NHK's tweeting analysis technology that makes use of the reputation of programs on Twitter to make programs

Tweeting the thoughts and thoughts about the program while watching TV is now an ordinary act, but if there are such tweets so much, NHK will be able to make use of something in program production We develop and operate technologies to analyze and classify tweets.

Exhibit item 14 Twitter analysis technology for reputation analysis | NHK STRL OPEN 2013
http://www.nhk.or.jp/strl/open2013/tenji/tenji14/index.html


NHK first collected tweets containing the keyword "NHK" on Twitter. Then most of them were retweeted NHK 's news, tweets about NHK itself, and the tweet where official program titles were specified was between 20 and 30%. So, for example, we developed a program name judgment algorithm that made it possible to pick up even if it was abbreviated, such as "N Spare" if it is NHK Special, "Craw now" if it is Close Up Hyundai, We also made it possible to estimate which tweets of the broadcasting times were tweeted.


This is the number of tweets related to NHK from April 28th to May 5th. You can see that there are many tweets especially on May 3rd.


On May 3, a continuous television novel "Ama-chan"It was arranged by time according to how much tweet about the word was done. The number of people tweeting in the time zone of this broadcasting is the most, and the number of tweets is increasing when other time of rebroadcasting comes.


Tweeted program ranking when it was limited to "May 3 22 o'clock". First place is "Bi-directional quizzes Unified TenkaIn this program, the number of tweets is growing, as viewers can be quiz show which can be participated by remote control. Also, "Ami-chan" is cutting into second place thanks to the rebroadcast at the BS.


Tweets are picked up from what you can search and display on Twitter and use the same data we see. In analysis and analysis, the tweet is classified into three categories of "opinion", "action", "information". Within this category, "opinion" is negative, positive, neutral, demand, expectation, action is recommendation / declaration · standby · view · miss · record · NOD, information is product · plan · content · self · link · news · It is divided into 20 items of official @ · indicator. For example, if it is tweet that "waiting until the cherry blossoms of Yae begin, I am also enjoying today" it is counted as "Positive" "Standby".

Looking at the average weekly weekly average of the Golden Time, about 40% of programs are related, and the rest are non-program related tweets. In addition, the tweet using the official hashtag name and the official program name is 13.5% of the total, 7.3% omitted the program name, those without the notation of the program name ("じ ゃ ぇ" 6.7%, such as what we call it by nickname).


In order to pick up as many tweets as possible, it is said to combine three types of judgment: detection of program name expression, program judgment by comparison between performers and EPG, and program judgment from tweet time.


As an image like this


Although the opinion obtained by this system is actually handed over to the program production team, since the system itself is for in-house use, it was said that it is not something to open for the public . In the future, we plan to increase detection accuracy and classification accuracy and realize more efficient totaling work.


This technology is being held at NHK Technical Research Institute until June 2NHK Institute of Technology Open 2013You can actually see it. It is unnecessary to register in advance and entry is free.