Address: [go: up one dir, main page]

Include Form Remove Scripts Session Cookies

Academia.edu no longer supports Internet Explorer.

To browse Academia.edu and the wider internet faster and more securely, please take a few seconds to upgrade your browser.

Log In
Sign Up

Figure from:Topic chains for understanding a news corpus

See full PDF downloadDownload figure

Fig. 4. Detecting focus shifts using difference of a word probability along the topic
chain. Each rectangle represents a topic node, and contains top probability words.
Edges connect two similar topics within a sliding window of size six, and the words next
to the edge are the named entities whose probabilities are changed the most between
two topics. xxM yyP zzT represent month, period, and topic number, respectively. — Figure 4 Detecting focus shifts using difference of a word probability along the topic chain. Each rectangle represents a topic node, and contains top probability words. Edges connect two similar topics within a sliding window of size six, and the words next to the edge are the named entities whose probabilities are changed the most between two topics. xxM yyP zzT represent month, period, and topic number, respectively.

Related Figures (4)

Table 1. Four topics discovered by LDA for the news dataset. Topics are randomly
chosen and are represented by top ten probability words. Topic 1 is about “soccer
game”, topic 2 is about “market” and “business”, topic 3 is about “smart phones”,
and topic 4 is about “research”. Each topic is a multinomial distribution over words.
Dirichlet with parameter @. From the generative process, we obtain the likelihood
of a document: — Table 1. Four topics discovered by LDA for the news dataset. Topics are randomly chosen and are represented by top ten probability words. Topic 1 is about “soccer game”, topic 2 is about “market” and “business”, topic 3 is about “smart phones”, and topic 4 is about “research”. Each topic is a multinomial distribution over words. Dirichlet with parameter @. From the generative process, we obtain the likelihood of a document:

Fig. 1. Comparison of negative log likelihood for six similarity metrics using a boxplot.
Negative log likelihood was computed for the corpus using the set of topics where
five topics were substituted with the five most similar topics from another time slice,
identified by each of the six similarity metrics. A better similarity metric gives a lower
negative log likelihood. JS divergence and Jaccard’s coefficient with 0.5 cumulative
probability mass achieve better performances than the other metrics. An asterisk (*)
next to a metric indicates statistically significant differences between the metric and
JS divergence using t-test, p < 0.01. — Fig. 1. Comparison of negative log likelihood for six similarity metrics using a boxplot. Negative log likelihood was computed for the corpus using the set of topics where five topics were substituted with the five most similar topics from another time slice, identified by each of the six similarity metrics. A better similarity metric gives a lower negative log likelihood. JS divergence and Jaccard’s coefficient with 0.5 cumulative probability mass achieve better performances than the other metrics. An asterisk (*) next to a metric indicates statistically significant differences between the metric and JS divergence using t-test, p < 0.01.

Fig. 2. Number of topic chains with different similarity cuts using JS divergence. The
number of topic chains is significantly changed at JS divergence of 0.4. We chose JS
divergence of 0.4 to construct topic chains. — Fig. 2. Number of topic chains with different similarity cuts using JS divergence. The number of topic chains is significantly changed at JS divergence of 0.4. We chose JS divergence of 0.4 to construct topic chains.

‘ig. 3. Six sets of topic chains constructed with sliding windows of sizes one to six. Fo!
‘ach set of topic chains, every topic chain starts at the same vertical position. Withir
ach topic chain, topics are temporally ordered, the oldest (first) topic at the top anc
roing down to the most recent topic at the bottom. The number of nodes indicates
he total number of topics that are connected with one or more similar topics. Blu
1odes are those that belong to the largest topic chain in the last set of topic chains
‘onstructed with the sliding window of size six. Blue nodes start out in the first set
of topic chains as small topic chains, and they agglomerate as the size of sliding win.
low increases. The full-size figure is available at http://uilab.kaist.ac.kr/research/topic
thain/ — ‘ig. 3. Six sets of topic chains constructed with sliding windows of sizes one to six. Fo! ‘ach set of topic chains, every topic chain starts at the same vertical position. Withir ach topic chain, topics are temporally ordered, the oldest (first) topic at the top anc roing down to the most recent topic at the bottom. The number of nodes indicates he total number of topics that are connected with one or more similar topics. Blu 1odes are those that belong to the largest topic chain in the last set of topic chains ‘onstructed with the sliding window of size six. Blue nodes start out in the first set of topic chains as small topic chains, and they agglomerate as the size of sliding win. low increases. The full-size figure is available at http://uilab.kaist.ac.kr/research/topic thain/

About
Press
Papers
Topics
We're Hiring!
Help Center

Find new research papers in:
Physics
Chemistry
Biology
Health Sciences
Ecology
Earth Sciences
Cognitive Science
Mathematics
Computer Science

Terms
Privacy
Copyright
Academia ©2025