Simple and sophisticated inning summary generation based on encoder-decoder model and transfer learning

Abstract

This paper describes an inning summarization method for a baseball game by using an encoder-decoder model. Each inning in a baseball game contains some events, such as hits, strikeouts, homeruns and scoring. Simplified description of the events leads to the improvement of readability of the inning information. Our method learns a relation between play-by-play data in each inning and inning reports. We also incorporate sophisticated expressions acquired from game summaries with the model. We call them Game-changing Phrase, GP. One problem in our task is the size of training data for the learning. To solve this problem, we apply a transfer learning approach into our method. In the experiment, we evaluate the effectiveness of our method with the transfer learning.