Extracting Important Tweets for News Writers using Recurrent Neural Network with Attention Mechanism and Multi-task Learning

31st Pacific Asia Conference on Language, Information and Computation (PACLIC 31), pages 363–369 Cebu City, Philippines, November 16-18, 2017 Copyright c⃝2017 Taro Miyazaki, Shin Toriumi, Yuka Takei, Ichiro Yamada and Jun Goto Extracting Important Tweets for News Writers using Recurrent Neural Network with Attention Mechanism and Multi-task Learning Taro Miyazaki, Yuka Takei, Ichiro Yamada, Jun Goto NHK Science & Technology Research Laboratories 1-10-11 Kinuta, Setagaya-ku, Tokyo, Japan {miyazaki.t-jw, takei.y-ek, yamada.i-hy, goto.j-fw}@nhk.or.jp Shin Toriumi Tokyo City University 3-3-2 Ushikubo-Nishi, Tsuzuki-ku Yokohama, Kanagawa, Japan g1683108@tcu.ac.jp Abstract Social media is an important source for news writers. However, extracting useful informa- tion for news writers from the vast amount of social media information is laborious. There- fore, services that enable news writers to ex- tract important information are desired. In this paper, we describe a method to extract tweets that include useful information for news writ- ers. Our method uses a Recurrent Neural Network (RNN) with an attention mechanism and multi-task learning that processes each character in the tweet to estimate whether the tweet includes important information. In our experiment, we compared two types of at- tention mechanism and compared their types with/without multi-task learning. By our pro- posed method, we obtained an F-measure of 0.627, which is 0.037 higher than that of base- line method. 1 Introduction Social media information is now an important source for news writers. People who encounter an incident can post what is happening before his/her eyes using photos and videos. These posts are im- portant primary information, so news writers want to gather them. However, extracting useful information from the vast amount of social media information is laborious. For this reason, services that enable news writers to extract information that can be used as a news source are desired. In fact, some services such as Spectee1 and FASTALERT2 have been launched 1http://www.spectee.com 2https://fa.xwire.jp in Japan. These services gather much information from social media and extract information that can be used as news sources. Information that news writers want to extract from social media includes many different topics such as fires, accidents, and other incidents. Therefore, ex- tracting information from social media by filtering with keywords is difficult. Assuming the words “de- lay” and “train” are included in the keywords, the tweet “xxx line is delayed by accident,” which can be used as a news source, can be extracted. How- ever, the tweet “I hope the train is delayed because I haven’t studied for today’s exam,” which cannot be used as a news source, is also extracted. To extract tweets that include important information, filtering by keyword is not enough because the output may include tweets that cannot be used as news sources. For this reason, we have been studying automatic extraction of useful information from social media. Our purposes are to reduce the amount of labori- ous work and extract information that cannot be ex- tracted by using queries. In this paper, we describe a method to extract tweets that include useful infor- mation for news writers. Generally, …

Content truncated to 500 words