Exploiting Meta Attributes for Identifying Event Related Hashtags
Sreekanth Madisetty and Maunendra Sankar Desarkar
Computer Science and Engineering, IIT Hyderabad, 502285, Hyderabad, Telangana, India
Keywords:
Social Media, Information Retrieval, Learning to Rank, Twitter.
Abstract:
Users in social media often participate in discussions regarding different events happening in the physical
world (e.g., concerts, conferences, festivals) by posting messages, replying to or forwarding messages related
to such events. In various applications like event recommendation, event reporting, etc. it might be useful to
find user discussions related to such events from social media. Finding event related hashtags can be useful for
this purpose. In this paper, we focus on the problem of finding relevant hashtags for a given event. Features are
defined to identify the event related hashtags. We specifically look for features that use similarities of the hash-
tags with the event metadata attributes. A learning to rank algorithm is applied to learn the importance weights
of the features towards the task of predicting the relevance of a hashtag to the given event. We experimented on
events from four different categories (namely, Award ceremonies, E-commerce events, Festivals, and Product
launches). Experimental results show that our method significantly outperforms the baseline methods.
1 INTRODUCTION
Nowadays people are getting more and more engaged
with various social media such as Facebook, Twitter,
MySpace, etc. They post opinions, anticipations, per-
sonal feelings, etc. on multiple different topics. The
discussion items may be from a diverse range of top-
ics such as events, product features, natural calami-
ties, government policies, etc. In this paper, we fo-
cus on user discussions that are related to events.
By the word event, we mean a real world incident
or occurrence which is pre-planned, takes place at a
certain time or duration and is of interest to several
people (Becker et al., 2011; Allan, 2012). Events
can be broadly categorized into two types: planned
events (e.g., concerts, shows, festivals, conferences,
sports events, movie launch) and unplanned events
(e.g., earth quakes, tsunami) (Sakaki et al., 2010).
Finding user discussions related to planned events
from social media can be helpful in various appli-
cations, e.g., event reporting, event recommendation,
etc. People often use hashtags in the tweets. If we
find relevant hashtags for the event, then we can eas-
ily identify tweets related to the event. For exam-
ple, #www2017 relates to world wide web confer-
ence 2017 event, #Ipl2017 relates to Indian Premier
League T20 cricket 2017, #JustinBieberIndia relates
to the concert by Justin Bieber in India. By using
these hashtags, tweets related to the corresponding
events can be retrieved. However, manual selection
of these hashtags is not a scalable approach. In this
work, we focus on the problem of automated identi-
fication of high precision hashtags for given planned
events.
Hashtags from social media can be identified for
various contexts, e.g., user interest, external news ar-
ticle, recent trend, etc. This problem is often viewed
as a context sensitive hashtag recommendation prob-
lem. Although there exist algorithms for hashtag rec-
ommendation for different contexts mentioned above,
there is no published work that considers planned
events as the external context and tries to identify
hashtags relevant to it. Towards this task, we first
use the event meta information to retrieve tweets pre-
cisely related to the event. We then identify a set of
candidate hashtags for the event from this retrieved
set of tweets. Next, we propose few features for
hevent, hashtagi pairs that attempt to measure relat-
edness between the event and hashtag. These feature
scores are then combined to estimate the relevance of
the hashtag with respect to the event. We evaluated
the performance of this approach on events from four
different categories (award ceremonies, e-commerce
events, festivals, and product launches). The experi-
mental results show that the algorithm is able to iden-
tify the hashtags that are truly relevant to the event.
Rest of the paper is organized as follows. Related
literature for current work is presented in Section 2.
Next in Section 3, problem statement of our work is
defined. Details of the proposed method are described
Madisetty S. and Desarkar M.
Exploiting Meta Attributes for Identifying Event Related Hashtags.
DOI: 10.5220/0006502602380245
In Proceedings of the 9th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (KDIR 2017), pages 238-245
ISBN: 978-989-758-271-4
Copyright
c
 2017 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved

in Section 4. Experimental evaluation of the method
is shown in Section 5. We conclude the work by pro-
viding directions for future research in Section 6.
2 RELATED WORK
As mentioned in the above section, there is no work
in literature that uses planned events as context for
the hashtag identification problem. Here, we dis-
cuss about some of the recent approaches for con-
text sensitive hashtag recommendation. Research on
hashtag recommendation has been receiving consid-
erable attention in recent years. A method for content-
based hashtag recommendation using Latent Dirich-
let Allocation (LDA) is described in (Godin et al.,
2013). However, the authors recommend the key-
words from the topic distribution of a tweet and take
the suggestions from the evaluators to know the qual-
ity of suggested keyword as a hashtag. Recommend-
ing hashtags for hyperlinked tweets is proposed in
(Sedhai and Sun, 2014). The authors showed that
functions of hashtags could be extended to the linked
documents from hyperlinked tweets. However, this
method works only for hyperlinked tweets whereas
less fraction of the tweets actually contain hyperlinks.
(Wang et al., 2013) proposed an adaptive crawl-
ing model that identifies emerging popular (having
high frequency) hashtags and monitors them to re-
trieve larger amounts of associated content for an
event. (Dovgopol and Nohelty, 2015) proposed an
approach for hashtag recommendation in Twitter by
using Naive Bayes approach. The authors considered
the hashtag as a class and words in the tweet are fea-
tures. Both (Wang et al., 2013; Dovgopol and No-
helty, 2015) have a strong bias towards the frequency
of hashtags in the tweets obtained for some event-
related seed queries fired to Twitter.
Hashtags are recommended for enterprise applica-
tions, emails, enterprise social networks, and special
interest group mail lists in (Mahajan et al., 2016). The
authors considered three scenarios, namely, Inline,
Post, and Auto-complete and used three types of fea-
tures, namely, temporal, structural, and content. A
method to recommend hashtags using attention-based
convolution neural network is described in (Gong and
Zhang, 2016). Real-time hashtag recommendation for
streaming news is proposed in (Shi et al., 2016). Se-
mantic similarity of a hashtag to existing news articles
is obtained by comparing the similarity of the article
with the tweet bag of the hashtags. The performance
of the algorithm would degrade if the tweet bags of
the hashtags are not known or are small in size. More-
over, the news articles are generally large, which is
not true for event descriptions. The focus on hashtag
semantics is limited in the existing work in literature.
3 PROBLEM DEFINITION
We now briefly define the problem addressed in this
paper: Given metadata of an event E, find a list of
hashtags relevant to the event E. Event metadata com-
prises of context features of the event such as title,
venue, time, location, and performer(s) of the event.
Event metadata can be obtained from several event
aggregation sites (e.g., Eventbrite, Eventful, last.fm).
The following is an example of event metadata in
JSON format.
{
"title": "Le ciel, la nuit et la pierre
glorieuse avignon"
"venue": "Jardin Ceccano"
"location": ""
"performers": "La Piccola Familia"
"date": "12th August 2016"
}
As it can be seen from the example, some of the meta-
data entries are missing.
4 METHODOLOGY
We use a two-phase approach for identifying relevant
hashtags for a given event. In the first phase, we re-
trieve a set of candidate hashtags for an event from
Twitter. This phase is described in Section 4.1. In the
second phase, we rank the hashtags from this candi-
date set according to their relevances with the event.
The method for finding relevance scores is presented
in Section 4.2.
4.1 Finding Candidate Hashtags
In this phase, given metadata of an event, we first
identify a set of tweets for the event from Twitter. We
use the precision query approach presented in (Becker
et al., 2012) for retrieving the tweets for the event.
Precision queries are queries which retrieve highly
relevant results for the specific information need. To
create precision queries for a given event E, different
combinations of its metadata features, namely, title,
location, and performer are used. A set of such preci-
sion queries (Q
E
) are submitted to the Twitter search
API. Hashtags that appear in the tweet bag (T B
E
) re-
turned by Twitter for this call are added to the candi-
date set. As the keywords of the precision query come
from the event title and venue, the retrieved tweets

generally match well with the event under considera-
tion. The candidate set thus generated contains a huge
number of hashtags.
4.2 Giving Scores to Candidate
Hashtags
The next phase of the algorithm assigns a relevance
score to each of these candidate hashtags. We iden-
tify a set of features that we consider important for
measuring this relevance for an hevent, hashtagi pair.
These feature scores are linearly combined to get the
final score of the hashtag for that event. In the follow-
ing discussion, we use EM (e.g., title, location, per-
former) to denote event metadata and HT to denote
the hashtag.
4.2.1 Features
• Frequency of Hashtag ( f
1
): This is the frequency
of the hashtag in tweet bag T B
E
of the event E.
Tweet corpus is different for different events. Let
the raw frequency of hashtag HT in tweet corpus
for event E be f req
HT,E
.
f
1
=



1 + log( f req
HT,E
) if f req
HT,E
> 0
0 otherwise
We have used log frequency of the hashtag.
• Bigram Feature ( f
2
): This feature computes the
number of common character-level bigrams present
in the hashtag HT and event metadata EM. If HT
B
is a set of Hashtag Bigrams and EM
B
is a set of
Event Metadata Bigrams then the value of this fea-
ture is computed as
f
2
= |HT
B
∩ EM
B
|
For example, Bigrams for hahstag #iPhone7 are #i,
iP, Ph, ho, on, ne. For the event metadata EM we
find the set of bigrams for the available event meta-
data component (e.g. title, performer, location.)
and take the union of these sets to get EM
B
.
• Trigram Feature ( f
3
): This feature counts the
number of common character-level trigrams present
in the hashtag HT and event metadata EM. If HT
T
is a set of hashtag trigrams and EM
T
is a set of event
metadata trigrams then the value of this feature is
computed as
f
3
= |HT
T
∩ EM
T
|
For example, trigrams for hashtag #samsung-
galaxyc7pro are #sa, sam, ams, msu, sun, ung, ngg,
gga, gal, ala, lax, axy, xyc, yc7, c7p, 7pr, pro.
• Bigrams of Top-K trigrams ( f
4
): Let S be the
set of Top-K word-level trigrams of an event E
which are obtained from T B
E
. S
KB
is the union of
character-level bigrams obtained from the elements
of S. Score according to this feature is computed as
f
4
= |HT
B
∩ S
KB
|
This feature specifies the number of bigrams that
are common in both hashtag HT and Top-K tri-
grams of tweet corpus of an event. We set K=30
in our algorithm.
• Subsequence Feature ( f
5
): This feature checks
whether HT is a subsequence of EM or not. String
A is a subsequence of string B if and only if A is
obtained by deleting some elements from B without
changing the order of remaining elements. For ex-
ample, if “Knowledge Discovery and Information
Retrieval” is event metadata EM then “KDIR” is a
subsequence of EM.
f
5
=

1 if HT is a subsequence of EM
0 otherwise
Except frequency feature, all other features try
to match the hashtag’s appearance or construct with
event metadata and try to capture semantic related-
ness between hevent, hashtagi pair.
4.2.2 Combining Feature Scores
Given an hevent, hashtagi pair, the different feature
scores can be obtained by following the descriptions
given above. Next, we want to find a weighted com-
bination of these individual feature scores to deter-
mine a single score for each hevent, hashtagi pair.
Given an event, hashtags with the highest values of
this score can be output as the relevant hashtags
for the event. We use a learning to rank algorithm
(SV M
Rank
) (Joachims, 2006) for finding the weights.
SV M
Rank
is a pairwise learning to rank approach. It
is a supervised machine learning algorithm. In our
setting, each instance of the supervised data has the
hevent, hashtagi feature scores and a relevance judge-
ment indicating the degree of relevance of the hashtag
for the event.
Given an event E, and a set of hashtags H =
{h
1
, h
2
, · ·· , h
n
}, the method attempts to construct the
pairwise ranking matrix R. It is constructed for an
event E with |H| × |H| dimensions. The (i, j)
th
en-
try of matrix R is 1 if h
i
is more relevant than h
j
for
the event E and 0 otherwise. For this, the method
learns a set of weights w over the hevent, hashtagi
features. If the feature vector for the event-hashtag
pair < E, h
i
> is denoted as Φ(E, h
i
), then the method
computes the relevance scores s(E, h
i
) = w
T
Φ(E, h
i
)

(a) NDCG (b) Precision
Figure 1: Comparing our proposed method with other alternative approaches (for all the events).
and s(E, h
j
) = w
T
Φ(E, h
j
). Then, h
i
is considered to
be more relevant than h
j
if s(E, h
i
) > s(E, h
j
). This
information can then be used to construct the rating
matrix
ˆ
R which is the prediction for the actual matrix
R for the set of hashtags available for the event. The
method learns the weight vector w by using the train-
ing data. It tries to identify the w that has low value
of this reconstruction error on R.
Given a ranking r, the corresponding pairwise
ranking matrix R can be constructed easily. SV M
Rank
models the learning of this R as minimizing the dis-
tance between the actual matrices R and the recon-
structed matrices
ˆ
R. The difference between the ac-
tual R and the predicted
ˆ
R can be computed as the
number of cells in which they disagree. One way to
minimize this disagreement count is to minimize the
Kendall Tau distance between the rankings r and r
0
.
Kendall Tau coefficient (τ) measures the difference
between two rankings. The pair h
i
6= h
j
is concordant
if r and r
0
agree on relative ordering of h
i
and h
j
and
discordant otherwise. τ between r and r
0
is calculated
as follows.
τ =
(#concordant pairs) − (#discordant pairs)
(#concordant pairs) + (#discordant pairs)
(1)
τ ranges between -1 and +1. The SV M
Rank
algo-
rithm tries to minimize the following loss function.
1
m
m
∑
i=1
−τ(r
f (E
i
)
, r
i
) (2)
where r
f (E
i
)
is predicted ranking for the event E
i
.
Minimizing the above loss function is same as min-
imizing discordant pairs for each event. This opti-
mization can be formulated as
minimize
w
1
2
w
T
w +C
∑
i, j,k
ε
i, j,k
(3)
subject to:
∀k and i 6= j ∈ {1, ..., n
k
} with h
ki
>
E
k
h
k j
(4)
w
T
Φ(E
k
, h
ki
) ≥ w
T
Φ(E
k
, h
k j
) + 1 − ε
i, j,k
(5)
ε
i, j,k
≥ 0 (6)
w is a weight vector, Φ(E, h) is a mapping onto
feature vectors that describe the similarity between
event E and hashtag h, C is a penalty parameter, and
ε
i, j,k
are (non-negative) slack variables.
Once the weight vector w is learned from the train-
ing data, candidate hashtags H = {h
1
, h
2
, ..., h
n
} for
any new event E can be ranked according to their rel-
evance scores s(E, h
i
) = w
T
Φ(E, h
i
).
5 EXPERIMENTS
In this section, we evaluate the performance of the
proposed method. The data for the experiment
was collected using Twitter streaming API. There
are four categories (Award ceremonies, E-commerce
events, Festivals, and Product launches) in the dataset.
The Award ceremonies category contains five events.
They are National film awards, Jio MAMI awards,
IIFA Utsavam awards, TSR-TV9 film awards, Zee ap-
sara awards. The E-commerce events contain four
events. They are Flipkart freedom sale, Super Sat-
urday Mumbai sale, Myntra fashion sale, Flipkart
Big billion days. The festival category contains in-
formation about Indian festivals, and seven events
are present in this category. They are Ram Navami,
Ganesh Chaturthi, Raksha Bandhan, Sri Krishna
Janmashtami, Hanuman Jayanthi, Bakrid, Ramzan.
Here, each festival is treated as an event. The Prod-
uct launches category contains information about new
product releases in the market, and seven events are

(a) NDCG (b) Precision
Figure 2: Category-wise comparison of NDCG and precision of top-k hashtags for two different categories using our method.
Table 1: Award Cermonies NDCG.
Rank K FreqPearson AlleqW AlldiffW
5 0.746 0.948 0.958
10 0.686 0.884 0.916
15 0.713 0.873 0.877
20 0.748 0.870 0.864
25 0.751 0.858 0.869
30 0.773 0.860 0.869
35 0.799 0.863 0.885
40 0.814 0.886 0.890
45 0.843 0.916 0.922
50 0.884 0.951 0.954
Table 2: E-commerce Events NDCG.
Rank K FreqPearson AlleqW AlldiffW
5 0.677 0.786 0.899
10 0.648 0.733 0.861
15 0.682 0.779 0.868
20 0.680 0.802 0.891
25 0.691 0.799 0.899
30 0.722 0.832 0.896
35 0.747 0.850 0.908
40 0.767 0.864 0.921
45 0.791 0.885 0.932
50 0.821 0.891 0.941
Table 3: Festivals NDCG.
Rank K FreqPearson AlleqW AlldiffW
5 0.870 0.971 0.967
10 0.708 0.900 0.983
15 0.652 0.810 0.989
20 0.674 0.764 0.950
25 0.726 0.697 0.960
30 0.753 0.676 0.967
35 0.799 0.641 0.943
40 0.819 0.618 0.921
45 0.854 0.613 0.907
50 0.873 0.594 0.907
Table 4: Product Launches NDCG.
Rank K FreqPearson AlleqW AlldiffW
5 0.675 0.830 0.936
10 0.662 0.808 0.878
15 0.654 0.801 0.893
20 0.678 0.824 0.898
25 0.706 0.844 0.905
30 0.729 0.859 0.913
35 0.741 0.873 0.918
40 0.747 0.884 0.937
45 0.781 0.897 0.946
50 0.801 0.906 0.953
present in this category. They are Reliance Jio, Moto
G5 launch, Le Tv Super3, Zopo F2 launch, Samsung
C7 Pro, Nubia Z11 mini, Swipe Elite Plus.
A pooling exercise was performed for generating
a labeled data set for evaluation. By using the fea-
tures defined in Section 4.2.1, 100 hashtags are re-
trieved for each event. All the hashtags thus retrieved
were given to 5 volunteers for relevance judgements.
Volunteers were asked to choose from three relevance
labels: 2 being highly relevant to the event, 1 being
moderately relevant to the event, 0 being irrelevant to
the event. For each hevent, hashtagi pair, the median
of labels entered by the volunteers for that pair was
used as the final label. However, for around 90% of
the hevent, hashtagi pairs, the same relevance label is
given by all the volunteers.
We compare the proposed method with the meth-
ods mentioned below.
• FreqPearson: It is the combination of frequency
and Pearson correlation feature (Wang et al.,
2013). Correlation between two hashtags is calcu-
lated by dividing the time frame into several time
slots, and the sequence is the frequency counts of
each time slot.
• AlldiffW: This is our proposed method which
combines the feature scores using the weights de-

termined by the SV M
Rank
algorithm. We used 10-
fold cross-validation to split the train and test sets.
• AlleqW: This is a standard baseline for our algo-
rithm. This is the combination of all features men-
tioned in Section 4.2 with equal weights given to
all features.
The performance of our method was evaluated using
the evaluation metrics NDCG, Precision. These met-
rics are widely used in Information Retrieval litera-
ture. For both these measures, higher values indicate
better performance.
5.1 Results and Discussions
We now present detailed experimental analysis of the
proposed method.
5.1.1 Comparison with other Methods
The comparison with other methods is presented in
Figure 1. The NDCG values are compared in Fig-
ure 1a and precision values are compared in Figure
1b. It is clear that the performance of the proposed
method is significantly better than the other methods
used for comparison. This is because the frequency of
the hashtags plays a significant role in the algorithm
(Wang et al., 2013) taken from literature. Hence,
they are more biased towards frequency. However,
along with frequency, we consider various other fea-
tures that attempt to measure the semantic relatedness
between the event and hashtag. The other methods
fail to capture semantic relatedness and hence keep
retrieving the hashtags that are more frequent but un-
related to the event. It can be observed that even
our baseline method achieves high scores than Freq-
Pearson. This signifies the usefulness of the semantic
features described in this work. The performance of
AlldiffW (weights are learned) is better than AlleqW
(uniform weights). This indicates the importance of
supervision along with semantic features.
5.1.2 Category-wise Comparison
Category-wise comparison of NDCG is presented in
Figure 2a and comparison of precision is presented
in Figure 2b. In Figure 2a, Award ceremonies cate-
gory is performing better than all other categories up
to NDCG@10. After that Product launches category
outperforms the other categories. In Figure 2b, Festi-
vals category is performing better than all other cate-
gories. Category-wise comparison of NDCG and pre-
cision with baseline method and FreqPearson method
is also described. NDCG comparison of Award cer-
emonies, E-commerce events, Festivals, and Product
launches is presented in Table 1, Table 2, Table 3, and
Table 4 respectively. Best performing method values
are put in bold. We observe that proposed method
is performing better at NDCG@5 for all categories
except Festivals category. AlleqW is performing bet-
ter for Festivals category which is also our baseline
method. For the remaining values proposed method
outperforms all other methods.
Category-wise precision comparison of Award
ceremonies, E-commerce events, Festivals, and Prod-
uct launches are presented in Table 6, Table 7, Table
8, and Table 9 respectively. Best performing method
values are put in bold. Similar to NDCG@5, AlleqW
precision value is higher than all other methods in
Festivals category for precision comparison as shown
in Table 8. For all other Precision@k where k = 5 to
50 our proposed method outperforms all other meth-
ods for all categories.We also presented the hashtags
obtained by different methods for different categories
in Table 5. Irrelevant hashtags are put in italic and
red color. We observe that the proposed method re-
trieves more relevant hashtags than other methods.
We applied the model learned from our data to iden-
tify relevant hashtags for the four festivals mentioned
in CLEF 2017 lab microblog dataset (Ermakova et al.,
2017). The identified hashtags are presented in Ta-
ble 10. Ground truth information is not available for
this dataset. Also, our volunteers are not able to pro-
vide relevance judgement for these hashtags due to
lack of knowledge about those festivals and the so-
cial/cultural contexts in which the candidate hashtags
can appear in the tweets related to these festivals.
However, by looking at the hashtags, it appears that
the hashtags are relevant to the event under consider-
ation.
Table 11 shows results of ablation experiments of
NDCG where features are added with equal weights
but remove one feature at a time. The most influen-
tial feature is Bigrams feature. This feature captures
the semantic similarity between event metadata and
hashtag. The second most important feature is Tri-
grams. Subsequence and frequency are the next im-
portant features.
6 CONCLUSION
In this paper, we focused on the problem of identify-
ing the relevant hashtags for planned events. We iden-
tified a set of features related to the hevent, hashtagi
pairs. We presented a model for combining feature
scores and learned the weights using learning to rank
algorithm.
We used our algorithm to retrieve hashtags from

Table 5: Top ten hashtags for one event of each category. Hashtags in italic and red colour are not relevant to the event.
Event FreqPearson Proposed Method Hashtags retrieved by our
method but missed by other
method
National Film
Awards (Award
Ceremonies)
#nationalfilmawards, #rustom,
#nationalaward, #24themovie,
#akshaykumar, #neerja, #bestac-
tor, #nationalawards, #dangal,
#zairawasim
#nationalfilmawards, #64thnation-
alfilmawards, #nationalfilmaward,
#nationalfilmawards2017, ##na-
tionalfilmawards, #64nation-
alfilmawards, #nationalaward,
#nationalawards, #nationalfil-
mawardsindia, #64thnationalfil-
maward
#64thnationalfilmawards, #64na-
tionalfilmawards, #nationalfil-
mawardsindia
Flipkart Big
Billion Days
(E-commerce
Events)
#bigbilliondays, #shoponbigbil-
liondays, #flipkart, #greatindian-
festival, #mobilesonbigbilliondays,
#bbd, #fashion, #unboxdiwal-
ibestoffers, #unboxdiwalisale,
#amazon
#bigbilliondays, #shoponbig-
billiondays, #mobilesonbigbil-
liondays, #bigbilliondays2016,
#flipkartbigbillionsale, #electron-
icsonbigbilliondays, #thebigbil-
liondays, #bigbilliondaystonight,
#bigbilliondaysareback, #bigbil-
liondayssneakpeek
#bigbilliondays2016, #flipkartbig-
billionsale, #electronicsonbigbil-
liondays, #bigbilliondaystonight,
#bigbilliondaysareback, #bigbil-
liondayssneakpeek
Janmashtami (Fes-
tivals)
#happyjanmashtami, #janmash-
tami, #krishna, #trlday4, #dahi-
handi, #krishnajanmashtami,
#trlday3, #lordkrishna, #happy,
#jaishrikrishna
#krishnajanmashtami, #happyjan-
mashtami, #happykrishnajanmash-
tami, #janmashtamicelebrations,
#happysrikrishnajanmashtami,
#happykrishnajayanthi, #hap-
pykrishnajanmashthami, #hap-
pykrishnashtami, #janmashtami,
#srikrishnajayanti
#happykrishnajanmashtami,
#janmashtamicelebrations,
#happysrikrishnajanmashtami,
#happykrishnajayanthi, #happykr-
ishnashtami, #srikrishnajayanti
Reliance Jio
Launch (Product
Launches)
#jio, #reliancejio, #relianceagm,
#jiodigitallife, #reliancejio4g, #jio-
fan, #jio4g, #reliance, #muke-
shambani, #airtel
#reliancejio4g, #reliancejio, #re-
liancejiolaunch, #relianceagm,
#reliance, #reliancejioishere,
#reliancejio’s, #reliancejio4g’s,
#reliancea, #reliancejio4
#reliancejiolaunch, #reliance-
jioishere
Table 6: Award Ceremonies Precision.
Rank K FreqPearson AlleqW AlldiffW
5 0.960 1.000 1.000
10 0.800 0.920 0.980
15 0.787 0.880 0.893
20 0.790 0.830 0.880
25 0.776 0.768 0.840
30 0.753 0.760 0.813
35 0.749 0.731 0.811
40 0.725 0.720 0.785
45 0.716 0.711 0.778
50 0.708 0.708 0.764
Table 7: E-commerce Events Precision.
Rank K FreqPearson AlleqW AlldiffW
5 0.400 0.750 0.950
10 0.275 0.600 0.800
15 0.300 0.567 0.733
20 0.288 0.513 0.700
25 0.290 0.460 0.660
30 0.267 0.450 0.600
35 0.257 0.443 0.564
40 0.256 0.413 0.538
45 0.261 0.389 0.511
50 0.265 0.365 0.480
Table 8: Festivals Precision.
Rank K FreqPearson AlleqW AlldiffW
5 0.880 0.971 0.967
10 0.680 0.900 0.983
15 0.560 0.810 0.989
20 0.530 0.764 0.950
25 0.496 0.697 0.960
30 0.440 0.676 0.967
35 0.429 0.641 0.943
40 0.405 0.618 0.921
45 0.387 0.613 0.907
50 0.368 0.594 0.907
Table 9: Product Launches Precision.
Rank K FreqPearson AlleqW AlldiffW
5 0.457 0.829 0.914
10 0.343 0.671 0.786
15 0.276 0.533 0.667
20 0.271 0.500 0.621
25 0.246 0.469 0.571
30 0.238 0.419 0.529
35 0.220 0.388 0.478
40 0.204 0.361 0.454
45 0.203 0.327 0.432
50 0.203 0.311 0.400
different events. Efficacy of the proposed method was
established with multiple evaluation metrics, namely,
NDCG, Precision. The work shows that identifica-
tion of semantic relatedness between the hashtag and
the event metadata helps in better retrieval of relevant
hashtags. As an extension to this work, we want to

Table 10: Top five hashtags from four different events of CLEF 2017 lab microblog dataset.
Festival 1: Anna Calvi,
charrues
Festival 2: La Piccola Fa-
milia, avignon
Festival 3: Suitable for par-
ties, transmusicales
Festival 4: Vanishing Point,
edinburgh
#vieillescharrues2015,
#annacalvi, #charrues,
#labelcharrues, #vieillechar-
rues2015
#piccolafamilia, #lapic-
colafamilia, #lafamilia,
#festivaldelafamilia, #frente-
nacionalxlafamilia
#transmusicales, #transmu-
sicales2015, #eventosmu-
sicales, #noticiasmusicales,
#rencontrestransmusicales
#vanishingpoint, #edinburgh-
festivalfringe, #thedestroyed-
room, #edinburghfringe2016,
##edinburg
Table 11: NDCG values obtained on four categories with one of the features removed.
Rank K AlleqW AlldiffW All-{Bigrams} All-{Trigrams} All-{Subsequence} All-{Frequency}
5 0.865 0.926 0.846 0.853 0.888 0.869
10 0.816 0.878 0.811 0.815 0.841 0.869
15 0.804 0.874 0.785 0.789 0.811 0.866
20 0.811 0.869 0.782 0.781 0.810 0.865
25 0.812 0.876 0.793 0.793 0.819 0.870
30 0.832 0.883 0.810 0.809 0.833 0.874
35 0.848 0.894 0.829 0.832 0.846 0.887
40 0.868 0.908 0.849 0.854 0.873 0.903
45 0.896 0.927 0.875 0.880 0.893 0.920
50 0.916 0.949 0.906 0.906 0.922 0.941
identify additional features for hevent, hashtagi pairs.
Also, we want to evaluate the proposed method’s per-
formance on other categories. Moreover, we would
like to see whether the proposed strategy is able to
retrieve hashtags for individual events which are part
of large-scale events (e.g., Rio Olympics, World Cup)
that are agglomerate of various individual events. It
would be an interesting work to use the proposed
method to retrieve relevant tweets for an event and
evaluate the quality of retrieved tweets.
REFERENCES
Allan, J. (2012). Topic detection and tracking: event-based
information organization, volume 12. Springer Sci-
ence & Business Media.
Becker, H., Iter, D., Naaman, M., and Gravano, L. (2012).
Identifying content for planned events across social
media sites. In Proceedings of the fifth ACM inter-
national conference on Web search and data mining,
pages 533–542. ACM.
Becker, H., Naaman, M., and Gravano, L. (2011). Be-
yond trending topics: Real-world event identification
on twitter. ICWSM, 11(2011):438–441.
Dovgopol, R. and Nohelty, M. (2015). Twitter hash tag rec-
ommendation. arXiv preprint arXiv:1502.00094.
Ermakova, L., Goeuriot, L., Mothe, J., Mulhem, P., Nie, J.-
Y., and SanJuan, E. (2017). Clef 2017 microblog cul-
tural contextualization lab overview. In International
Conference of the Cross-Language Evaluation Forum
for European Languages, pages 304–314. Springer.
Godin, F., Slavkovikj, V., De Neve, W., Schrauwen, B., and
Van de Walle, R. (2013). Using topic models for twit-
ter hashtag recommendation. In Proceedings of the
22nd International Conference on World Wide Web,
pages 593–596. ACM.
Gong, Y. and Zhang, Q. (2016). Hashtag recommendation
using attention-based convolutional neural network.
In IJCAI, pages 2782–2788.
Joachims, T. (2006). Training linear svms in linear time. In
Proceedings of the 12th ACM SIGKDD international
conference on Knowledge discovery and data mining,
pages 217–226. ACM.
Mahajan, D., Kolathur, V., Bansal, C., Parthasarathy, S.,
Sellamanickam, S., Keerthi, S., and Gehrke, J. (2016).
Hashtag recommendation for enterprise applications.
In Proceedings of the 25th ACM International on Con-
ference on Information and Knowledge Management,
pages 893–902. ACM.
Sakaki, T., Okazaki, M., and Matsuo, Y. (2010). Earthquake
shakes twitter users: real-time event detection by so-
cial sensors. In Proceedings of the 19th international
conference on World wide web, pages 851–860. ACM.
Sedhai, S. and Sun, A. (2014). Hashtag recommendation
for hyperlinked tweets. In Proceedings of the 37th
international ACM SIGIR conference on Research &
development in information retrieval, pages 831–834.
ACM.
Shi, B., Ifrim, G., and Hurley, N. (2016). Learning-to-rank
for real-time high-precision hashtag recommendation
for streaming news. In Proceedings of the 25th Inter-
national Conference on World Wide Web, pages 1191–
1202.
Wang, X., Tokarchuk, L., Cuadrado, F., and Poslad, S.
(2013). Exploiting hashtags for adaptive microblog
crawling. In Proceedings of the 2013 ieee/acm inter-
national conference on advances in social networks
analysis and mining, pages 311–315. ACM.