Distant-supervised Language Model for Detecting Emotional Upsurge on Twitter

Distant-supervised Language Model for Detecting Emotional Upsurge on Twitter Yoshinari Fujinuma†,‡∗ †University of Tokyo fujinumay@gmail.com Hikaru Yokono‡ Pascual Mart´ınez-G´omez§,‡ ‡National Institute of Informatics §Ochanomizu University {yokono, pascual, aizawa}@nii.ac.jp Akiko Aizawa†,‡ Abstract Event-specific twitter streams often reveal sudden spikes triggered by users’ upsurge of emotions to crucial moments in the real world. Although upsurge of emotion is usually identi- fied by a sudden rise in the number of tweets, the detection for diverse event streams is not a trivial task. In this paper, we propose a new method to extract spiking tweets with up- surge of emotions based on characteristic ex- pressions used in tweets. The core part of our method is to use a distant-supervised language model (Spike LM) built from tweets in spikes to capture such expressions. We investigate the performance of detecting emotional spik- ing tweets using language models including Spike LM. Our experimental results show that the natural language expressions used in emo- tional upsurge fit specifically well to Spike LM. 1 Introduction Twitter is one of the most popular micro-blogging platforms in recent days. There are over 500 mil- lion tweets posted per day1 including real-world events described on Twitter which range from short and daily life events (e.g. falling to the ground) to long and widely-broadcasted events (e.g. a match in World Cup). Such tweets are good sources to detect users’ reactions toward real-world events. ∗This work was done while the first author was at the Uni- versity of Tokyo and National Institute of Informatics. The first author is currently at Amazon Japan K.K. 1https://blog.twitter.com/2013/new-tweets-per-second- record-and-how Figure 1: The number of tweets per minute (TPM) during Japan vs. Cameroon for hashtags related to World Cup 2010. People behave unusually when they encounter ex- citing moments in an event, for example, yell out or dance with each other after their favorite soccer team scores a goal. On Twitter, this action is often re- flected by a large number of posts within a short time period. When Japan scored a goal against Cameroon in World Cup 2010, there were a maximum of 2, 940 tweets per second (TPS), which marked the record TPS for goals at that time.2 It is significantly larger than the average of 750 TPS.2 In this paper, we call such bursty traffic as “numerical spikes”. Figure 1 shows the number of tweets per minute during the match of Cameroon vs. Japan, and Table 1 shows the examples of tweets sampled from both numeri- cal spikes and other parts. Detecting emotional upsurge is important for both extracting emerging important real-world events and important moments of them. We call an upsurge that are caused by Twitter users’ emotional spike as 2https://blog.twitter.com/2010/big-goals-big-game-big- records PACLIC 29 488 29th Pacific Asia Conference on Language, Information and Computation pages 488 - 495 Shanghai, China, October 30 - November 1, 2015 Copyright 2015 by Yoshinari Fujinuma, Hikaru Yokono, Pascual Mart´ınez-G´omez and Akiko Aizawa Moment Example of a Tweet English Translation Emotional Upsurge Japan won the match やったああああああああ ああああああああああ ああ #jpn …

Content truncated to 500 words