Discovering NBA Game Stories from Twitter

DEIM Forum 2014 B6-1
Discovering NBA Game Stories from Twitter
Zhichao ZHANG† , Hisashi KOGA† , and Youhei OGYU†
† Graduate School of Information Systems, University of Electro–Communications 1–5–1, Chofugaoka,
Chofu-shi, Tokyo, 182–8585 Japan
E-mail: †{zhang,koga,ogyu}
Abstract This paper proposes a system which generates the game summary for basketball games by quoting the
representative tweets during the game period for the first time, whereas the previous similar researches treated
soccer and American football. Basketball is a more challenging subject than football or or soccer for making a
summary. This paper introduces a unique idea to put the discussions at the break times between quarters and at
the end of the game into the game summary, since spectators state the game status at that moment and comments
upon the activities of the players and the teams then. Without specifying keywords, our heuristics attempt to such
discussions as the peaks of long duration in the tweet volume graph. Experimentally, our summary covers up to
87% of the items written in the NBA Official Game Summary.
Key words Twitter Mining, Sport Event, Summary Generation
1. Introduction
than that of soccer game. The score of a basketball game
can be 110-100, while the score of a soccer game can be
Twitter has become the most popular micro-blog which
only 1-1. Therefore, one shoot/goal is less important for
has more than 0.5 billion users. A lot of tweets are pub-
basketball than for soccer. This paper introduces a unique
lished every day in Twitter, containing a myriad of infor-
idea to generate the game summary effectively for basketball
mation about what the users are doing and watching, which
games, while paying attention to the increase of the tweet
can be seen as describing opinions about various events. Re-
volume in the same way as the previous researches. After
cently, mining information about such events from the twit-
discovering the important moments from the tweet volume
ter stream has been a significant research topic. This re-
graph which records the number of tweets per time unit, our
search approach is categorized into two kinds: the first kind
method yields the game story by choosing the representative
attempts to discover the event occurrence without knowing
tweets for each important moment. Notably, our method
event types [2], while the second one intends to obtain the de-
works in an almost unsupervised way without requiring the
tailed descriptions about a specific event like an earthquake,
event keywords specific to basketball games such as dunk, 3-
a typhoon etc. [1] from the twitter stream and to summarize
pointer and so on. Only the two team names associated with
the game have to be given in order to extract the relevant
In the line of the second kind, some previous litterateurs
set of tweets from the twitter stream.
dealt with sport games. Chakrabarti et al. [3] and Sport-
The rest of the paper is organized as follows: In Sect. 2,
Sense [4] studied American football, while [5] examined the
we brief the related works which focus on the event detection
soccer’s world cup. All these methods identify the important
from the twitter stream. Section 3 analyzes the character-
moments in sport games by picking up the time instances
istics of the tweets regarding to the NBA games. Section
when the tweet volume per time unit rapidly increases. For
4 describes our summary generation system for the NBA
the sport games, a remarkable event causes a sudden increase
games. Section 5 reports the experimental results. Finally,
of the tweet volume, because many Twitter users comment
Section 6 concludes this paper and discusses the future work.
on it. Basically, important events in sport games can be discovered by checking if the tweet volume suddenly increases.
2. Related Works
Our research purposes to generate the summary for bas-
This section briefly refers to the previous researches which
ketball games from the tweets regarding to the NBA (Na-
attempted to detect events for sport games from the twitter
tional Basketball Association) games. Basketball has differ-
ent properties from American football and soccer. For ex-
Zhao et al. [4] developed a system named Sportsense which
ample, the goal frequency of basketball game is much higher
displays the major events rates fans’ excitement level in the
middle of the NFL American football games. They first build
event templates by learning event examples which shows the
change of the tweet volume when the events happen. Once
the event templates are complete, the same type of events
can be detected on-line for the ongoing game by matching to
the event templates. However, this event detection method
requires supervision, since the keywords related to events, for
example ”touchdown” or ”TD” for short, must be specified
so as to collect the event examples. The researchers must
have the domain knowledge about the NFL games in order
to predetermine the event keywords. Chakrabarti et al. [3]
use Twitter to generate summaries of long running, structure rich events under the circumstances that multiple event
instances share the same underlying structure. Specifically,
they learned the structure and the vocabulary of events for
American football with a modified Hidden Markov Model
Figure 1 the Start, Peak, and End Times of a Spike
equals its word frequency. Finally, the top N sentences that
do not share any non-stop word stemmed tokens are output
as the game summary. For American football game, [5] sets
N to 3, which means that only a few sentences are enough
to cover the contents of an important moment.
3. Tweets for Basketball Games
(HMM). Here the tweets for many games need to be prepared to compose the learning data. Moreover, the learning
process for the HMM is time-consuming.
On the other hand, Nichols et al. [5] detected events and
generated a journalistic summary from the tweets at a World
Cup soccer game.
They neither count on predetermined
event keywords nor learn from the tweets for multiple games.
Namely, their method works in an unsupervised way. We explain their method in details below, since we will extend it
for basketball games in this paper.
After obtaining the tweets regarding to a given game using basic keyword filtering via the twitter API, they first
draw a tweet volume graph whose x-axis presents the time
in minute and the y-axis denotes the number of tweets per
minute. From this graph, they extract spikes each of which is
defined by the triple <Start Time, Peak Time, End Time>
as shown in Fig. 1. The tweet volume starts increasing at
the start time, reaches the peak at the peak time and stops
decreasing at the end time. Among the derived spikes, only
those whose slope between the start time and the peak time
goes beyond a threshold are memorized as the important
moments M1 , M2 , · · · , Mm , where m denotes the number of
important moments. In [5], the threshold is empirically set
to 3 times as large as the median of all the slopes for American football. Similar approaches are utilized by some other
previous researches [6] [7] [8].
After the important moments in the game are identified,
the game summary is constructed by selecting the N representative sentences from the set of tweets posted at each im<j<
portant moment Mj (1 =
= m). In this process, first, the
This section first explains how to collect the tweet dataset
regarding to the NBA games and then describe their features.
3. 1 Collection of Tweets
We rely on the Twitter Streaming API (https://dev.twitter
.com/docs/streaming-apis) to gather the tweets on the NBA
games. This service allows developers to pull tweets in realtime which contain specified keywords. In our case, we collect the tweets for a certain NBA game between two teams
by setting their team names as the keywords. For example,
we can collect the tweets for the game between the Miami
Heat and San Antonio Spurs, by specifying the set of hashtags: ”#Heat”,”#Spurs”, ”#MIA” and ”#SAS” as the
keywords. Here ”MIA” and ”SAS” are abbreviation of the
two team names. We get the team names and their abbreviations from the section of ”teams” from the NBA official
website ( In order to get the tweets at the
game period, we keep only the tweets with the timestamp
between the begin time and the end time of the game, which
we also learn from the official website. Our dataset consists
of tweets for 30 games from the regular season, playoffs and
the finals of NBA 2012-2013 season. We show two examples
of the tweets below.
@TrappedlnThe225 - Down by 6 still in the game. #Heat
null Sat May 25 10:01:36 JST 2013
@dudeimspacely - #Pacers 28#Heat 22 End of lst Quarter. #NBAPlayoffs null Sat May 25 10:01:36 JST 2013
3. 2 Feature of the Tweets for NBA Games
Figure 2 illustrates the tweet volume graph which records
the number of tweets per minute for the game between Lak-
longest sentence of each tweet included in Mj is abstracted.
ers and Spurs in Apr. 27th, 2013. After examining several
Let the set of such longest sentences for Mj be Lj . Then,
tweet volume graphs for different NBA games manually, we
the score of a sentence in Lj is computed by summing up the
have noticed the next primary features of the tweet dataset
scores of its word tokens. Here, the score of a word token
for the NBA games.
information is missed.
To contain such game status information, this paper
uniquely pays attention to the discussions at the break times
between quarters and at the end of the game, which always
contains important information such as the game status and
the comments on the previous quarter. For instance, during
the halftime of the NBA games, spectators give their impression of the first half and, therefore, we should not neglect
the tweets published then so as to obtain useful informaFigure 2 Tweet Volume Graph for a Basketball Game
tion. Note that the tweets in the break times are ignored
or slighted in the previous researches for soccer and football.
( 1 ) The graph fluctuates more often in the NBA games
than in the games of NFL and soccer, which implies that it
contains more spikes. This is probably because the basketball games are accompanied by more continuous actions than
American football and soccer. We also noticed that spikes
We show an example of the tweet issued at a break time below. This tweet was posted at the break time after the 3rd
quarter of a game between the Indiana Pacers and the Miami
Heat. It surely exhibits the game status at that moment and
is significant.
with sharp slopes tend to be related to real-time actions such
as splendid slum dunks and turnovers.
( 2 ) The contents of the tweets contained in one spike
are very diverse for the NBA games, since multiple events
such as slum dunk, assist, three pointer, game winner etc.
can happen in just one minute.
Our system in the next section exploits the two above features to generate an attractive game story for an NBA game.
4. Our Event Detection System
At the end of the 3rd, the Pacers are ahead of the
Heat by 13. Hibbert, George lead with 22 points
Unfortunately, the discussions at the break times cause
only gentle excitement and result in the peaks with gentle slopes in the tweet volume graph. Then, how can we
find such discussions from the graph without knowing specific keywords? In this paper, we propose a simple efficient
heuristic approach to regard the spikes having long duration
The purpose of this system is discovering useful information as much as possible, and displaying them to the NBA
fans as the games story. So the input of our system is the
stream of tweets regarding to a certain NBA game, and the
output is the summary, i.e. the story of this game. We develop this system by tailoring the method of Nichols et al. [5]
for the basketball games. It operates the next three steps in
order so as to output the final games story:
( 1 ) Our system first determines the important moments
by choosing some of the spikes in the tweet volume graph.
( 2 ) Next, the tweets that can describe the contents of
as the discussions between the break times. The rationale
of this idea is as follows. At the beginning of a break time,
people start talking about the previous quarter and a uphill
slope is formed. Then, since they gradually leave from their
PCs or smart phones, a downhill slope is formed. Interestingly, this downhill slope becomes a very long tail, as it is
never interrupted until the next quarter starts.
We take both the slope and the duration of a spike into
consideration. Particularly, we evaluate the value of a spike
P according to Eq. (1). A spike is evaluated higher, as this
formula becomes larger for the spike.
the important moments are chosen from the set of tweets
published at the important moments.
( 3 ) The final game summary is derived by excluding the
similar tweets from the tweets chosen at the previous step.
We explain the three steps from now on.
4. 1 Decision of Important Moments
Like [5], the important moments are derived by searching
peculiar spikes from the tweet volume graph. Although [5]
seek spikes whose slopes are steeper than a certain threshold
as the important moments for soccer, considering spikes with
sharp slopes is insufficient for basketball, because they are
related to the real-time actions such as beautiful slum dunks
as pointed out in Sect. 3.2, Thus, the general game status
Score(P ) = fs ×
Slope(P )
Area(P )
+ fa ×
fs and fa are weighting parameters to control the contribution of slope and area. We currently set both fs and fa to
0.5. MaxSlope denotes the biggest slope and MaxArea is the
biggest area size of the peaks over the whole tweet volume
graph. Though the area size of a spike appears in Eq.(1)
instead of the spike duration, be aware that the area size of
the spike is roughly proportional to its duration. In the same
way as [5], we choose the peaks with the score higher than a
threshold θ as the important moments. In the experiments
at Sect. 5, θ is set to 0.1.
4. 2 Selection of Tweets for Important Moments
(4) Harden with the airball with 53 seconds left. Did I say
Next, for an important moment, we select the set of tweets
soft? I meant flaccid. #Rockets
which can describe its contents. In this process, a tweet is
(5) James Harden picks up his 5th foul, he will sub out
ranked according to the relevance of words that the tweet
This example shows that the diversity of the top 10 key-
contains. Here, the relevance of a word is determined by its
words is high. The top 10 keywords are divided into several
frequency in the group of tweets that belong to the impor-
types: (1) players’ names, (2) event names and (3) general
tant moment. In counting the word frequency, we exclude
information words. Despite only the 10 words are considered
the English stop words by utilizing English stop-word dic-
in selecting tweets for this important moment, the contents
tionaries opened to the public on the web. In addition, the
of the chosen tweets have high diversity. We observe the
two team names which compete in the concerned NBA game
similar tendency for other important moment examples. We
are also discarded. Because the team names are used as the
guess the reason of this phenomenon as follows.
filtering keywords for the Twitter Streaming API, they are
( 1 ) On condition that one event is associated with one
contained in almost all the tweets. Thus the team names
player, the top 10 keywords for one important moment in-
cannot describe the contents of the important moments well.
cludes several player’s names for the most cases. Thus, mul-
After calculating the frequency of all the words over the
tiple events are discovered with the top 10 keywords.
tweets belonging to the important moment, we get the top
( 2 ) One word out of the top 10 keywords may corre-
K words which have the highest word frequency at the im-
spond to multiple events. For example, a player’s name can
portant moment. In our current implementation, K = 10.
be related to several different events.
Next, the score of a tweet is computed. We regard a tweet
( 3 ) Even for the identical event, multiple tweets describ-
is more important as it contains more highly-ranked key-
ing it can supply different information, while they contain the
words. The score of a tweet t is denoted by V (t) in Eq.
common keywords.
4. 3 Removal of Similar Tweets
V (t) =
The similar tweets are excluded from the tweets chosen at
the previous step. Removal of the similar tweets is necessary here, since a lot of spectators issue very similar tweets
Here, n is the number of the top K words in t, and score(i)
on the same event and displaying such similar tweets annoys
presents the value of the top i-th word wi . For K = 10,
the NBA fans.
score(i) is set to 20 − i. Hence, the word with higher fre-
This step first uses clustering to classifying the set of tweets
quency has a higher score. In particular, the top keyword
into several clusters of similar tweets and then outputs one
is assigned about twice as large score as the 10th keyword.
representative tweet per cluster. As a clustering algorithm,
Finally, the set of tweets t for which V (t) becomes greater
we use average linkage method, one of the well-known ag-
than some threshold τ are passed to subsequent processing
glomerative hierarchical clustering algorithms. The agglom-
discussed in Sect 4.3.
erative hierarchical clustering begins with one-point clusters
To confirm if we can successfully get the relevant tweets us-
and recursively merges the most similar pair of clusters, until
ing the top 10 keywords, and if the diversity of these tweets
the number of clusters finally reduces to one. In the agglom-
is high, we preliminarily apply our method to several im-
eration step, the clustering algorithm searches the closest
portant moments. Here we report one case example for one
pair of clusters and merges them into a new single cluster.
important moment of the game between Houston Rockets
The hierarchical clustering algorithm is advantageous in that
and Oklahoma City Thunder on Apr. 30th, 2013. For this
the number of clusters do not have to be specified a priori.
case, the top 10 keywords are as follows:
This nice feature is suitable for our case, since a single impor-
durant, kevin, left, dunk, seconds, pointer, lead, driving,
harden, cuts.
The five tweets with the highest score for this important
moment is shown below.
tant moment contains multiple events for basketball games
as stated in Sect. 3.2, so that it is impossible to grasp the
proper number of tweet clusters beforehand.
In the average linkage method, the distance between two
(1) @okcthunder: Kevin Durant with a 3-pointer and a
clusters is defined as the average distance between any mem-
driving dunk in 29 seconds. #Thunder cuts #Rockets lead
ber (tweet in our case) of one cluster to any member of the
to 2. 105-103. 1:13 left in Gam
other cluster. Here, the distance D(ti , tj ) between a pair of
(2) harden missed 3-pointer clutch shots in a row. watch
tweets ti and tj is defined as the Jaccard distance in Eq. (3).
durant pull up and win it with a 3 at the buzzer. #Thunder
(3) James harden blowing the games for #ROCKETS
D(ti , tj ) = 1 −
|Si ∩ Sj |
|Si ∪ Sj |
where Si and Sj symbolize the set of the words in ti and tj
the Indiana Pacers on June 2nd, 2013. First, the ground
respectively. The Jaccard distance is derived by subtracting
truth important moments discovered manually are shown on
the Jaccard coefficient from 1. The Jaccard coefficient be-
the tweet volume graph for this game in Fig. 3. There, the
tween two sets A and B is defined as
and measures the
extent of the overlap between them.
spikes surrounded by the green rectangle correspond to the
ground truth important moments. To see what kind of infor-
If we stop merging clusters before the cluster number de-
mation the ground truth important moments include, Table
creases to 1, multiple clusters are extracted. Our implemen-
1 lists a typical tweet example chosen by us for the 14 im-
tation ceases merging clusters, when the distance between
portant moments. Among them, IM1, IM5, IM9 and IM14
the two clusters to be merged exceeds a threshold value
correspond to the discussions at the break times between
D = 0.925. Since we determine this value of D only empiri-
quarters and at the end of the game.
cally, the algorithm to derive an optimal value of D remains
to be developed in future.
After having multiple clusters in the above way, we determine one representative tweet for every cluster with more
than 3 members. Namely, small clusters are not adopted,
since they are not admitted by many spectators. Consider a
cluster C consisting of m tweets (m > 3). A tweet in C which
is the most similar to other tweets in the same cluster is appointed to the representative of C. Here, the similarity of a
tweet ti in C to other tweets is measured by
Figure 3 Ground Truth Detected Manually
|Si ∩ Sj |
j=1,j =i
|Si ∪ Sj |
which sums up the Jaccard coefficients between ti and all the
other tweets tj in C.
Table 1 Examples of Tweets for each Important Moment
Paul George’s dunk on bosh #nasty #heat vs pacers
At the end of the first quarter, the #Pacers trail the
Finally, the representative tweets of all the clusters for all
the important moments constitute our game story.
Heat 23-21
are shooting lightsout at 85% (6-7) and #Pacers are
5. Experiments
With the dataset in Sect. 3, we experimentally evaluate
the performance of our system.
5. 1 Accuracy of Detected Important Moments
shooting 50% (1-2). #Heat lead 25-24
2 missed dunks #pacers ouch
D-Wade steal, ends up with a #LeBron JAMI #Heat
lead at the half 40-39 0ver the #pacers
One of the novel ideas in our method is that it takes both
nals#heat #game6
the slope only. This policy is abbreviated as SL hereafter.
#Pacers are dominating the boards. They’ve got 43
rebs (13 0ff 30 def) vs 28 rebs (10 0ff 18 def) for #Heat.
The accuracy of an algorithm is measured by how many percentages of the real important moments, i.e., the spikes with
The pacers are playing so good defensively. Let’s go
Heatl!! Let’s go Heat!!! Letis go Heat!! #eastern fi-
by comparing it with another policy which only considers
SL is derived by setting fs = 1.0 and fa = 0.0 in Eq. (1)
#pacers taking an 11 point lead against the #heats on
the 3rd quarter #awesome #enjoyingthegame #keep-
the slope and the area of the spikes into account to detect the
important moments. Therefore, we evaluate our algorithm
Ridiculous 3 point FGyo early in this game. #Heat
Lance Stephenson leads w/ 11
useful information are identified. We manually derive the
At the end of the 3rd, the #Pacers are ahead of the
Heat by 13. Hibbert, George lead with 22 points each
real important moments for all the 30 games from their tweet
Three pointer! lets go miller #heat
volume graphs as the ground truth.
Intense #heat and #pacers!! win or go home game
Let’s go #Heat#NBA #HEATNATION
As the result, 14.86 genuine important moments are discovered manually per game on average. On average, our
tantrum and likely closing the door on a comeback.
algorithm finds 13.23 important moments contained in the
ground truth, while SL finds 10.56. Thus, the recall of our
#heat #pacers
algorithm reaches 89% whereas that of the SL is 71%. Thus,
To be more comprehensive, we give an explanation with
respect to one game instance between the Miami Heat and
#Pacers go on a 9-0 to restore order and now lead 81-6
w/ 3:55 left. Roy Hibbert 24pts, 9rebs, Paul George
our algorithm can detect important moments more accurately than SL.
LeBron James just got a technical for throwing a
25pts 8 rebs. #heatvpacers
It’s going to game 7 #Pacers
Our algorithm detects 13 ground truth important mo-
Table 2 Sentences in the Official Game Story and the Matched
Tweets in our Summary
ments. It misses one true important moment, that is, IM10
Official Game Story
that describes an important three pointer by the Miami
Heat’s player Mike Miller. Note that all the discussions in
Miami Heat 91-77 in Game 77
the break times are found by our algorithm. On the other
hand, SL discovers 10 important moments of which 9 match
to the ground truth. It misses IM1, IM8, IM9 IM10 and
The Pacers limited Miami
to 36-percent shooting and
IM12. Importantly, IM9 which describes the summary of
the third quarter and corresponds to the break time is not
Our Summary
The pacers defeated the #Pacers win # Heat by 91-
dominated things inside
The Pacers won the re- Rebound 53-33, score in
recognized by SL considering the slope only. In this way, SL
bounding battle 53-33 and the paint 44-22, this the
leaks more meaningful information than our algorithm.
outscoring the Heat 44-22 reason why the #Pacers
5. 2 Quality of Our Summary
The design purpose of our system is to describe the NBA
in the paint
with 7 rebounds and 6 as- and 6 assists in game 6
game story by exploiting the useful information contained in
the representative tweets. To examine if our system fulfills
won this game
LeBron James scored 29 LBJ got 29 pts, 7 rebounds
In quarter 3, the Pacers At the end of the 3rd, the
this goal, we evaluate the representative tweets outputted
outscored Miami 29-15, in- #Pacers are ahead of the
by our system. In particular, we compare our summary that
cluding a 12-0 burst early Heat by 13.IND 68, MIA
consists of the chosen tweets with the NBA official Game
in the quarter , took a 55
68-55 lead into the fourth
Story and examine how our summary agrees with it. The
NBA official Game Story is a detailed game summary writ-
ten by the editors of the NBA official web site which contains
shots, both 3-pointers, for changed #heat
the meaningful information about the games, such as game
status, representative events, good move, bad move, quota-
Miller made his only two Mike miller is a game
the Heat.
Had a putback dunk and LeBron James just got a
tions from player interviews telling the state of the game and
then Miami unraveled com- technical for throwing a
so on.
pletely. Called for an offen- tantrum
sive foul
As an example for one game, Table 2 shows the 13 items
described in the official Game story and the matched tweets
ana’s lead to 81-68 with up 81-68.
in our summary side by side. This game was played between
the Miami Heat and the Indiana Pacers on June 2nd. Empty
entries in the table mean that our summary misses the item
corresponding to the row. Since there exist two empty entries, our summary covers the 11 items. Therefore the coverage rate of our summary is
= 84.6%. Whereas our
summary cannot cover all the items described in the NBA
official Game Story, it succeeds in acquiring useful information not written there. Table 3 displays the examples of the
tweets with useful information that is not stated in the NBA
George Hil pushing Indi- 4 mins left and #pacers are
3:55 remaining.
Dwayne Wade and Chris D.Wade and C.Bosh toBosh combined for 15 points tally got 15 points
on 4-for-16 shooting.
10 West missed his first seven Wake up, Mr. West!!!
shots and finished 5-for-14
for Indiana.
11 Joel Anthony came back
into the rotation
12 Paul George (28 points) George 28 pts, 8 rebound
and Roy Hibbert (24)
Official Game Story, but contained in our summary. These
and 5 assists, Roy Hibbert
20 pts, 11 rebounds.
tweets are either more detailed descriptions about the game
13 Miami was without for- This game just shows how
than the NBA official Game Story or comments on the teams
ward Chris ”Birdman” An- Miami would have lost the
and the players.
last game if Birdman got
Figure 4 summarizes the coverage rate for multiple games,
i.e., the 7 games of the NBA eastern semi-final between the
Finally, we show that our approach to compile the rep-
Heat and the Pacers. Here, the x-axis denotes the game ID
resentative tweets for the clusters produced by the hierar-
and the y-axis shows the coverage rate against the NBA Of-
chical clustering algorithm is effective to make the summary
ficial Games Story. The mean coverage rate reaches up to
diverse. The tweets with respect to the NBA games are clas-
87%. Since our summary mines a lot of useful information
sified into the following 6 types:
not contained in the NBA Official Game Story, we consider
that this coverage rate is acceptable.
( 1 ) General status: showing the score of the game. For
example, the scores of the two teams at half time and at the
Table 3 Tweets with Useful Information that is not Stated in the
NBA Official Game Story
That dunk by Paul George over bosh!!! #pacers
Led by @PaulGeorge24’s 9pts/4rebs, the #Pacers trail
types. Table 4 exhibits the tweet examples in our summary
for each information type. On the other hand, when we chose
IND 21
randomly the same number of tweets from the whole set of
#Pacers only down 2 thru the 1st. Despite #Heat go-
tweets associated with IM5, only 2 information types were
ing 6-7 from downtown. #GoPacers #HEATvsPAC-
covered, because too personal tweets were put into the sum-
ERS #NBAPlayoffs
mary. The above fact provides some support for the claim
At least 15 points left on the board due to missed
8-1 free throw and 10-3 foul advantage for the #pacers
in the first half. Where are all the people claiming the
nba wants the #Heat to win
at the halftime, our summary covered all the 6 information
the Heat by 2 at the end of the 1st quarter.?MIA 23,
dunks layups #Pacers
As for IM5 in Table 1 which corresponds to the discussion
Hibbert with an easy layup, #pacers lead the #Heat
that the agglomerative hierarchical clustering helps augment
the diversity of our summary.
Table 4 Types of the Representative Tweets for IM5
Type of information
General status
51-42 midway thru the 3rd quarter
A tale of two 3rd quarters in game 5 Miami won the
Detailed information
Paul George Wt 3 t put #pacers 7 up
10 Bosh is so bad its 3 min left in the game and dude
1 8-1 free throw and 10-3 foul advantage for the pacers in the first half
29-15 #ECF #heat #pacers #games6
Omg the heat coming back 2 fast. C’mon #pacers
Score of half time: 40-39. Pacers down
by 1 point
third quarter 30-13 in game 6 Pacers won the quarter
Instance of detected information
2 Wade has only 1 point and Bosh
with 3
General events
is sitting on the bench. How is this guy getting paid
1 D-Wade steal, L-James dunk
15m??? #Heat #NBAPlayoffs
2 Sam Young hits the Pacers’ first 3pointer of the night
11 This is a good #nbaplayoffs game #heatvspacers!
#Pacers came to play.
12 This game is really intense.#Heat #Pacers
Special events
broke the record for the number of
missed dunks in a half
Comments on teams
1 Bad sign for Pacers. Down by 1 at
the half. Should be up by at least 8.
2 Pacers end yet another quarter terribly
3 Both teams are lucky as hell.
Comments on players
1 Wake up Mr West
2 NOTHING from DavidWest
Figure 4 Coverage Rate of our Summary against the NBA Offi-
6. Conclusion
cial Games Story
This paper proposes a system which generates the game
end of the game.
( 2 ) Detailed information: showing the detailed informa-
tion of the game, e.g., the number of rebounds, assists, and
the points that a specific player achieves.
( 3 ) General events: These are common events which are
observed in many games, e.g., brilliant block, slam dunk,
three pointer and game winner.
( 4 ) Special events: These are rare events that may not
happen in every game, e.g., breaking the record of three point
and long consecutive win in the NBA history.
( 5 ) Comments on the performance of the teams.
( 6 ) Comments on the performance of players.
summary for NBA basketball games by quoting the representative tweets issued during the games for the first time,
while the previous researches dealt with soccer and American
football. It is more challenging to make a game summary for
basketball than for football or soccer, since one shoot/goal
is less important for basketball. To enrich the summary, we
actively put not only the real-time actions such as beautiful
slum dunks which are described in the steep spikes in the
tweet volume graph, but also the discussions at the break
times between quarters which usually contain both the game
status and the comments on the previous quarter. We propose a heuristic approach to focus on the spikes of long duration, even if they have rather gentle slopes in order to gather
such discussions. Our algorithm successfully finds the ground
truth important moments in the game more accurately than
the one which considers the slope of the spikes only. Furthermore, selecting the representative tweets per cluster which is
derived by the hierarchical clustering algorithm applied to
the tweets associated with an important moment increases
the diversity of our summary. As the result, our summary
covers up to 87% of the items in the NBA Official Game
Story. Although our summary can not cover all the items
stated in the NBA Official Game Story, it also contains a lot
of useful information not contained there.
There remains problems to be solved in future: First, we
need to make the evaluation method more reliable, since the
current evaluation method depends on the manual works by
ourselves much and is subjective to some extent. Next, our
current system presumes an offline environment. We will extend it, so that meaningful tweets are detected real-time from
the tweet stream and displayed to the NBA fans. Increasing
the readability of our summary should be also pursued, since
our current system displays the representative tweets as they
This work is supported by the Ministry of Education, Culture, Sports, Science and Technology, Grant-in-Aid for Scientific Research (C) 24500111, 2013.
[1] T. Sakaki, M. Okazaki and Y. Matsuo, “Earthquake Shakes
Twitter users: Real-time Event Detection by Social Sensors”, in Proc, of WWW’10, pp. 851–860, 2010.
[2] S. Petrovic, M. Osborne and V. Lavrenko, “Streaming first
story detection with application to Twitter”, in Human
Language Technologies: The 2010 Annual Conference of the
North American Chapter of the Association for Computational Linguistics (NAACL-HLT), pp.181–189, 2010.
[3] D. Chakrabarti and K. Punera, “Event Summarization using Tweets,” in Proc. of ICWSM, 2011.
[4] S. Zhao, L. Zhong, J. Wickramasuriya and V. Vasudevan,
“SportSense: Real-Time Detection of NFL Game Events
from Twitter,” CoRR abs/1205.3212 (2012).
[5] J. Nichols, J. Mahmud, and C. Drews, “Summarizing Sporting Events Using Twitter,” in Proc. of IUI’12, pp. 189–198,
[6] D.A. Shamma, L. Kennedy and E.F. Churchill, “Tweet the
Debates: Understanding Community Annotation of Uncollected Sources,” in Proc. of the first SIGMM workshop on
Social media, 2009.
[7] D.A. Shamma, L. Kennedy and E.F. Churchill, “Peaks and
Persistence: Modeling the Shape of Microblog Conversations,” in Proc. of CSCW’11, pp.355-358, 2011.
[8] J. Weng and F. Lee, “Event Detection in Twitter,” in Proc.
of ICWSM, 2011.