approximated by:
P(w_{N + 1} = v | w_1, …, w_N) = (1 / S) * sum_{s=1}^S sum_{t=1}^T
(N_{v|t}^{(s)} + beta * n_v) / (N_t^{(s)} + beta) * m_t
where m, beta, and n are hyperparameters, and superscript “{(s)}”
indicates that the superscripted quantity is that obtained using a
single set of token–topic assignments z_1^{(s)} through z_N^{(s)}
drawn from P(z_1, …, z_N | w_1, …, w_N, H).
a) Explain why this approximate predictive probability does not
involve any document-specific counts or alpha.
b) Explain why it’s possible to compute this approximate predictive
probability as stated above, i.e., by averaging over multiple
samples from P(z_1, …, z_N | w_1, …, w_N, H), without
encountering the label switching problem.
b) What is the corresponding approximate predictive probability of a
single new token w_{N + 1} of type v in an EXISTING document?
c) Is it possible to compute this approximate predictive probability
without encountering the label switching problem? Why?
d) What is the corresponding approximate predictive probability of a
single new token w_{N + 1} of type v for which z_{N + 1} = t?
e) Is it possible to compute this approximate predictive probability
without encountering the label switching problem? Why?
** Question 4
According to LDA, the predictive probability of new tokens w’_1
through w’_{N’} belonging to D’ new documents given observed tokens
w_1 through w_N and assumptions H can be approximated as follows:
P(w’_1, …, w’_{N’} | w_1, …, w_N, H) ~= (1 / S) * sum_{s=1}^S
P(w’_1, …, w’_{N’} | w_1, …, w_N, z_1^{(s)}, …, z_N^{(s)}, H)
where z_1^{(s)} through z_N^{(s)} comprise a single set of
token–topic (i.e., token–component) assignments drawn from P(z_1,
…, z_N | w_1, …, w_N, H). This approximation involves an
intractable sum over token–topic assignments z’_1 through z’_{N’},
thereby necessitating the following approximation:
P(w’_1, …, w’_{N’} | w_1, …, w_N, z_1^{(s)}, …, z_N^{(s)}, H) ~=
prod_{n=1}^{N’} (1 / R) sum_{r=1}^R sum_{t=1}^T P(w’_n, z’_n = t |
w’_1, …, w’_{n – 1}, z’_1^{(r)}, …, z’_{n – 1}^{(r)}, w_1, …,
w_N, z_1^{(s)}, …, z_N^{(s)}, H),
where z’_1^{(r)} through z’_{n – 1}^{(r)} comprise a single set of
token–topic (i.e., token–component) assignments drawn from P(z’_1,
…, z’_{n – 1} | w’_1, …, w’_{n – 1}, w_1, …, w_N, z_1^{(s)},
…, z_D^{(s)}, H). Lastly, if token w’_n is of type v, then
P(w’_n, z’_n = t | w’_1, …, w’_{n – 1}, z’_1^{(r)}, …, z’_{n –
1}^{(r)}, w_1, …, w_N, z_1^{(s)}, …, z_N^{(s)}, H) =
(({N’_{v|t}^{
>>> corpus = preprocess(‘questions.csv’, ‘stopwordlist.txt’, extra_stopwords)
>>> train_corpus = corpus[:-100]
>>> assert train_corpus.vocab == corpus.vocab
>>> test_corpus = corpus[-100:]
>>> assert test_corpus.vocab == corpus.vocab
>>> V = len(corpus.vocab)
>>> T = 10
>>> alpha = 0.1 * T
>>> m = ones(T) / T
>>> beta = 0.01 * V
>>> n = ones(V) / V
>>> lda = LDA(train_corpus, alpha, m, beta, n)
>>> lda.gibbs(num_itns=250, random_seed=1000)
>>> print lda.log_predictive_prob(test_corpus, num_samples=5)
When testing your code, may find it useful to save your model as
follows after performing Gibbs sampling:
>>> lda.gibbs(num_itns=250, random_seed=1000)
>>> lda.save(‘model.dat’)
You can then use your saved model as follows:
>>> lda = LDA.load(‘model.dat’)
>>> print lda.log_predictive_prob(test_corpus, num_samples=5)
b) What is the approximate log predictive probability, computed as
described above, i.e., using R = 5 random samples after
initializing the random number generator with a value of 1000 and
performing 25 Gibbs sampling iterations?
** Question 5
CSV file ufos.csv contains descriptions of 5,000 UFO sightings. Each
sighting is represented by a unique ID, the state in which the
sighting occurred, the shape of the sighted UFO, and a description.
a) Run your Gibbs sampling code as follows to produce lists of the
most probable word types according to the approximate posterior
distribution over phi_1 through phi_100 after 250 Gibbs sampling
iterations. Examine these lists and, well, enjoy 🙂
>>> extra_stopwords = [‘light’, ‘lights’, ‘object’, ‘objects’, ‘sky’]
>>> corpus = preprocess(‘ufos_small.csv’, ‘stopwordlist.txt’, extra_stopwords)
>>> V = len(corpus.vocab)
>>> T = 100
>>> alpha = 0.1 * T
>>> m = ones(T) / T
>>> beta = 0.01 * V
>>> n = ones(V) / V
>>> lda = LDA(corpus, alpha, m, beta, n)
>>> lda.gibbs(num_itns=250)
* References
[1] http://www.python.org/
[2] http://numpy.scipy.org/
[3] http://www.scipy.org/
[4] http://ipython.org/
[5] http://research.madsci.org/dataset/