- LDAvis is an interactive visualization tool built using R and D3 to help users interpret topics estimated using Latent Dirichlet Allocation (LDA). - It aims to answer questions about the meaning of each topic, the prevalence of each topic, and how topics relate to each other. - The tool visualizes term relevance, topic prevalence, and inter-topic distances to help users understand the topics in a corpus.

1. LDAvis:
A method for visualizing and interpreting topics
Carson Sievert Iowa State University
Kenneth E. Shirley AT&T Labs Research
Illvi2014

2. • LDAvis, a web-based interactive visualization of topics
estimated using Latent Dirichlet Allocation that is built using
a combination of R and D3.
• We introduce LDAvis that attempts to answer a few basic
questions about a fitted topic model：
I . I N T R O D U C T I O N

3. A. What is the meaning of each topic?,
B. How prevalent is each topic?,
C. How do the topics relate to each other?
I . I N T R O D U C T I O N

4. I I . R E L A T E D W O R K
LDA Turbo-
Topics
liftPMI

5. I I . R E L A T E D W O R K
LDA
Latent Dirichlet allocation (LDA)
is an example of a topic model and
was first presented as a graphical
model for topic discovery by
David Blei, Andrew Ng, and
Michael I. Jordan in 2003.

6. I I . R E L A T E D W O R K
Turbo-
Topics
Blei and Lafferty (2009) developed
“Turbo Topics”, a method of
identifying n-grams within LDA
inferred topics。
the resulting output is still simply a
ranked list containing a mixture of
terms and n-grams。

7. I I . R E L A T E D W O R K
Turbo-
Topics

8. I I . R E L A T E D W O R K
Pointwise Mutual Information(PMI) ：
PMI
Independence ： P(x , y) = P(x) × P(y)
PMI = log1 = 0

9. I I . R E L A T E D W O R K
lift
Lift：
Φ kw denote the probability of term
w ∈ {1, ..., V } for topic k ∈ {1, ..., K},
where V denotes the number of
terms in the vocabulary, and let pw
denote the marginal probability of
term w in the corpus.

10. I I . T O P I C M O D E L V I S U A L I Z A T I O N
• A number of visualization systems for topic models have been
developed in recent years.
• But, the visualization elements are limited to barcharts or
word clouds of term probabilities for each topic, pie charts of
topic probabilities for each document, and/or various
barcharts or scatterplots related to document metadata.

11. I I . T O P I C M O D E L V I S U A L I Z A T I O N
• Chuang et al. (2012b) develop such a tool, called “Termite”,
which visualizes the set of topic term distributions estimated
in LDA using a matrix layout.
• http://vis.stanford.edu/papers/termite

12. I I . T O P I C M O D E L V I S U A L I Z A T I O N
• 在代表主題內容的關鍵字選擇上，Termite定義了saliency 變數： saliency(w)
= P(w)× distinctiveness(w)。
• w 代表一個關鍵字； P(w)代表w的頻率；distinctiveness(w)代表w在主題間
的差異性，包含w的主題越多，distinctiveness(w)值越低。
• 在關鍵詞排列上，Termite 把經常連續出現的詞，如social和networks 排在
一起以便增強語義讓使用者較好理解。

13. I I I . R E L E V A N C E O F T E R M S T O T O P I C S
• Here we define relevance, our method for ranking terms
within topics, and we describe the results of a user study to
learn an optimal tuning parameter in the computation of
relevance.
• We define the relevance of term w to topic k given a weight
parameter λ (where 0 ≦ λ ≦ 1).

14. I I I . R E L E V A N C E O F T E R M S T O T O P I C S
• “λ” determines the weight given to the probability of term w
under topic k relative to its lift.
• Setting “λ” = 1 results in the familiar ranking of terms in
decreasing order of their topic-specific probability, and
setting “λ”= 0 ranks terms solely by their lift.

15. I I I . R E L E V A N C E O F T E R M S T O T O P I C S
• We fit a 50-topic model to the 20 Newsgroups data and each
term in the vocabulary (which has size V = 22, 524) for a
given topic.
• Figure shows this plot for Topic 29, which documents posted
to the “Motorcycles” Newsgroup, but also from documents
posted to the “Automobiles” Newsgroup and the “Electronics”
Newsgroup.

17. I I I . U S E R S T U D Y

13, 695 documents
20 Newsgroups

18. I I I . U S E R S T U D Y
• Some of the LDA-inferred topics occurred almost exclusively
(> 90% of occurrences) from a single Newsgroup, such as
Topic 38. which was came from the documents posted to the
“Medicine” (or “sci.med”) Newsgroup.
• Other topics occurred in a wide variety of Newsgroups. One
would expect these “spread-out” topics to be harder to
interpret than the “pure” topics like Topic 38.

19. I I I . U S E R S T U D Y
• we recruited 29 subjects among our colleagues (research
scientists at AT&T Labs with moderate familiarity with text
mining techniques and topic models).
• each subject completed an online experiment consisting of 50
tasks k (for k ∈ {1, ..., 50}).
• Task k was to read a list of five terms, ranked from 1-5 in
order of relevance to topic k, where “λ” ∈ (0, 1) was randomly
sampled to compute relevance.

20. I I I . U S E R S T U D Y
• we expected the proportion of correct responses to be roughly
1/3 no matter the value of λ used to compute relevance.
• In fact, seven of the topics were correctly identified by all 29
users, and one topic was incorrectly by all users.
• we estimated a topic specific intercept term to control the
difficulty of the topic (not just due to its tokens variety, but
also to account for the inherent familiarity of each topic to
our subject.)

21. I I I . U S E R S T U D Y
• The estimated effects of λ and λ² was statistically significant
(χ² p-value = 0.018).
• there was roughly a 67% baseline probability of correct
identification. As Figure 3 shows, for these topics, the
“optimal” value of λ was about 0.6.
• λ：0 ≒ 53 ％ and λ：1 ≒ 63 ％
• We view this as evidence that where λ < 1 can really improve
topic interpretability.

22. I I I . U S E R S T U D Y

23. I V . T H E L D A V I S S Y S T E M
Installation： pip install pyldavis

24. I V . T H E L D A V I S S Y S T E M
http://nbviewer.jupyter.org/github/bmabey/pyLDAvis/blob/
master/notebooks/pyLDAvis_overview.ipynb

25. I V . T H E L D A V I S S Y S T E M
http://nbviewer.jupyter.org/github/effytseng/dh_ldavis/blob/ma
ster/LDA-practice-1.html#topic=0&lambda=1&term=
http://nbviewer.jupyter.org/github/effytseng/dh_ldavis/blob/ma
ster/python-lldavis_test1-24-2.html
http://nbviewer.jupyter.org/github/bmabey/pyLDAvis/blob/mast
er/notebooks/Gensim%20Newsgroup.ipynb
．用 gensim 處理文本：
．自由中國 ldavis 圖表：
．自由中國 l-ldavis 圖表：

26. I V . T H E L D A V I S S Y S T E M
• the areas of the circles are proportional to the relative
prevalences of the topics in the corpus
• The default for computing inter-topic distances is Jensen-
Shannon divergence.
• The default for scaling the set of inter-topic distances defaults
to Principal Components.

27. I V . T H E L D A V I S S Y S T E M

28. I V . T H E L D A V I S S Y S T E M

29. I V . T H E L D A V I S S Y S T E M

30. V . D I S C U S S I O N
A. What is the meaning of each topic?,
B. How prevalent is each topic?,
C. How do the topics relate to each other?

31. V . D I S C U S S I O N
• For future work, we anticipate performing a larger user study to
further understand how to facilitate topic interpretation in fitted
LDA models
• In addition to relevance. The need to visualize correlations
between topics can provide insight into what is happening on the
document level without actually displaying entire documents.
• Last, we seek a solution to the problem of visualizing a large
number of topics (say, from 100 - 500 topics) in a compact way.