News2Images: Automatically Summarizing News Articles into Image-Based Contents via Deep Learning.

Co mpact representation is a key issue for effective information delivery to users in mobile content-providing services. In particular, it is more severe when providing text documents such as news articles on the mobile service. Here we propose a method for generating compact image-based contents from news documents (News2Image). The proposed method consists of three modules for summarizing news into a few key sentences based on the sematic similarity and diversity, converting the sentences into images, and generating contents consisting of sentence-embedded images. We use word embedding for document summarization and convolutional neural networks (CNNs) for sentence-to-image transformation. These image-based contents improve the readability, thus effectively delivering the core contents of the news to users. We demonstrate the news-to-image content generation on more-than one million Korean news articles using the proposed News2Image. Experimental results show our method generates better image-contents semantically related to the given news articles compared to a baseline method. Furthermore, we discuss some directions for applying News2Images to a news recommendation system.