Tweet Extraction for News Production Considering Unreality

31st Pacific Asia Conference on Language, Information and Computation (PACLIC 31), pages 370–375 Cebu City, Philippines, November 16-18, 2017 Copyright c⃝2017 Yuka Takei, Taro Miyazaki, Ichiro Yamada and Jun Goto Tweet Extraction for News Production Considering Unreality Yuka Takei, Taro Miyazaki, Ichiro Yamada, Jun Goto NHK Science & Technology Research Laboratories 1-10-11 Kinuta, Setagaya-ku, Tokyo, Japan {takei.y-ek, miyazaki.t-jw, yamada.i-hy, goto.j-fw}@nhk.or.jp Abstract Acquiring information on incidents and acci- dents from social media can be useful for broadcasters to report news faster. However, many tweets including words related to inci- dents and accidents are actually irrelevant to re- al events, for example, “Backdraft's explosion scene was impressive!!!” Social media contains many comments on events in unreal worlds such as movies, animations and dramas, and it is time-consuming to discriminate these tweets manually. This work presents a method for au- tomatically extracting useful tweets for news reports by focusing on “unreal” information. We first prepare unreal tweets as learning data and use a distributed representation and features that can determine if a tweet is real or unreal. By adding the features of a neural network, we generate a learning model that can effectively discriminate whether a tweet includes infor- mation on actual incidents or accidents. Results of evaluations revealed that the proposed meth- od achieved a 3.8-point higher F-measure than the baseline method. 1. Introduction Social networking services (SNSs) enable us to easily transmit information anywhere in real-time. The large amount of information transmitted on SNS, known as “Social Big Data,” is a valuable information source for grasping newsworthy oc- currence (Vieweg et al., 2010; Kanouchi et al., 2015), and broadcasters monitor social media such as Twitter to collect information about incidents and accidents. By obtaining information directly from witnesses of such events, broadcasters can report news more quickly and effectively. They use various tools to manually search for tweets that have potential news value, using keywords to find tweets indicating incidents and accidents. However, a lot of effort is required to find valuable infor- mation from among the large number of tweets sent every day. Methods have been reported for automatically extracting tweets with potential news value by us- ing machine learning (Freitas et al., 2016; Mizuno et al., 2016; Doggett et al., 2016). However, many tweets irrelevant to actual incidents or accidents include relevant words, which worsen the extrac- tion results. Examples include tweets about events in current animations and TV programs, such as “ド ラえもん「のび太の家火事になる・前編」 (Doraemon - Nobita's House Fire・Part 1).” Many viewers tweet while watching TV to share their opinions with other people. Therefore, many tweets include names of animations (which we call “virtual proper nouns”) and TV programs. In addition, words in Japanese idioms could also suggest incidents or accidents, such as “火の無いところに煙は立たない (Where there’s smoke, there’s fire).” Furthermore, there are tweets that include hypothetical expres- sions that assume an incident or accident occurring, such as “火事になったら、どこに逃げるべきだろう (If a fire occurs, where should I escape to?).” All three tweets include the word “fire” but do not indicate the …

Content truncated to 500 words