Detecting Topics Popular in the Recent Past from a Closed Caption TV

Corpus as a Categorized Chronicle Data

Hajime Mochizuki

and Kohji Shibano

Institute of Global Studies, Tokyo University of Foreign Studies, Tokyo, Japan

Research Institute for Languages and Cultures of Asia and Africa, Tokyo University of Foreign Studies, Tokyo, Japan

Keywords:

Topic Detection, Closed Caption TV Corpus.

Abstract:

In this paper, we propose a method for extracting topics we were interested in over the course of the past 28

months from a closed-caption TV corpus. Each TV program is assigned one of the following genres: drama,

informational or tabloid-style program, music, movie, culture, news, variety, welfare, or sport. We focus on

informational/tabloid-style programs, dramas and news in this paper. Using our method, we extracted bigrams

that formed part of the signature phrase of a heroine and the name of a hero in a popular drama, as well as

recent world, domestic, showbiz, and so on news. Experimental evaluations show that our simple method is

as useful as the LDA model for topic detection, and our closed-caption TV corpus has the potential value to

act as a rich, categorized chronicle for our culture and social life.

1 INTRODUCTION

Corpora have become the most important resources

for studies and applications related to natural lan-

guage. Various studies and applications of corpus-

based computational linguistics, knowledge engineer-

ing, and language education have been reported in re-

cent years (Flowerdew, 2011)(Newman et al., 2011).

However, the proportion of spoken language in cor-

pora is quite low. There are only a few “spoken

language corpora,” such as the Corpus for Sponta-

neous Japanese (Maekawa et al., 2000) and a part of

the British National Corpus (Corpus, 2007), that can

be used for research purposes. Under these circum-

stances, we have been continuing to build a large-

scale spoken language corpus from closed-caption

TV (CCTV) data. We have been collecting the CCTV

corpus since December 2012, and its size has reached

over 116,000 TV programs, over 44 million sen-

tences, and 475 million words as of March 2015.

After amassing this spoken-language corpus, we

will be positioned to employ it in a wide variety of

research areas as a language resource. TV is a major

form of media and is familiar in our daily lives. Fur-

thermore, TV broadcasting stations assign each TV

program at least one of twelve genre labels, such as

“Animation,” “Drama,” and “News.” Therefore, we

expect that the CCTV corpus will act as a rich, cate-

gorized chronicle for our culture and social life.

In our research, we aimed to extract recent and

popular or interesting topics from the CCTV cor-

pus. The variety of information changes rapidly in

our modern society, and we often cannot remember

something from recent TV episodes or concerning

topics of interest to us. For example, few people can

quickly describe the prevailing societal issues beyond

the most recent two months. It is unexpectedly difﬁ-

cult to remember what dramas, songs, and topics were

interested.

In this paper, we propose a simple method for

extracting topics that were popular over the past 28

months from the CCTV corpus.

Our research is related to trend or topic de-

tection (Mathioudakis and Koudas, 2010)(Glance

et al., 2004)(Weng and Lee, 2011)(Wang et al.,

2013)(Mochizuki and Shibano, 2014) (Lau et al.,

2012)(Fujimoto et al., 2011)(Keane et al., 2015) and

buzzword extraction (Nakajima et al., 2012) studies.

Many of these studies aim to extract popular topics or

buzzwords from a large number of texts in consumer-

generated medias (CGM), such as Weblog articles or

tweets, but we analyze closed-caption text from the

mass media.

Because of the wide-spread use of the Internet,

there are people of the opinion that social media plays

a central role in public culture and social movements.

However, we believe that TV programs still have

a strong inﬂuence on the general public. Despite the

342

Mochizuki, H. and Shibano, K..

Detecting Topics Popular in the Recent Past from a Closed Caption TV Corpus as a Categorized Chronicle Data.

In Proceedings of the 7th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2015) - Volume 3: KMIS, pages 342-349

ISBN: 978-989-758-158-8

cultural contribution people make with Twitter mes-

sages about speciﬁc topics, the spread of that topic to

the many others who do not use Twitter will be lim-

ited. On the other hand, if such a topic is reported by

TV programs even once, the topic will be well known

to the general public, including those who do not usu-

ally use the Internet. Topics thought to be valuable

to the public will be repeatedly reported in many TV

programs. Words related to the topic will therefore

occur frequently in the voice data of these TV pro-

grams.

In this paper, we focus on three genres:

information/tabloid-style programs, referred to as

“genre I”; drama, referred to as “genre D,”; and news,

referred to as “genre N.” Among these three gen-

res, we center genre I as a nuclear genre, because

information/tabloid-style programs consist of miscel-

laneous topics that are also included in genres N or

D. We expect that a word that occurred frequently in

genre I and which was treated by many TV stations

has a high possibility of being a popular topic that

captured the attention of the public. After detecting

a frequent word in genre I, we decide on the sort of

topic by investigating the appearance tendency of the

same word in genres N and D.

To evaluate the effectiveness of our simple

method, we extracted topics by Latent Dirichlet Allo-

cation (LDA), which is one of the most popular meth-

ods for topic detection(Blei et al., 2003), and com-

pared both methods’ topics. From these results, we

will show that our simple method is as useful as LDA,

and that our TV closed-caption corpus is useful mate-

rial that includes various kinds of popular topics.

Unfortunately, it would be impossible to release

the CCTV corpus itself because of a legal issue. As an

alternative method, we should mention how to build a

corpus from TV data for researchers who wish to use

CCTV data. Therefore, we will describe the details of

the CCTV corpus in the next section.

2 COLLECTING CLOSED

CAPTION TV DATA

2.1 Japanese Television Services with

Closed Caption

In Tokyo, there are seven major broadcasting sta-

tions that organize the nationwide Japanese network:

(1) NHK-G, (2) NHK-E, (3) NTV, (4) TBS, (5) TV

Asahi, (6) Fuji TV, and (7) TV Tokyo. One of the

characteristics of Japanese TV stations is that they

provide a wide variety of programs belonging to dif-

ferent genres. According to the classiﬁcations pro-

vided by the EPG (Electric Program Guide), there

are at least 12 genres: “Animation,” “Sport,” “Cul-

ture and Documentary,” “Drama,” “Variety,” “Film,”

“Music,” “Hobby and Educational,” “Inform Tabloid-

Style,” “Welfare,” and “Other.” According to a re-

port by the Ministry of Internal Affairs and Com-

munications detailing the achievements of closed-

caption TV in 2012

, the amount of programmingthat

included closed-caption data reached approximately

50%.Therefore, a large number of resources for build-

ing a spoken language corpus are currently available

in Japan.

We can use deﬁnition of (ARIB, 2009) to extract

closed-caption data. The following three procedures

are necessary for building a CCTV corpus: (1) Record

all TV programs with closed caption data in a TS

(Transport Stream) format during a 24-hour period;

(2) automatically extract the closed-caption data in

ASS (Advanced SubStation Alpha) format from the

TS data; (3) ﬁlter the ASS-format ﬁle to extract a

plain text format ﬁle and execute a morphological an-

alyzer; and (4) convert the video data in TS format

into MP4 format. Some open-source software appli-

cations are available for these purposes.

2.2 Recording All TV Programs

Television Services with Closed

Caption

We used the freeware packages EpgDataCap

Bon

and EpgTimer to record all TV programs in

Tokyo.EpgTimer can be used to retrieve the EPG

(Electronic Program Guide) list and set the timer to

record. EpgDataCap

Bon is executed by EpgTimer

to record the programs, generating a TS-format ﬁle

and a program information ﬁle for each program. The

TS-format ﬁle is a full segment broadcasting video

ﬁle. The program information ﬁle includes certain in-

formation, such as the program name, broadcast time,

station name, and a program description.

2.3 Extracting Closed Caption Data

The next procedure is to extract the closed caption

data in ASS format from the TS-format ﬁle. The

closed caption data are mixed with video data and

transmitted through digital terrestrial broadcasting.

Therefore, we must use a special program to sepa-

rate the closed caption data from the TS format ﬁle.

Caption2Ass

PCR, a freeware program, is available

http://www.soumu.go.jp/menu

news/s-news/

01ryutsu09

02000071.html

Detecting Topics Popular in the Recent Past from a Closed Caption TV Corpus as a Categorized Chronicle Data

343

for this purpose. An example of an ASS format ﬁle is

shown in Figure 1 and an example of a screen image

of closed-caption TV in Figure 2.

Figure 1: Example of an ASS format ﬁle.

Figure 2: Example of a screen image of closed-caption TV

data. This example is a screen shot from the animation pro-

gram “the Pocket Monsters” broadcasted in Tokyo on May

2, 2013.

As shown in Figure 1, the ASS ﬁle consists of

at least three parts: “[Script Info],” “[V4+ Styles],”

and “[Events].” The ﬁrst, the ”[Script Info]” sec-

tion, contains information about the closed-caption

ﬁle, such as its title, creator, type of script, and dis-

play resolution. In this example, the second line

shows that the text ﬁle was generated by Aegisub,

open-source software used to create closed-caption

data. The eighth and ninth lines show that the

display resolution is 1920x1080 pixels. The sec-

ond section, “[V4+ Styles],” provides a list of

style deﬁnitions, such as font, font size, and pri-

mary/secondary/outline/background color. In this

study, we use the style “Name” as a signiﬁer for

closed-captiontexts. We only use closed-caption texts

in which the style is classiﬁed as “Default.” We ignore

the “Rubi” style because “Rubi” indicates Japanese

kana, which are used to show the pronunciation of

Kanji characters. The third section, “[Events]” is the

list of closed-caption text that the TV station intends

to display with a particular timing. The creator of

closed caption text can specify certain attributes, such

as the style to use for this “event,” the position of

the text on the TV screen, display timing, and text

to display. Information about timing is divided into

start time and end time, and is formatted as h:mm:ss.

The bottom two lines in Figure 1 are examples of

closed-caption texts. In these examples, we use the

two Japanese strings in these lines as the content of

our corpus.

2.4 Filtering ASS ﬁle to Generate Plain

Text and Create Morphemes Data

For post-processing, we ﬁltered the ASS-format ﬁles

to generate plain language text ﬁles without meta-

symbols. These plain texts are composed of Japanese

sentences, and are ﬁnally divided into morphemes

tagged with the part-of-speech information. We used

MeCab (Kudo et al., 2004) as the morphological an-

alyzer. Examples of a plain text and morphemes-

format ﬁle are shown in Figure 3.

Figure 3: An example of plain text extracted from an ASS

ﬁle and its morpheme data processed using MeCab.

RDBPM 2015 - Special Session on Research and Development on Business Process Management

344

2.5 Creating the Closed Caption TV

Corpus

We have currently recorded a total of 116,583 pro-

grams, or 73,832 hours and 13 minutes of program-

ming. Each program is classiﬁed into at least one

genre.

Table 1 shows the total number of morphemes

by genre. The scale of our corpus at this point is

475,096,596 morphemes from 44,951,572 sentences.

We can state that our corpus is presently one of the

largest spoken Japanese corpora. In this study, we

selected and used the sub-corpora from three genres:

drama, labeled D; news, labeled N; and information

and tabloid-style programs, labeled I. The sizes of the

sub-corpora are 12,061 programs, 22,043 programs,

and 16,624 programs, respectively. Approximately

43% of our corpus is made up of 50,728 total pro-

grams.

3 METHOD FOR DETECTING

RECENT TOPICS

Our purpose in this paper is to detect the topics and

episodes of recent interest to us from a large-scale TV

closed-caption corpus. We focus on words related to

genres I, N, and D. We start with the words treated

in genre I and investigate the appearance tendency of

the same words in genres N and D. When a drama

is especially successful, topics related to it also tend

to become popular and to appear in genres I and D.

The words that appear with the same tendency in both

genres I and N tend to express topics related to news

articles. We estimate that this phenomenon can be

observed by comparing the word distribution among

the three genres.

For example, attractive heroes or main characters

in a hit drama would also become popular. Words

related to the characters, such as their names, fash-

ions, and signature dialogue, may frequently oc-

cur not only in dramas (genre D) but also in the

information/tabloid-style genre (genre I).

On the other hand, words related to an unpopu-

lar drama might not appear in genres other than dra-

mas if the drama went unnoticed by other TV pro-

grams. We adopted the following process to discover

topics of past interest to us. Besides, we use bigrams,

which are pairs of adjacent two words, instead of sin-

gle words, in this paper. We think bigrams are more

suitable for forming topics than single words, because

single words tended to be too small as topic words in

our previous work.

(1) Counting the total number of bigrams and the

number of each word every month. We express

the monthly distributions for each bigram by its

monthly proportion.

(2) Picking out bigrams from genre I that appear in

speciﬁc months. These bigrams have the possibil-

ity of being the characteristic bigrams in speciﬁc

topics.

(3) Investigating how the bigrams selected in Step (2)

appear in other genres (genre N and D). It is ex-

pected that if one bigram is related to a popular

drama, the distribution of the bigram in genre D is

similar to the distribution in the genre I. Similarly,

it is also expected that if one bigram is related to

an interesting topic, such as a signiﬁcant event or

crime, the distribution of the bigram in genre N is

similar to the distribution in genre I.

(4) Checking whether the selected bigrams from Step

(3) can be considered to express the topics of in-

terest to us in the past. During this step, we must

conﬁrm the answer by checking texts that contain

the target bigrams in the corpus for the present.

In Step (1), the total occurrence number of bigram

i is calculated with Equation 1.

TotalFreq

∑

freq

i, j

(1)

where freq

i, j

refers to the frequency of the bigram i in

month j. The ratio

i, j

that is the proportion of bigram

i in month j, is calculated with Equations 2 and 3.

biasedFreq

i, j

= freq

i, j

(2)

ratio

i, j

biasedFreq

i, j

∑

biasedFreq

i,n

(3)

where N is the total number of bigrams in the sub-

corpus, and N

is the total number of bigrams in the

j-th month in the sub-corpus. The j varies from 1 to

28 because our corpus has terms of 28 months. There-

fore, each variable of frequency ratio for each bigram

in a sub-corpus has 28 values.

In step (2), we set three thresholds: the bigram

frequency, the monthly frequency, and the frequency

ratio per month.

For the ﬁrst threshold, we select bigrams that have

a frequency of more than one per million bigrams.

For the second threshold, we use a metric simi-

lar to the document frequency (DF) in the informa-

tion retrieval domain. The DF for term i refers to the

number of documents that contain the term i. The DF

is considered to represent the level of a term’s speci-

ﬁcity in the entire document set. We use a monthly

frequency, referred to as MF, instead of the DF. If bi-

gram i appears in all months of our corpus, the MF

Detecting Topics Popular in the Recent Past from a Closed Caption TV Corpus as a Categorized Chronicle Data

345

Table 1: Total number of morphemes for each genre.

Genre No. of Progs. Total Hrs (h:m) No. of Sents. No. of Morphes

A: Animation 9,959 3,612:49 2,961,139 22,053,412

S: Sport 4,313 3,609:09 1,991,883 24,087,726

C: Culture/Documentary 14,313 6,411:42 3,236,149 38,479,292

D: Drama 12,061 10,675:25 7,698,865 66,055,199

N: News 22,043 15,056:48 6,087,279 109,170,849

V: Variety 19,462 17,201:08 12,687,446 107,019,473

F: Film 711 1,413:04 867,138 6,373,679

M: Music 1,927 1,283:59 496,616 5,374,042

H: Hobby/Educational 12,930 3,956:44 3,057,684 27,480,760

I: Information/Tabloid Style 16,624 9,621:34 5,218,918 61,728,141

W: Welfare 1635 693:56 374,138 4,553,119

O: Other 605 300:16 274,317 2,720,904

Total 116,583 73,836:34 44,951,572 475,096,596

for bigram i is 28, because our corpus was collected

over 28 months. We seek bigrams with monthly fre-

quencies of less than 18 months.

For the third threshold, we seek biased bigrams

with a monthly frequency proportion of at least 10%.

This 10% threshold means that the frequency of the

bigram is biased by approximately twice that of the

average proportion, because the average frequency

proportion for 18 months is approximately 5.5%. We

think that it is natural that when a topic captured pub-

lic attention, words related to the topic will occur fre-

quently. We extract the bigrams that satisfy all of

these conditions.

In Step (3), we consider only the bigrams selected

in Step (2). For the other genres, we use the sub-

corpora from genres D and N. We check whether the

distribution for a selected bigram in Step (2) is sim-

ilar to its distributions in both genres D and N. We

remove the bigrams that have fewer similarities. We

can set the similarity criteria. In this study, we use the

following criteria to select bigram i.

(3-1) The monthly frequency of bigram i from the

genre I sub-corpus is the same as or larger than

in genres D or N.

(3-2) At least one of the monthly frequency propor-

tions is over 10% in the genre I sub-corpus.

(3-3) Pearson’s correlation coefﬁcient r between gen-

res I and D, or between genres I and N, equals

0.7 or higher. We calculate and check r for each

bigram.

During Step (4), we check whether the bigrams

selected during Step (3) can be considered to express

recent popular topics. We expect that the bigrams re-

lated to a hit drama or various kinds of news can be

detected during these steps.

4 EXPERIMENTAL EVALUATION

To determine the effectiveness of our method for de-

tecting past topics of interest, we performed an ex-

periment and examined how well it performed with

bigrams.

4.1 Counting the Bigram Frequency

We counted the frequency of bigrams in our corpus

prior to the experiments.

The numbers of bigram types in genres I, D, and

N are 3,158,472, 3,101,213, and 3,541,631, respec-

tively. The numbers of bigram tokens are 65,256,640,

71,908,595, and 112,222,068, respectively. Among

them, the numbers of types that have a frequency of

more than one per million are 83,748 in genre I, 77917

in genre D, and 88,484 in genre N. These ratios are

2.7%, 2.5%, and 2.5%, respectively. Similarly, the

numbers of types that have a frequency of more than

one per million in genres I, D, and N are 52,077,048,

57,784,431,and 92,761,338, respectively.Theseratios

are 79.8%, 80.4%, and 82.7%, respectively.

As shown in Table 3, the total frequency for the

top 2.5% or 2.7% of bigrams is about 80%. Approxi-

mately 2.5% or 2.7% of all three genres’ vocabularies

have a coverage of 80% in the sub-corpora.

4.2 Experiment Involving Extracting

for Bigrams between Genres I and D

The ﬁrst experiment we report assessed the topics that

were mined with our method using the frequency of

bigrams in genres I and D. First, we extracted candi-

date bigrams from the results in Step (2), as described

in Section 3.

RDBPM 2015 - Special Session on Research and Development on Business Process Management

346

Table 2: The extracted bigrams and dramas.

Drama (bigrams): Example(Pearson’s r), ...

Ama-chan (13): Aki/chan(0.82), j/eje(0.86),

kita/sanriku(0.87), no/Ama(0.84)

Massan (11): Massan/no(0.94), Ellie/san(0.91),

Whisky/wo(0.89), Ellie/chan(0.91)

Gochiso-san (7): Me/iko(0.84), Yuta/ro(0.87),

Kazue/san(0.72)

Hanako-to-Ann (6): Hana/chan(0.82),

Renko/san(0.73), Daigo/san(0.80)

Table 3: The extracted bigrams and news.

Category (bigrams): Example(Pearson’s r), ...

World (67): Islamic/State(0.92), Snowden/shi(0.98),

Ebola-hemorrhagic-fever/no(0.93), Pro/Russian(0.95),

Korean-air-lines/no, Jordan/government(0.95)

Domestic (85): unsafe/drug(0.82), d/rone(0.95),

bird/ﬂu(0.98), Hokuriku/super-express(0.94),

Obuchi/san(0.85), Kako/sama(Princess Kako)(0.80)

Showbiz (31): Tsunku/san(0.93),

ghost/writer(0.98), Takakura/san(0.99)

Sport (51): Hanyu/senshu(0.94), a triple/axel(0.97),

Nishikori/senshu(0.92), Hakone/ekiden(0.87),

Zacch/Japan(0.87), World-Cup/ni(0.73)

Weather (64): strong/typhoon (0.96),

bofu/setsu(blizzard)(0.77), hot/summer(0.91)

Culture (28): obake/yashiki(a haunted house)

(0.83), hallo/ween(0.90), nengajo(new year card)/wo

(0.77), Xmas/present(0.93),Shikinen/Sengu(0.82)

Economy (10): Kumiko/shacho (chairwoman)(0.99),

Ootsuka-Kagu (furniture shop)/no(0.99),

oil/price(0.76), kabunushi/sokai (stockholders’

meeting)(0.90)

Science (9): STAP/cell (0.87),

(Obo)kata/Haruko(0.98), Professor/Amano(0.98)

For this experiment, ﬁrst we extracted 3,507 bi-

grams from genre I as candidate bigrams related to

popular dramas. Next, we counted the frequency of

the same 3,507 bigrams in genre D and selected 422

bigrams that had frequencies of more than one per

million. In Step (3), we investigated how these 422

bigrams appeared in genres I and D. We found bi-

grams that had similar distributions in genre pairs I

and D. In fact, we calculated and checked Pearson’s

correlation coefﬁcient r for each bigram between gen-

res I and D. From the result in Step (3), we found 56

bigrams which their r equal 0.7 or higher.

In Step (4), the ﬁnal step, we checked whether the

bigrams selected during Step (3) could be considered

related to recently populardramas. Out of 56 bigrams,

41 were related to dramas. We could not ﬁnd relations

to speciﬁc dramas for 15 bigrams.

The results show that thirteen, eleven, seven and

six bigrams were related to the great hit dramas, Ama-

chan, Massan, Gochiso-san and Hanako-to-Ann, re-

spectively. The remaining four bigrams were related

to three different dramas that were less popular. Table

2 shows the titles of the four dramas and examples of

bigrams.

From a different viewpoint, 29 bigrams were

names or nicknames of characters in dramas. Five

bigrams were related to key items of dramas such as

whisky and shiosai. Two bigrams were the titles of

dramas. Only three bigrams were a part of a signa-

ture phrase of a heroine, and only one bigram was the

name of a location in a drama.

In Table 2, Ama-chan was the greatest hit drama

in 2013, and Aki-chan is the heroine of Ama-chan.

In Japan, everybody knows her. Jejeje is a signa-

ture phrase of hers, and was also the winner of the

Keywords-of-the-Yearcontest for 2013. It can be said

that our method in the ﬁrst experiment tended to yield

parts of character names or signature phrases in past

dramas, including greatly popular dramas. However,

the effectiveness of our method was limited to only

four dramas that were broadcasted by NHK-G from

Monday to Saturday over six months.

4.3 Experiment Involving Extracting

for Bigrams between Genres I and N

The second experiment we report assessed what top-

ics were mined with our method using bigrams in

genres I and N. Like Experiment 1, ﬁrst we extracted

3,507 bigrams from genre I as candidate bigrams re-

lated to news topics.

Next, we counted the frequency of the same 3,507

bigrams in genre N and selected 911 bigrams that had

frequencies of more than one per million. In Step (3),

we investigated how these 911 bigrams appeared in

genres I and N. In fact, we calculated and checked

Pearson’s correlation coefﬁcient r for each bigram be-

tween genres I and N. From the results of Step (3), we

found 454 bigrams whose r equaled 0.7 or higher.

In Step (4), the ﬁnal step, we checked whether

the bigrams selected during Step (3) could be consid-

ered related to recent interesting news. Out of 454 bi-

grams, 345 could be found with relations to any news

topic. We could not determine the relation to speciﬁc

news for 109 bigrams because their meanings were

too general.

Table 3 shows eight categories of news topics and

examples of bigrams. The results show that sixty-

seven, eighty-ﬁve, thirty-one, ﬁfty-one, sixty-four,

twenty-eight, ten, and nine bigrams were related to

Detecting Topics Popular in the Recent Past from a Closed Caption TV Corpus as a Categorized Chronicle Data

347

world, domestic, showbiz, sports, weather, culture,

economy and science news topics, respectively.

It seems that many extracted bigrams reﬂect var-

ious aspects of the present society. Unfortunately,

many of them were not good news. Many topics from

the world news were related to terrorism, the Islamic

State, and Edward Snowden. The reason why there

was a large number of topics related to the weather

reﬂects recent extraordinary weather throughout the

world. Many topics from science news were related

to the STAP cells scandal. However, we extracted

some pleasurable topics, such as the opening of the

Hokuriku Shin-kansen, the new Japanese railroad, and

excellent athletic performances by Yuzuru Hanyu and

Kei Nishikori.

The results of these two experiments are not in-

consistent with our memories and sense impressions.

It can be said that our method of using genres I and N

can yield recently popular news topics.

Table 4: The topics extracted by LDA.

score words

genre I

0.021 America, president, North Korea, Russia,

Obama, Ukraine, ...

0.007 Isramic, Jordan, Syria, Iraq, Turkey, Japanese

0.030 Live, Debut, concert, AKB, Fun, dance, Idol

0.022 Mt. Ontake-san, East Japan, earthquake

0.009 Virus, inﬂuenza, allergy, vaccine, asthma,

a sideeffect,...

0.009 Massan, Ellie, Wisky, ...

0.022 Drama, Kanbei, scene, Taiga, Hero, ...

genre D

0.014 Massan, Ellie, Wisky, Scotland,...

0.160 date, party, girls,...

0.103 policemen, PC, police, net, stalker,...

0.023 Kanbei, Hanbei, onago, gozaru, child,...

0.018 Idol, gege, GMT, memory, mother, cafe,...

genre N

0.199 truck, bicyle, car, police, taxi,

intersection, driver,...

0.096 Obama, president, America, Washington,

Snowden,...

0.078 Tokyo Denryoku, atomic energy, tank,

radioactive rays,...

0.075 goal, soccer, team, league, World cup,...

0.055 Russia, Ukraine, America, President, Europe

0.037 Virus, inﬂuenza, Ebola, vaccine, WHO,...

0.018 Islamic, Jordan, Japanese, pilot, jornalist,

4.4 Supplementary Experiment Using

LDA

This research is potentially related to event or topic

detection. Latent Dirichlet Allocation (LDA) is a

well-known and popular topic model for event or

topic detection(Blei et al., 2003) . It is a powerful

model for analyzing massive sets of text data. Many

works on topic detection have used LDA (Lau et al.,

2012)(Fujimoto et al., 2011)(Keane et al., 2015).

Here, we also tried to apply LDA to our event de-

tection. We tried to detect topics in text sets of genres

I, N, and D by the LDA of the Mallet language toolkit

(McCallum, 2002). We set the number of topics to be

used at 100, and the other parameters used the Mal-

let defaults. Part of the extracted words are shown in

Table 4. In the Table, “score” means Dirichlet coefﬁ-

cient.

Because this method of extracting topics is differ-

ent from our method, it cannot compare the results di-

rectly. We investigated similar word groups between

genres I and D, and genres I and N. The results be-

tween genres I and D show that there are two simi-

lar word groups, related to the TV series Massan and

Kanbei. The results between genres I and N show that

there are three similar word groups, related to Isramic,

Ebola, and Russia and Ukraine.

Relatively few word groups related to two genres

were detected in this supplementary experiment using

LDA.

5 DISCUSSION

In the ﬁrst two experiments, there was no overlap in

the 1,333 bigrams made up of the 422 bigrams in

genre D and the 911 bigrams in genre N. Therefore,

the total number of candidate bigrams was 1,333 out

of 3,507, and we extracted 510 bigrams whose Pear-

sons r in genre I equaled 0.7 or higher. Finally, 395

out of 510 bigrams that had relations to topics from

speciﬁc dramas or news articles were found.

All four dramas related to the extracted bigrams

in genre D were big hits from the recent 28 months.

Many extracted bigrams from genre N were related to

topics that captured public attention during these 28

months.

These results have encouraged us to use our

method of determining the sort of topic after detect-

ing a frequent word in genre I by investigating the ap-

pearance tendency of the same word in genres N and

D. The results also support our expectation that the

CCTV corpus could act as a rich categorized chroni-

cle data.

RDBPM 2015 - Special Session on Research and Development on Business Process Management

348

On the other hand, it seems that a method based

on high frequency is too strict for genre D. Since there

are a few hit dramas that were unable to be detected

by our method, the threshold of frequency should be

lowered in genre D.

In the third supplementary experiment, relatively

few word groups related to two genres were detected.

Though our method is not a more sophisticated model

than LDA, the results seem to show that suitable top-

ics can be detected with our method.

6 CONCLUSION

In this paper, we described methods for detecting past

topics of interest in a CCTV corpus. The results show

that some bigrams related to hit dramas and popular

or interesting news were extracted.

Experimental evaluations show that our simple

method is as useful as the LDA model for topic de-

tection, and our CCTV corpus has the potential value

to act as a rich, categorized chronicle for our culture

and social life.

To improve the accuracy and availability of our

method, we intend to pursue the following future

projects to further our work in detecting recent top-

ics: (1) Reduce the threshold of frequency in genre D

when the similarity between I and D is calculated; (2)

Compare genre I with genres other than D and N; (3)

Count the total number of bigrams and the number of

each bigram every week, rather than every month; (4)

Use a longer unit instead of a bigram.

ACKNOWLEDGEMENTS

This research was supported by JSPS KAKENHI

Grant Numbers 26240051 and 15H02794.

REFERENCES

ARIB (2009). Service information for digital broadcasting

system (in japanese). In Association of Radio Indus-

tries and Businesses.

Blei, D. M., Ng, A. Y., and Jordan, M. I. (2003). Latent

dirichlet allocation. The Jornal of Machine Learning

Research archive, 3.

Corpus, B. N. (2007). The British National Corpus, ver-

sion 3 (BNC XML Edition). Oxford University Com-

puting Services on behalf of the BNC Consortium.

http://www.natcorp.ox.ac.uk.

Flowerdew, L. (2011). Corpora and Language Education.

Palgrave Macmillan.

Fujimoto, H., Etoh, M., Kinno, A., and Akinaga, Y. (2011).

Topic analysis of web user behavior using lda model

on proxy logs. Advances in Knowledge Discovery and

Data Mining, LNCS, 6634/2011:525–536.

Glance, N. S., Hurst, M., and Tomokiyo, T. (2004). Blog-

Pulse: Automated Trend Discovery for Weblogs. In

Proceedings of WWW2004 Workshop on the Weblog-

ging Ecosystem: Aggregation, Analysis and Dynam-

ics.

Keane, N., Yee, C., and Zhou, L. (2015). Using topic mod-

eling and similarity thresholds to detect events. In the

3rd Workshop on EVENTS at the NAACL-HLT 2015,

pages 34–42.

Kudo, T., Yamamoto, K., and Matsumoto, Y. (2004). Ap-

plying conditional random ﬁelds to japanese morpho-

logical analysis. In EMNLP 2004, the Conference on

Empirical Methods in Natural Language Processing,

pages 230–237.

Lau, J. H., Collier, N., and Baldwin, T. (2012). On-line

trend analysis with topic models: #twitter trends de-

tection topic model online. In COLING 2012, pages

1519–1534.

Maekawa, K., Koiso, H., Furui, S., and Isahara, H. (2000).

Spontaneous Speech Corpus of Japanese. In Proceed-

ings of the Second International Conference on Lan-

guage Resources and Evaluation LREC2000, pages

947–952.

Mathioudakis, M. and Koudas, N. (2010). TwitterMonitor:

Trend Detection over the Twitter Stream. In Proceed-

ings of the ACM SIGMOD International Conference

on Management of Data, SIGMOD 2010, pages 1155–

1158.

McCallum, A. K. (2002). Mallet: A machine learning for

language toolkit. http://mallet.cs.umass.edu.

Mochizuki, H. and Shibano, K. (2014). Building very

large corpus containing useful rich materials for lan-

guage learning from closed caption tv. E-Learn

2014, World Conference on E-Learning in Corpo-

rate, Government, Healthcare, and Higher Education,

2014(1):1381–1389.

Nakajima, S., Zhang, J., Inagaki, Y., and Nakamoto, R.

(2012). Early detection of buzzwords based on large-

scale time-series analysis of blog entries. In ACM Hy-

pertext 2012, 23rd ACM Conference on Hypertext and

Social Media, pages 275–284.

Newman, H., Baayen, H., and Rice, S. (2011). Corpus-

based Studies in Language Use, Language Learning,

and Language Documentation. Rodopi Press.

Wang, J., Zhao, X. W., Wei, H., Yan, H., and Li, X.

(2013). Mining New Business Opportunities: Iden-

tifying Trend related Products by Leveraging Com-

mercial Intents from Microblogs. In Proceedings of

the 2013 Conference on Empirical Methods in Natu-

ral Language Processing, pages 1337–1347.

Weng, J. and Lee, B. S. (2011). Event detection in twitter.

In Proceedings of the Fifth International AAAI Con-

ference on Weblogs and Social Media, pages 401–408.

Detecting Topics Popular in the Recent Past from a Closed Caption TV Corpus as a Categorized Chronicle Data

349