International Conference on Advances in Electrical, Electronics, Information, Communication and Bio-Informatics (AEEICB16)
Trending Topic Analysis Using Novel Sub
Topic Detection Model
Halima Banu S and S Chitrakala
Department of Computer Science and Engineering, Anna University, Chennai, Tamil Nadu, India.
(e-mail: halimabanu91@gmail.com , au.chitras@gmail.com).
Abstract—Twitter collects millions of tweets every day which
in turn serves as a rich information delivering platform.
However, some users, especially new users, often find it difficult
to understand trending topics in Twitter when confronted with
overwhelming and unorganized tweets. Previously, there has
been attempts to provide a short snippet to summarize a topic
but, this does not scale up to user’s expectation as it does not
provide any analyzed summary. This work aims to develop a
Trending Topic Analysis System to analyze trending topics by
performing Topic based sentiment classification thereby
summarizing public views on selected trending topics and to
generate extractive sub summaries of topics over the time period
using a novel sub topic detection approach. Different from the
traditional summarization framework, conflicting summary
generation could be avoided with sentiment classification
enhanced by common and tweet specific feature extraction
thereby sorting the data into separate sentiment corpus. Volume-
based followed by topic modelled approach of detecting sub topic
in the corpus help detect subtopics under the trending topic more
efficiently. A new approach called Foreground Dynamic Topic
Modelling (VF-DTM) is proposed which distills the noisy content
and extracts the foreground tweets from the corpus and then
build the intended model on it. Efficiency of Sub topic detection
in turn increases the quality of the extractive sub summaries
generated.
Index Terms—Trending Topic Analysis, Novel Sub-Topic
Detection, Foreground Dynamic Topic Modelling, VF-DTM
I. INTRODUCTION
Over the recent years Twitter has grown from a vague
invention to become a mainstream medium for dissemination
of messages and the public discussion of news and events. The
rapid proliferation of Twitter posts presents a big obstacle for
efficient information acquisition. It is impossible for a user to
get an overview of important topics on Twitter by reading all
tweets every day. In addition, because of information
redundancy and the informal writing style, it is time
consuming to find useful information about a topic from a
huge number of tweets. The tremendous volume of tweets
raises a needful framework called summarization as the key to
facilitating the requirements of topic exploration, navigation,
and search from hundreds of thousands of tweets. Specifically,
a summary that provides representative information of topics
with no redundancy and overlapped content is desired.
978-1-4673-9745-2 ©2016 IEEE
Analysis of topics is truly strengthened by performing a
sentiment classification but summarization applications lack
this component and as a result produce conflicting summaries.
Topics discussed in Twitter are very diverse and
unpredictable. Sentiment classifiers always dedicate
themselves to a specific domain or topic. Namely, a classifier
trained on sentiment data from one topic often performs
poorly on test data from another. Most of the recent
applications for sentiment analysis deal with a single topic
under interest. Another concern regarding sentiment
classification is if carried out as a supervised approach it will
incur heavy manual effort in annotation.Contributions of this
paper focus in addressing the following issues: (1) Avoidance
of generation of conflicting summaries (2) Topic Evolution
with efficient sub topic detection model to uncover hidden sub
topics.
Rest of the paper is organized as follows: In Section 2 we
discuss the literature survey pertaining to Sentiment
Classification and Sub Topic Detection for Twitter
Summarization. System architecture and details about the
proposed model in described in Section 3. Experimental
Results are discussed in Section 4 followed by Conclusion and
Future works in Section 5.
II. LITERATURE SURVEY
A. Sentiment Classification
One of the major issues of Sentiment classification using
machine learning approaches is the need for labeled training
data. Twitter dataset lack labeled data which often poses a
huge effort in manual annotation.
H.Rui et all [1] used supervised algorithms namely Naïve
Bayes Classifier and Support Vector Machine Classifier for
sentiment classification to investigate if Twitter WOM affects
movie sales and if so the approach involved in it.
Contributions included measuring the impacts of WOM from
people with different degrees of social connectivity exhibited
in a social network. But the system incurred heavy labor work
in labeling the huge training data set.
Furthermore twitter data are very diverse; therefore to train
a universal classifier was near impossible.
International Conference on Advances in Electrical, Electronics, Information, Communication and Bio-Informatics (AEEICB16)
Some works have been done for cross domain sentiment
analysis for review datasets like in [7], S.J.Pan et all used
Spectral FeatureAlignment algorithm (SFA) for classifying
sentiment polarity, or bridging the gap between the domains
and to aligndomain-specific words from different domains into
unified clusters. System significantly outperformed previous
approaches to cross-domain sentiment classification but could
not be applied for twitter dataset.
Xiaojun Wan [2] used Co training algorithm for training
and SVM Classifier was used for sentiment analysis. Major
focus was on cross lingual sentiment classification which
leveraged English corpus for Chinese sentiment classification.
But the approach could not suite well for the nature of tweets.
S.Lui et all [3] performed semi supervised multi class
model for topic adaptive sentiment classification. The
classifier was initially built using common features which did
not adapt well for cross domains. Semi supervised learning
was used to adaptively learn the build a topic adaptive
classifier. But the system needed topic labeled data in order to
apply the respective classifier model. Obtaining labelled data
poses a major constraint which in turn reduces the adaptability
of this system to any kind of data to be analyzed.
B. Sub Topic Detection for Twitter Summarization
The popularity of micro-blogging services, such as Twitter
has caught increasing attention from worldwide researchers.
There exist some pioneering researches working on Twitter
summarization.
D.Wen et all [4] accomplished Summarization using a non-
parametric Bayesian model applied to Hidden Markov
Models. A novel observation model was designed to allow
ranking based on selected predictive characteristics of
individual tweets. Major focus was to investigate the
possibility of using a temporal probabilistic data model known
as Hierarchical Dirichlet Process Hidden Markov Model
(HDP-HMM) to process a stream of tweets pertaining to a
single subject and cluster the tweets into groups or rankings
based on the value of the individual tweets .But Summaries
generated did not project the temporal nature.
F.Lui et all [5] proposed a Concept Based optimization
framework for topic Summarization using Integer Linear
Programming. Target data comprised of original and
normalized tweets along with web content relevant to the
topic. The focus was not on developing new summarization
systems but rather utilizing and integrating diverse text
sources to generate more informative summaries but there
were lack of series of sub events identification to show topic
evolving process.
B.O’Conor et all [6] explained Twitter topics by presenting
a simple list of messages. An exploratory search application
for Twitter called TweetMotif grouped messages by frequent
significant terms and retrieved results were further narrowed
down through a search interface. But the system lacked
temporal aspect in summaries generated and topic
development was not observed since sub topic detection
process was not employed. D.Gao et all [8] generated
sequential summaries on a topic using Stream and Semantic
based approaches for sub topic detection. But the approach fell
short of readability of summaries generated. System failed to
978-1-4673-9745-2 ©2016 IEEE
capture the opinion expressed in the data thus leading to
conflicting summaries. Also, sub topic detection mechanism
needs more enhancement.
III. SYSTEM ARCHITECTURE
Trending Topic Analyzer is implemented through topic
adaptive sentiment classification and multi tweet extractive
summarization. This work aims at generating chronologically
ordered sub summaries which throws light in understanding
the topic evolution of the trending topics. Topic under study is
assumed to contained several hidden sub topics which are
revealed using the proposed novel model called Foreground
Dynamic Topic Modelling (VF-DTM). This enables an end
user to completely analyze a trending topic to a greater level
of detail.
A. Target Data Preparatory Steps
Target data collection is performed by extracting the
trending topic on a region basis. Only those topics which
involves a topic development of user focus change will be
considered for summarization in this work. Based on the
selected topics tweets are extracted from the Twitter and
stored in database. Necessary preprocessing such as URL
removal, Slang Word Replacement, Non English Tweet Filter,
Stemmer and Stop Word Removal is performed to prepare the
target useable data. Along with the preprocessing phase, the
proposed system also handles other language tweet translation
which is essential to prevent discarding of public opinion
about the trending topic.
B. Topic Based Sentiment Classification
Our sentiment classification is inspired by the method
proposed by Shenghua Liu et al. in [3] for labelling the corpus
with their corresponding sentiments. Topic based sentiment
classification is divided into two phases namely Feature
Extraction and Model Building.
Feature extraction involved in this system creates a
significant impact in the performance of sentiment
classification task. Tweets have a special nature unlike normal
English sentences like @ symbol used to refer to a user,
emoticons expressed in the tweets etc. Incorporating such
features while performing feature extraction enhances the
overall process. Features extracted could be classified into two
broad categories namely Common Features and Tweet
Specific Features. With the extracted features classifier model
is built based on each topic without exhaustively labelling the
training set. Support Vector Machine is used as the classifier
whose appropriateness to this work will be examined in later
sections. Common features are a combination of common
sentiment words and topic sentiment words which will be used
to train the classifier.With POS tagging for tweets on a topic
and removing the common sentiment words, proposed
framework selects the frequent adjectives (aj), verbs (v),
nouns (n) and adverbs (av) as candidates of topic-adaptive.
Figure 1 below shows the system architecture depicting the
proposed system model. Following Algorithm 1 describes the
steps involved in tagging the tweets.This forms the topic
adaptive words. Common adaptive word set is built using
generic positive and negative oriented words.
International Conference on Advances in Electrical, Electronics, Information, Communication and Bio-Informatics (AEEICB16)

end for
Fig.1 System Architecture.
for every user in addUserList do
extractSentiment(user)
Algorithm 1: Common Feature Extraction extractPostTime(user.tweet)
Input : Pre-processed tweets end for
Output : Feature Vector
for each pre-processed tweet in topic e do Based on the post time of each user , feature vector will be
words [] =POSTagger (tweet) built based on the sentiment of the user‘s tweet and its post
frequencyCount (words []) time. Emoticons present in the tweet text serve as a simple
targetTagList (v, n, av, aj) way to predict the sentiment of the tweet. Therefore using a
for each word in targetTagList emoticon dictionary as s basis, emoticon feature value will be
if (word not in CommonSentimentWordList) computed.
addTopicAdaptiveWordList(word) Once the initial features are extracted , classifier model is
end if built using a collaborative training approach where the
end for unlabelled data and the features extracted will be iteratively
end for augmented with the initial labelled set.

Tweet specific features include the @-network features,
user sentiment features and emoticons. These features captures
the special nature of tweets as features. In twitter, @ symbol
in a tweet is used to refer to a user whom a person would
tweet to e.g. @abc means that the tweet is referring the user
abc. Using this symbol useful feature can be extracted there by
eliminating the need to label the entire dataset.Algorithm 2
details about the tweet specific feature extraction where child
edges (CEdge) and parent edges (PEdges) will be created for a
directed graph model according to the respective criteria.
Algorithm 2: Tweet Specific Feature Extraction
Input : Pre-processed tweets
Output : Feature Vector
for each tweet in topic e in training set L do
initialize (vertices-tweet)
initialize (d-edges- @-relations)
end for
for every vertexi in vertices do
for every vertexj in vertices do
if(postTime(vertexi)<postTime(vertexj))
createCEdge(vertex i,vertexj)
else
createPEdge(vertex i,vertexj)
end if
end for
end for
for each tweet in topic e in training set L do
if(user.tweetCount>2)
C. Tweet Summarization Framework
Topic evolution is achieved using a two-step process in the
summarization framework as used in [8] namely Sub Topic
Detection and Sub Summary Candidate Selection. Tweets are
prone to contain noisy data due to various reasons such as
limited length of a tweet text, informal language used by the
user tweeting etc.
Sub Topic Detection process in the proposed framework
involves a Volume based approach followed by the novel
Topic Modelling approach called the Foreground-Dynamic
Topic Modelling. Tweet stream that is collected may contain
many irrelevant data which may tamper the quality of the
extractive summary generated.
This raises a need to clean and extract only the relevant
corpus. Noisy and irrelevant tweets given a trending topic is
handled by a bi-level process of executing Volume based
approach and; the output from the previous step i.e. the tweet
set collected will be fed into the novel topic modelling
approach.
Sub topics are captured using eq. (3.10) where the word
distribution evolves over the time period.
βt,k | βt−1,k ~ N(βt−1,k ,σ 2 I ) (1)
Tweet distribution is obtained using eq. (3.11) where the
tweet at time interval at time interval t depends on t-1.
a t | α t −1 ~ N (α t −1 , δ 2 I ) (2)

addUserList(user)
end if
978-1-4673-9745-2 ©2016 IEEE
International Conference on Advances in Electrical, Electronics, Information, Communication and Bio-Informatics (AEEICB16)

VI. EXPERIMENTAL RESULTS

Twitter API allows to track and download the trending
topics given the WOEID (Where On Earth Identifier) of a
region. Based on the selection criteria, only those topics which
involves topic development or user focus change will be
considered and hence downloaded. Tweets from July 2015 to
September 2015 are collected. Usually a ratio of 7:3 will be
followed in hand classifying the training set and the remaining
thirty percent would be for testing. But the classification
Fig.2 Foreground Dynamic Topic Modelling
approach followed in this work is topic based and it is
adaptive in nature. Therefore, excessive manual labelling is
Algorithm 3: Foreground Dynamic Topic Modelling VF-DTM
avoided reducing the initial training set to about 10 to 20
Input : Sentiment labelled tweets detected from Volume percent. The classifier modelled is independent of the topic
based approach and K sub topics under study. Table 1 justifies the superiority of SVM
Output : Tweets modelled under each Topic classifier.
begin
populate foreground and background sets TABLE I
APPROPRIATENESS OF CLASSIFIER USED
capture sub topics with eq(1)
capture sequential structure between models as Classifier Name Accuracy
eq.(2s)
for each tweet do SVM 0.800
capture variational state space model as in eq(3) Ada-Boost 0.676
for each word do Decision Tree 0.710
compute eq.(4)
Naïve Bayes 0.713
update variational parameters
Random Forest 0.722
end for
end for
end Weka [10] implementation of classifiers namely Ada-Boost,
Decision Tree, Naïve Bayes and Random Forest are used for
Variational state space model is used approximate comparison. Now, the topic wise sentiment labelled data will
posterior inference using eq. (3) and (4). be considered for summarization task. Performance of the
^ extractive summary is measured using two popular metrics for
β t | β t ~ N ( β t ,ν t 2 I ) (3) summarization namely Coverage and Novelty. The coverage
wt ,n | β t ~ Mult (π ( β t )) (4) is defined using eq. (6):

 Count
exp( β t , W )
where, π ( β t ) = match ( N − gram)
 exp( β t , W )

S H S
w 1 1 d j ∈D N − gram∈di ,di
Coverage = .
where, at -per tweet distribution at time t, ν t -sub topic | DH |
d i ∈D H
wij
 Count ( N − gram)
distribution. At time t, z t -sub topic at time t, β t -per word d j ∈D S N − gram∈d iH

distribution at time t, wt is the specific word. Before (6)

generating the extractive summary for the topic modelling, we | j − i | +1, j ≠ i
where, wi, j = 
first need to sort the subtopics detected from F-DTM 1, j = i
according to their temporal information. Particularly, we use
Reason candidate model [9] to output the relationships where, D H and D S are human and system generated sub
between tweets and subtopics to assign each tweet to the summaries, i and j are the index of the sub summary. wi, j is
subtopic that it most likely belongs to. Most probable tweets the weight added corresponding to the matches between the
will be associated to each sub topic by the extracting the word system and human generated sub summaries. Novelty is
distribution only from the foreground set and not the defined using eq. (7):
background set. This process is applicable for mining topics
for specific set of data collection. The tweets being assigned Novelty =
1
| D | −1  (I
i >1
d i − I d i ,d i −1 ) (7)
signify that they best represent the topic under consideration.
Finally using eq. (5) , scores of the tweets are computed where, I di is the Information Summary and I d i , d i−1 is the
and the highest scored tweets will be considered under each overlapped information. This is used to calculate the average
sub topic as the extractive sub summaries. increment of information content of two adjacent sub
Score(Tweet ) =  p ( word | z t ), tweet i ∈ Z t
words∈tweeti
(5) summaries. Performance of baseline systems namely LDA,
DTM and Heuristic systems are compared based on the
Redundancy in each sub summary is handled using Maximal metrics namely Coverage and Novelty of the summary
Relevance Model as in [8]. generated and is shown in Table 2 below.
TABLE 2

978-1-4673-9745-2 ©2016 IEEE

International Conference on Advances in Electrical, Electronics, Information, Communication and Bio-Informatics (AEEICB16)

COMPARISON WITH BASELINE SYSTEMS proceeding with building the required model has proved to be
Proposed an efficient choice.
Trending Topic Metric Heuristic LDA DTM
VF-DTM
Coverage 0.270 0.310 0.330 0.39
#Yakub Memon V. CONCLUSION
Novelty 0.390 0.735 0.795 0.800
#Black Day Coverage 0.264 0.305 0.330 0.380 Thus the proposed Trending Topic Analysis System has
For Indian
Democracy
Novelty 0.390 0.717 0.771 0.794 been developed that is able to analyse twitter trending topics
in a constructive manner. Tweet corpus will be sentiment
labelled with common and tweet specific feature extraction,
thus enhancing the classification task. Sentiment classified
tweets are forwarded into novel summarization framework in
which the proposed Foreground Dynamic Topic Modelling
(VF-DTM) approach enables in detecting sub topics
accurately by avoiding noisy data .Then the system extracts
the most significant tweets to generate extractive summaries
using a graph based approach using proposed salient features
of the tweets. The system can be further enhanced by
normalizing the selected significant tweets for better
understanding and can be modelled to generate abstractive
summaries of the topic which also contributes to the
readability of the summary

Fig.3 Comparison with Baseline Systems
From Figure 3 it is learnt that Coverage performance of all
the baseline systems are noticeably less than the proposed
system. And for Novelty, Heuristic systems show a very poor
performance , LDA gives an acceptable performance whereas
DTM and the proposed system almost achieves near
performance of 7% increase compared to the other two, still
the proposed system VF-DTM is 2% higher than DTM.
Sub topic segmentation also contributes to the quality of
the summary generated. Evaluation of sub topic segmentation
is performed using average cosine similarity due to the fact
that the sub topics tend to cluster the tweet stream into
segments and each such segment will focus on one sub topic.
Cosine similarity for two successive sub topics are evaluated
to depict the segmentation performance of the proposed
system. Table 3 displays the average cosine similarity values
exhibited by various baseline systems for various sample
topics.
TABLE 3
SUB TOPIC SEGMENTATION PERFORMANCE
Sub Topic Detection Cosine Similarity
Models
Heuristic Baseline 0.841
LDA 0.688
OPAD 0.693
DTM 0.677
Proposed VF-DTM 0.6585
Performance of Volume based approach in sub topic
segmentation proves that detecting topic merely based on
volume will affect the segmentation performance leading to
unidentified topics in the corpus. DTM uses semantic
approach and not merely volume but the performance of
proposed system (VF-DTM) is still better than DTM due to
the reason that the model building process is not interrupted
with noisy content. Therefore distilling noise and further
978-1-4673-9745-2 ©2016 IEEE
REFERENCES
[1] Rui, Huaxia, Yizao Liu, and Andrew Whinston. “Whose and what
chatter matters? The effect of tweets on movie sales.” Decision Support
Systems55.4 (2013): pp.863-870.
[2] Wan, Xiaojun. “Co-training for cross-lingual sentiment classification.”
Proceedings of the Joint Conference of the 47th Annual Meeting of the
ACL and the 4th International Joint Conference on Natural Language
Processing of the AFNLP: Volume 1-Volume 1. Association for
Computational Linguistics, 2009.pp.235-243
[3] S.Liu, X.Cheng, F.Li, and F.Li, “TASC: Topic-Adaptive Sentiment
Classification on Dynamic Tweets”, IEEE Transactions on Knowledge
and Data Engineering, Vol. 27 ,no. 6,pp. 1696 - 1709, Jun. 2011 .
[4] Wen, Dunwei, and Geoffrey Marshall. “Automatic Twitter Topic
Summarization.” Computational Science and Engineering (CSE), 2014
IEEE 17th International Conference on. IEEE, 2014, pp-207-212
[5] Liu, Fei, Yang Liu, and Fuliang Weng. “Why is SXSW trending?:
exploring multiple text sources for Twitter topic summarization.”
Proceedings of the Workshop on Languages in Social Media.
Association for Computational Linguistics, 2011,pp.66-75
[6] O'Connor, Brendan, Michel Krieger, and David Ahn. “TweetMotif:
Exploratory Search and Topic Summarization for Twitter.” ICWSM.
2010,pp.384-385
[7] Pan, S. J., Ni, X., Sun, J. T., Yang, Q., & Chen, Z. (2010, April). “Cross-
domain sentiment classification via spectral feature alignment.”
InProceedings of the 19th international conference on World wide
web (pp. 751-760). ACM.
[8] D.Gao, W.Li, X.Cai, R.Zhang, and Y.Ouyang, “Sequential
Summarization: A Full View of Twitter Trending Topics”, IEEE
Transactions on Audio, Speech, and Language Processing, Vol. 22, no.
2,pp. 293-302,Feb. 2014.
[9] S.Tan, Y.Li, H.Sun, Z.Guan, and X.Yan, “Interpreting the Public
Sentiment Variations on Twitter”, IEEE Transactions on Knowledge
and Data Engineering, Vol. 26, no. 5, pp. 1158-1170, May. 2014.
[10] M. Hall, E. Frank, G. Holmes, B. Pfahringer, P. Reutemann, and I. H.
Witten, “The weka data mining software: An update,” ACM SIGKDD
[11] Chen, Yan, Xiaoming Zhang, Zhoujun Li, and Jun-Ping Ng. “Search
engine reinforced semi-supervised classification and graph-based
summarization of microblogs.” Neurocomputing 152 (2015): 274-286.