=Paper=
{{Paper
|id=None
|storemode=property
|title=Analyzing Twitter for Social TV: Sentiment Extraction for Sports
|pdfUrl=https://ceur-ws.org/Vol-720/Zhao.pdf
|volume=Vol-720
}}
==Analyzing Twitter for Social TV: Sentiment Extraction for Sports==
https://ceur-ws.org/Vol-720/Zhao.pdf
Analyzing Twitter for Social TV: Sentiment Extraction for Sports
Siqi Zhao and Lin Zhong Jehan Wickramasuriya and Venu Vasudevan
Dept. of Electrical & Computer Engineering Applied Research
Rice University Motorola Mobility
6100 Main Street 600 N. U.S. Highway 45
Houston, TX Libertyville, IL
{siqi.zhao, lzhong}@rice.edu {jehan, venu.vasudevan}@motorola.com
Abstract a game becomes more interesting when the scores are
close. When there are multiple games ongoing at the
As TV watchers tweet about how they feel and what same time, e.g., up to 10 games take place on Sunday
they see, they produce valuable information not only afternoon in the NFL regular season; it is nontrivial for a
about the TV program but also how engaged they are to watcher to determine which game is the most interesting.
the program. We have already built a web service, On the other hand, without knowing what is going on in
SportSense, which recognizes major events in the US the program or how engaged the TV audience is to a
National Football League (NFL) games within 40 se- program at a certain time, an advertiser can only fill ad-
conds after an event takes place by analyzing data re- vertisement slots based on general demographics of the
trieved from Twitter in real-time. In this paper, we report audience and the overall rating of the program.
our effort to extend SportSense to extract TV watchers‟
sentimental reaction to major events in live broadcast We envision a social TV system in which watchers
sports games in real-time and present our ongoing work and advertisers can see how other watchers are enjoying
that leverages SportSense for a social TV system that a live broadcast program in real-time and what is hap-
enables TV watchers to better select interesting pro- pening in the program. In this system, a watcher can
grams in real-time and to produce personalized program switch to a program when peers are excited by it or a
summaries and enables advertisers to customize ads certain event just takes place. Similarly, an advertiser
based on recognized events and extracted audience sen- can pay more to get a time slot when the watchers are
timents. highly engaged or even use a commercial that resonates
with the event that just happened.
Categories and Subject Descriptors While there are previous efforts based on video
H5.m [Information Interfaces and Presentation]: analysis toward this goal [1, 2], we are interested in us-
Miscellaneous ing Twitter to recognize major events and extract audi-
ence sentiments in real-time. Compared to video, Twitter
General Terms has many unique advantages toward the envisioned so-
Algorithms, Design, Experimentation, Measurement. cial TV system. (i) Tweets posted by TV watchers di-
rectly reflect what they feel and think about the program.
Keywords While video analysis allows one to learn about the pro-
Twitter, social TV, sentiment extraction. gram, Twitter analysis directly allows us to learn about
the program audience. With over 200 million active
users, Twitter ensures good coverage of popular TV pro-
1. Introduction grams [3]. (ii) Second, Live video streaming is available
Live broadcast TV programs such as sports games, only at a high financial, bandwidth, and computing cost.
reality shows, and concerts often attract a large number In contrast, public tweets are free and are easy to retrieve
of loyal watchers, creating tremendous commercial op- due to their brief, textual nature. (iii) Thirdly, the textual
portunities. Key to their attractions is their inherent un- nature of tweets makes them amenable to lexicon-based
certainty or unpredictability. On the other hand, the un- analysis. As a result, end users can easily personalize
predictability challenges both TV watchers and advertis- Twitter-based event recognition and sentiment extraction
ers. On one hand, while some TV watchers tune to a live by using the right keywords. This is significantly easier
broadcast program for its entire duration, many jump than personalizing a video analysis tool to recognize a
from one channel to another occasionally looking for the new event. (iv) Finally, Twitter allows tweets to be re-
most interesting one at a moment because a live broad- trieved by tweeters. Therefore, end users of the Twitter-
cast program is not always equally interesting through- empowered social TV system can personalize their sys-
out its duration. This is particularly true for sports where
tem by tuning to a personalized collection of tweeters, below 5%. Further, we found that the team names appear
creating a personalized social information channel. in over 60% of the game-related tweets.
To demonstrate the feasibility of using Twitter for 2.2 Event Recognition
the envisioned social TV system, we have designed and
SportSense recognizes NFL events by analyzing the
realized a web service, called SportSense that recognizes
content of game-related tweets collected using the lexi-
major events in a sports game in real-time. During the
con-based heuristic and examining the post rate of
2010-2011 NFL season, SportSense is able to recognize
tweets with keywords related to game events.
major events such as touchdowns and interceptions with
SportSense employs a simple method to detect an event
90% accuracy within 40 seconds after an event happens.
by calculating the post rate increase as the ratio of the
Through a web site, the service visualized the results and
post rate in the second half of a sliding time window to
provided a popularity “thermometer” for games that
that in the first half. To overcome an undocumented
were played at the same time.
Twitter limit to the Streaming API of 50 tweets per se-
In this workshop paper, we present our ongoing cond, SportSense focuses on tweets with a keyword re-
work in extracting audience sentiments from Twitter lated to the game event only.
analysis and realizing Twitter-enabled, social, electronic
The size of the window will have a significant im-
program guide (EPG) and audience-aware advertisement
pact on the tradeoff between the delay and accuracy of
auction. Although we focus on NFL games in this work,
event detection. A shorter window will lead to a smaller
most of the techniques can be readily applied to many
delay but may have a poor performance when the post
other sports games that have a similarly sized fan popu-
rate is low and, therefore, there are not many tweets
lation and have similar frequencies of major events, e.g.,
posted in the time window.
soccer, baseball, and basketball.
To achieve the best tradeoff, SportSense selects the
The rest of this paper is organized as follows. Sec-
window size adaptively. The sliding window has a vari-
tion 2 introduces the background of SportSense which is
able size of 10, 20, 30, or 60 seconds; and each window
a web service that utilizes Twitter to recognize sports
is divided into two sub-windows of equal length. At eve-
events in real-time. Section 3 provides detailed explana-
ry second, SportSense will start from the shortest win-
tion of our approach on Twitter analysis for sentiment
dow, 10 seconds, to examine the post rate of the event
extraction. Then, Section 4 presents the applications that
keywords in the window and the post rate ratio of each
can adopt proposed approach to enhance the TV watch-
event type between the two halves. If both the post rate
ing experience. Section 5 demonstrates the effectiveness
and the ratio exceed their corresponding thresholds, the
of our approach for various types of sports, followed by
corresponding event will be recognized; otherwise the
the related work and conclusion in Section 6 and Section
window size will increment. The threshold for the post
7, respectively.
rate of tweets will filter out the random presence of the
2. Background of SportSense event keywords when people predict events, or discuss
about past events. From our observation, the random
We first describe SportSense [4, 5], a web service
presence has a low post rate and low frequency, i.e. at
that recognizes major events from live broadcast NFL
most 2 tweets in a second and usually null. The value is
games in real-time by analyzing Twitter. SportSense
set to half of the window size which is 5, 10, 15, and 30
heavily leverages the bounded vocabulary of sports
tweets. The post rate ratio between two sub-windows
events and employs lexicon-based heuristics to retrieve
will examine the increase of the post rate. This ratio is
game-related tweets in real-time and recognize events.
the post rate in the second half window to the post rate
2.1 Lexicon-based Game Tweets Separation of those in the first half. The value is set to 1.3 in this
SportSense uses a predetermined lexicon that in- case; that means the post rate in the second half window
cludes game terminology and team names to retrieve needs to be at least 1.3 times of the post rate in the first
tweets related to NFL games through the Twitter half to proceed.
Streaming API. SportSense relies on the team names to Half of events can be detected using the window
collect data when multiple games take place at the same size less than 20 seconds and more than 2/3 of events
time and attribute these tweets to games based on the can be detected using the window size less than 30 se-
mentioned team names. This lexicon-based approach is conds. Since we halve the window to detect events, the
effective to separate the game related tweets. By manual- delay of the system is half of the window size. As a re-
ly examining random 5% of the tweets, about 2,000, sult, this method will introduce less than 10 second delay
posted during the 2010 NFL Super Bowl by the NFL for half of the events.
followers, we find that extraction by 10 keywords in-
cluding game terminology and team names achieves a
2.3 Web Service Realization
false negative rate below 9% and a false positive rate We have implemented the solution described above
as a real-time web service that visualizes event recogni-
2
Figure 1: Visualized results of real-time event recognition and sentiment extraction at SportSense.com
tion results through a website throughout the 2010-2011 be retrieved from the website by either team name or
NFL season. The implementation is coded in PHP and game schedule.
consists of the backend for data collection and analysis
2.4 Recognition Performance
and the frontend for web visualization.
The web service has been active since Week 8 of the
The backend consists of two parallel modules and a
2010-2011 NFL season. Using data from the NFL web-
MySQL database. The data collection module collects
site as the ground truth we examined the events recogni-
game-related tweets through Twitter Streaming API.
tion performance of SportSense for games in the last two
Collected tweets are saved in the MySQL database. The
weeks (27 games in Week 16-17), playoffs (5 games),
event recognition module will retrieve tweets from data-
and the Super Bowl. We considered not only how many
base, separate tweets to games, recognize events, and
events have been correctly recognized but also how
generate the results in Google Chart format.
many are missed. Note that four events are targeted:
As analyzed above, the backend can introduce many touchdown, interception, field goal, and fumble.
seconds of delay to event recognition. To minimize this SportSense recognized all the events in the Super Bowl
delay, we created multiple threads to maximize the par- game without false positives and perform decently for
allelism of data analysis. The data collection module other games. In particular, SportSense recognized 92%
employs one thread to retrieve tweets from Twitter, de- (151 out of 163) touchdowns, 75% (51 out of 68) inter-
code and save the tweets into the MySQL database. The ceptions, 74% (84 out of 113) fumbles, and 67% (61 out
event recognition runs as another thread that retrieves of 91) field goals in these 33 games. The average delay
data from the database and analyze them for event in event recognition is about 40 seconds which is ac-
recognition. ceptable compared to a delay of about 90 seconds for the
The frontend visualizes the analysis results using ESPN web page in updating the score.
Google Charts API through a website. For ongoing
games, the website shows a color-coded bar chart for the
3. Sentiment Extraction
“hotness” of all concurrent games according to the post Our ongoing work seeks to extend SportSense for
rate of tweets related to each game. For each game, the sentiment extraction. That is, we are interested in know-
website provides three line charts that draws post rate of ing how excited game watchers are and how positive
tweets related to the game as the excitement level and their feelings are toward the game and each team of the
denotes recognized events, as shown in Figure 1. The game. We leverage techniques created by sentiment
first chart shows the excitement level of all game watch- analysis research. Our current approach encompasses
ers; the second and third charts show that of game four steps: pre-process game-related tweets, detect sen-
watchers rooted for each side of the game, respectively. timental tweets, and classify sentimental polarity. They
The line charts and recognized events for past games can are summarized in Figure 2 and described in details be-
low.
3
Sentiment Extraction Table 1: Occurrences of words indicating positive
1. Data Collection and negative sentiments in 2 million game tweets
Collect tweets in real-time using Twitter
Streaming API for 2010-2011 season NFL Word Mentioned Word Mentioned
games. Go 186,451 Lose 62,710
2. Pre-Processing Win 164,654 Beat 45,406
Remove URLs, @username, and stop words. Like 81,762 F*** 42,046
Identify all capital words, emoticons, Lol 80,703 Love 35,823
question marks, and exclamation marks. Good 64,733 Great 27,184
3. Extracting Sentimental Tweets
Produce a list of 20 frequent sentimental
words based on 50 games data. Table 2: Occurrences of POMS words
Identify the sentiment orientation of these
20 words and take them as seed words. Word Mentioned Word Mentioned
Grow the positive and negative lexicon by Good 64,733 Sad 5,215
mining synonyms and antonyms of the 20 seed
Ready 18,230 Fight 5,016
words in the WordNet.
4. Classifying Sentiment Polarity Sorry 11,096 Active 4,884
If (the tweet contains one of the senti- Blue 8,005 Angry 1,363
mental words)
Tweet orientation = word orientation
If (but clause or negation words appear) other punctuations and stop words to avoid noise. Such
Flip the tweet orientation. removal brings more benefit than harm to our analysis
Figure 2: Lexicon based Sentiment Extraction algo- because tweets are brief and usually semantically simple.
rithm. 3.3 Extracting Sentimental Tweets
3.1 Data collection To decide whether a tweet is sentimental, we study
the lexicon-based approach reported in prior work [8].
For the 2010-2011 season NFL games, we collected
The approach forms a lexicon with sentimental words
tweets during game time using the Streaming API and
and detect if any of the sentimental words appears in the
game keywords identified from the 2010 Super Bowl.
text.
We collected the tweets and their metadata such as tweet
source, created time, location, and device. These tweets First we apply information retrieval techniques to
were analyzed for event recognition in real-time through generate a list of 20 most frequent words using over two
a web service described previously. For the regular sea- million game-related tweets from 50 games played in 4
son games and playoffs, we collected more than 19 mil- weeks of the 2010-2011 NFL season. This list contains
lion game-related tweets over a period of 9 weeks in- the most frequent appeared words according to the term
cluding 100 games, from 3.5 million users. We collected frequency.
about 1 million game-related tweets from over half a We then identify the part of speech of each word,
million users for 2011 Super Bowl. The evaluation of i.e. noun, verb, adjective, etc. Previous work only con-
our solutions was performed in real-time when a game sidered adjectives, e.g., [8], and sometimes sentimental
was ongoing and was repeated with trace-based emula- verbs, e.g., [9]. We expand the sentiment verbs in sports
tion off-line if necessary. domain because action words are an essential part of
sports. The most frequently mentioned action verbs are
3.2 Pre-Processing
“go”, “win” and “beat”, etc. For example, go team A
Tweets are very noisy to extract sentiments because and beat team A emphasize sentiments toward team A
the tweets are short, informal, and ungrammatical by its through the verbs.
nature. We leveraged pre-processing procedures in [6, 7]
Finally, we produce a list of sentimental words us-
to extract the relevant signals from the tweets. First of
ing adjectives and verbs from the list of most frequent
all, URLs and replies to users (@username) often appear
words as seed. Similar to sentiment detection in product
in tweets. We remove them as being irrelevant to the
review [8], we assume synonyms share the same senti-
sentiment of a tweet. Moreover, we identify the all capi-
ment orientation while antonyms share the opposite sen-
tal words, emoticons, question marks, and exclamation
timent orientation. That means each sentimental word
marks. Capital words are commonly used in tweets to
has either positive or negative sentiment orientation. We
represent emotions; emoticons and exclamation marks
manually decide the sentiment orientation of the most
are important indicators of the presence of sentiments.
frequent appeared words and utilize a large lexical data-
We include the emoticons in the sentiments lexicon to
base, WordNet [10], to search the synonyms and anto-
determine the sentiment polarities. We label the all capi-
nyms of the seed words to grow the lexicon. In addition,
tal words and exclamation marks as emphasis of the sen-
we also include the most frequent appeared slang and
timents in the tweets. We label question marks as uncer-
abbreviations in the sentiment lexicon such as lol and
tainty about events or sentiments. Finally, we remove
wtf, etc, which are heavily used and have clear senti-
4
50% PIT GB
Positive sentiment
40%
30%
20%
10%
0%
1 61 121 181 241
Time (minute)
Figure 3: The trend of positive sentiments for both teams in the Super Bowl between Pittsburgh Steelers
(PIT) and Green Bay Packers (GB).
ments. We create our lexicon based on millions of game When we determine the sentiment orientation of a
related tweets. Therefore we believe it is accurate and tweet, we consider special conditions including but
comprehensive to determine the sentiment orientation in clause, negation, and multiple sentimental words. When
sports domain. the tweet contains the but clause which starts with “but”
In all, we extract and collect 104 sentimental words or “however”, the sentiment orientation in the clause is
(61 positive, 43 negative) and 16 emoticons. In our ex- regarded as the tweet sentiment. When the negation
periment including 50 games in 4 weeks, we find more word, i.e. “no” or “not”, appears within a threshold dis-
than one third of the tweets contain an entry from the list tance, e.g. 3 words, around the sentimental words, the
of sentimental words and, therefore, are considered sen- sentiment orientation is the opposite of its original [8].
timental. Among the sentimental tweets, 87% are posi- When multiple sentimental words appear in the tweet,
tive-only, 11% negative-only and only 2% contain both. the number of positive/negative words determines the
The top 10 frequently mentioned sentimental words are sentiment orientation. If positive or negative words dom-
either positive or negative as listed in Table 1. inate the tweet, the orientation of the tweet is regarded as
positive or negative. If the tweet contains equal number
3.4 Classifying Sentiment Polarity of sentimental words, we consider the tweet as mixed of
We are only interested in if a tweet reflects positive positive and negative sentiments. But we show that very
or negative sentiments. Positive and negative sentiments few, only 2%, sentimental tweets have mixed senti-
include positive and negative emotions, evaluations, and ments.
stances, respectively. The sentiments are obvious in most
3.5 Results
tweets. For example: “Good defence Titans!!!!!” and
“TOUCHDOWN COLTS!!” convey the positive senti- Next, we demonstrate our sentiments extraction re-
ments. In contrary, “The Titans are undisciplined and sults. Since positive sentiment is the dominant polarity
bad on Defence” expresses the negative sentiments di- which weighs almost 90% of sentimental tweets, we
rectly. present the results using the percentage of positive sen-
timents in each minute. However, we still count the neg-
The primary reason that we only consider positive
ative sentiment. When we calculate the percentage of
and negative sentiments is that the most frequent ap-
positive sentiment, we first subtract the number of nega-
peared sentimental words are either positive or negative.
tive tweets from the number of positive tweets. Since the
An apparent alternative to our positive-negative catego-
number of the negative sentiment is stable and infre-
rization is POMS, a well-established psychometric in-
quent, the negative sentiment does not impact the gen-
strument. POMS (profile of mood states) assessment is a
eral trend of the sentiment.
factor-analytically derived inventory that measures six
identifiable mood or affective states. For example, Bol- We take the results of 2011 Super Bowl, which is
len et al [18] discovered the correlation between public shown in Figure 3, as an example to examine the effec-
mood and events in the social political, cultural and eco- tiveness of our approach. In the beginning of the game,
nomic sphere by applying POMS to Twitter analysis. We the percentages of positive sentiment of both teams are
examine the occurrences of POMS terms in the same 50 about the same. Along with the game progress, the
games that we extract positive and negative lexicons. trends of positive sentiment start to fluctuate. Note the
Although POMS contains 6 factors that may provide trends of two teams tend to fluctuate apart when the
more information about sentiments, we found that the events take place because the events impact oppositely
POMS terms are rarely mentioned comparing to other on fans‟ sentiments of two teams. More interestingly, the
frequent sentimental words except “Good” (see Table 2). percentage of positive sentiment towards the scoring or
the winning team will not increase a lot but percentage
5
towards the opponent will drop considerably. It is be-
cause the number of tweets related to the scoring or win-
ning team also increases significantly that stabilizes the
percentage. When the game is over, it is not surprise that
the positive sentiment of the winning team dominates.
We obtained the similar results in another 30 games
in the NFL regular season and playoffs. Thus, we con-
clude that our approach is adequate to extract the senti-
ments, especially when the events happen.
4. Twitter-Enabled Applications
Program event recognized and sentiments of TV
watchers extracted by SportSense can be used in various
ways to enhance the TV watching experience. We next
discuss a few from our ongoing work. Figure 4: Screen snapshot of e*PG
4.1 Personalized EPG Overlays (e*PG) ing that to specific information that can be associated
Users consume large amounts of video content, with programming is a useful one due to the fact that
ranging from broadcast TV to Internet video sources. there is a lot of information that is not captured by tradi-
Along with the increasing selection of content, discovery tional metadata services (EPG being a concrete exam-
of that content is becoming a pain point for users. The ple). Tweets become the direct information reflecting the
general problem is given a dilemma of selecting a pro- game events and audience sentiments. In addition, Twit-
gram (channel) among a large number of live programs, ter users spread out all over the broadcasting area such
how does the user decide what is worth watching? This that the result is more representative.
problem is particularly prevalent in the sports domain
4.2 Dual-Screen Advertising
where it is not uncommon for fans to track multiple
games that are occurring simultaneously. For example, Any additional knowledge regarding TV watchers
up to 10 games are played on every Sunday afternoon at (a.k.a. audience measurement) is important to advertis-
the same time during the NFL regular season. Further- ers, networks and stations. Advertisers want to know the
more, generically switching between games may cause size and characteristics of the audience they are reaching
users to miss particular moments of interest. Even when they purchase ads on a particular program. The
though static EPG information can provide a certain lev- network or station cares about audience size and compo-
el of information in regard to programming, this doesn‟t sition because they determine the amount it can charge
apply to live programming where events of interest are for commercial time [11]. Methods such as program-
dynamic and vary among users. ming rating and commercial rating are applied to meas-
ure TV and commercial audience. They also require de-
Given the tools developed for SportSense, we utilize
ploying people meter to collect measuring samples [11].
the real-time excitement level of each program to enable
Not only are such methods expensive, but also they rate
users to switch to the most exciting ongoing program
a large chunk of a program or usually an entire program
(game). Existing work has applied visual, audio or
because commercials are delivered in the pre-planned
webcast analysis to learn about the program content and
break time of a program. As more and more commer-
its excitement level. Our work opens a novel, orthogonal
cials are embedded during the program without a break,
direction by utilizing social data to infer about the audi-
there is a need to determine how engaged the audience is
ence reaction in real-time and in turn augment that in-
at a much finer grain. Because many TV watchers tweet
formation in the form of overlays for traditional EPG as
while watching TV, we can estimate the audience ex-
shown in Figure 4. These overlays are generated using a
citement level by analyzing tweets from TV watchers in
combination of sentiment extraction and event recogni-
real-time. The analysis not only tells us how engaged the
tion from the collected Twitter data (which is performed
audience is overall but also tell which sector of the audi-
in real-time) as well as external information about the
ence is more engaged. For example, in a sports game, we
specific program/user as it is available. For example, if
are able to tell what events are taking place and which
it‟s known who my favourite teams/players are, the over-
side of fans is exciting. Advertisers can bid on the time
lays can indicate when any of them are participating in
slots based on the sentiments of their target customers.
active games being displayed on the EPG. Furthermore,
we can utilize general excitement levels on Twitter to Given the fact that we can utilize Twitter data to ex-
indicate if games are „interesting‟ – i.e. in overtime, or tract sentiment/events about programming that is being
the scores are close. The general notion of mining Twit- consumed by users, we applied this information as part
ter for audience sentiment around TV and then translat- of a framework for coordinating concurrent dual-screen
6
100% 50%
Inter Bayern UCONN BUTLER
Postitive sentiment
Postitive sentiments
80% 40%
60% 30%
40% 20%
20% 10%
0% 0%
1 60 119 1 60 119
Time (minute) Time (minute)
Figure 5: The trend of positive sentiments for both Figure 6: The trends of positive sentiment for both
teams in a UEFA Champions League game between teams in the NCAA men’s basketball tournament
Inter Milan and Bayern Munich. final between Univ. of Connecticut and Butler Univ.
campaigns for advertising. In a world where prolifera- Figure 5 shows the sentiment trends extracted our
tion of second screens is increasing substantially and real-time Twitter analysis for both teams in the Champi-
where a brand advertiser still has a sizable investment in ons League game. The only goal from Bayern Munich
the primary screen, e.g., TV, and uncontrolled second happened at the last minute of the game. Therefore, the
screen can be a liability or distraction. Our system for sentiments regarding to Bayern Munich became ex-
synchronized, dual-screen advertising offers TV adver- tremely positive in the end. Although there was no goal
tisers brand and message protection through cross-screen in other part of the game, events such as yellow card or
context sharing & content coordination. Combining the goal opportunity could impact the sentiments. Figure 6
data from aggregated social networks like Twitter, ana- illustrates the sentiments of both teams in the NCAA
lyzing and extracting things like sentiment and events men‟s basketball tournament final. Since scoring events
allows us to monetize any TV-related context as a bid- are common in the basketball game, the sentiments are
dable keyword for advertisers. For example Goldman & changing smoothly during this game. When the Univer-
Sachs can synchronize the second screen experience, sity of Connecticut kept leading, its sentiments became
e.g., tablets and smartphones, with an event correspond- more positive until the end of the game.
ing to the primary screen, e.g., touchdown = suc- Although these two games are insufficient to prove
cess+fame, thus playing an ad on the companion device the effectiveness of our approach on soccer and basket-
when such an event has occurred in the game being ball, the results illustrate the feasibility of applying our
watched on the primary screen. The system also enables approach on various sports games.
advertisers to prevent things like brand dissonance be-
tween screens and provide brand/message protection. 6. Related Work
The system architecture provides enhancements to a Several concurrent projects also study tweets about
traditional campaign manager to support the extra intra- sports games, however they do not provide real-time
program information gleaned from Twitter, as well as a event detection. Hannon et al [12] used post rate of
messaging and notification system to execute second- tweets to produce video highlights of the World Cup off-
screen campaigns on desired devices such as tablets or line. They did not recognize game events nor did they
smartphones. produce highlights in real-time. Chakrabarti and Punera
5. Other Sports Games [13] assumed that a game event is already recognized
and focused on describing the event using Hidden Mar-
As we have already shown the effectiveness of the kov Models trained with tweets collected from events
sentiments extraction on the NFL games, we next briefly happened in the past. Therefore, our focus on real-time
demonstrate the generality of our approach on various event recognition is complementary, and addresses a
types of sports. We apply our approach on two of the more difficult and fundamental problem.
most popular sports, soccer and basketball as well. We
Existing works on sentiments measurement and
choose one game from UEFA (the Union of European
opinion detection focus on product review and tweets
Football Associations) Champions League between Inter
moods modelling. Hu and Liu [8] mined and summarize
Milan and Bayern Munich, and one game from NCAA
the customer reviews of a product. Pang et al [14] and
(National Collegiate Athletic Association) men‟s basket-
Zhuang [15] focus on sentiments classification in movie
ball tournament between University of Connecticut and
reviews. Jansen et al [16] investigated Twitter as a form
Butler University. Both games are broadcasted on TV
of electronic word-of-mouth for sharing consumer opin-
and have a large number of watchers.
ions concerning brands. Extracting sentiments from sport
7
games-related tweets are significantly different from [4] S. Zhao, L. Zhong, J. Wickramasuriya, and V.
product or movie reviews, because reviews have formal Vasudevan, "Human as Real-Time Sensors of Social
format and rich context information but tweets are collo- and Physical Events: A Case Study of Twitter and
quial without context. Furthermore, as people are emo- Sports Games," Technical Report, Rice University,
2011.
tional during sports games, their sentimental expressions [5] SportSense, "http://sportsense.us."
are diverse and unexpected. [6] V. Pandey and C. Iyer, "Sentiment Analysis of Mi-
Bollen et al. [17] modelled public mood and emo- croblogs."
tion according to people‟s Twitter posts. Pandey and [7] A. Pak and P. Paroubek, "Twitter as a corpus for
Iyer, Barbosa and Feng [18] proposed machine learning sentiment analysis and opinion mining," in Proc. of
approaches to classify sentiments on tweets. They focus LREC 2010.
[8] M. Hu and B. Liu, "Mining and summarizing cus-
on tweets with certain expressions over a long time, e.g.
tomer reviews," in Proc. ACM SIGKDD, 2004.
one year in [17]. For sentiment extraction from game- [9] T. Nasukawa and J. Yi, "Sentiment analysis: captur-
related tweets, we must extract sentiments in real-time ing favorability using natural language processing,"
without the expression or structures leveraged by [17]. in Proc. ACM K-CAP, 2003.
[10] G. A. Miller, "WordNet: a lexical database for Eng-
7. Conclusion lish," in Commun. ACM. vol. 38, 1995, pp. 39-41.
In this work, we described our ongoing effort in ex- [11] M. B. George Belch, Advertising and Promotion: An
tracting real-time audience sentiments by analyzing Integrated Marketing Communications Perspective:
Twitter. We showed that the limited vocabulary of sports McGraw-Hill Irwin, 2009.
games makes lexicon-based analysis methods highly [12] J. Hannon, K. McCarthy, J. Lynch, and B. Smyth,
"Personalized and automatic social summarization of
effective. For several major sports games, including the
events in video," in Proc. ACM IUI, 2011.
US NFL, UEFA Champions League, and NCAA basket- [13] D. Chakrabarti and K. Punera, "Event Summariza-
ball tournament, SportSense is able to not only recognize tion using Tweets," in Proc. AAAI ICWSM, 2011.
major game events in real-time but also capture the sen- [14] B. Pang, L. Lee, and S. Vaithyanathan, "Thumbs
timent toward each side of a game during the game. We up?: sentiment classification using machine learning
described our ongoing work in leveraging the real-time techniques," in Proc. ACM EMNLP ACL-02 Volume
event recognition and sentiment extraction of SportSense 10, 2002.
for a social TV system with socially informed electronic [15] L. Zhuang, F. Jing, and X.-Y. Zhu, "Movie review
program guide (e*PG) and dual-screen advertising. Our mining and summarization," in Proc. ACM CIKM,
2006.
work demonstrates the potential of Twitter as a key in-
[16] B. J. Jansen, M. Zhang, K. Sobel, and A. Chowdury,
formation source toward social TV systems. "Twitter power: Tweets as electronic word of
8. References mouth," Journal of the American Society for Infor-
mation Science and Technology, vol. 60, pp. 2169-
[1] L. He, E. Sanocki, A. Gupta, and J. Grudin, "Auto- 2188, 2009.
summarization of audio-video presentations," in [17] J. Bollen, A. Pepe, and H. Mao, "Modeling public
Proc. ACM Multimedia, 1999. mood and emotion: Twitter sentiment and socio-
[2] S. Nepal, U. Srinivasan, and G. Reynolds, "Automat- economic phenomena," in Proc. AAAI ICWSM, 2011.
ic detection of 'Goal' segments in basketball videos," [18] L. Barbosa and J. Feng, "Robust Sentiment Detection
in Proc. ACM Multimedia, 2001. on Twitter from Biased and Noisy Data," in Proc.
[3] M. Shiels, "Twitter co-founder Jack Dorsey rejoin ACM COLING, 2010.
company, http://www.bbc.co.uk/news/business-
12889048," 2011.
8