Generating Abstractive Summaries of Sports Games from Japanese Tweets

Microblogs are one of the most important resources for natural language processing. In this paper, we deal with a summarization task of sports events on Twitter. Although many researchers have proposed methods to summarize tweets about sports events, the methods were based on extractive approaches. We focus on an abstractive approach to generate a better summary as compared with the extractive approaches. First, our method detects burst situations in which many users post tweets when a sub-event in a game occurs. Tweets in the burst situations are the inputs of our method. Next, it extracts sub-event elements (SEEs) that contain actions in a game, such as “Player A made a pass to Player B” and “Player B made a shot on goal.” Then, it identifies the optimal order of the extracted SEEs by using a scoring method. Finally, it generates an abstractive summary on the basis of the ordered SEEs, such as “Playler B made a shot on goal from the Player A's pass.” In the experiment, we show the effectiveness of our method as compared with related work based on an extractive approach.