Generation of Conceptual-Level Text Cloud with Graph Diffusion

Generation of Conceptual-Level Text Cloud with Graph Diffusion Ying-Chun Lin, Po-An Yang, Yen-Kuan Lee and Kun-Ta Chuang Dept. of Computer Science and Information Engineering National Cheng Kung University yclin@netdb.csie.ncku.edu.tw, yangpoan@gmail.com, yenkuanlee@gmail.com, ktchuang@mail.ncku.edu.tw Abstract In this paper, we explore a novel framework to generate a well-known text cloud visualiza- tion with the conceptual sense. The traditional text cloud is usually generated according to the word occurrence, possibly including the idf-based concept for word weight. The solution is applicable for the long articles. However, for a set of short sentences such as daily news titles, we cannot easily understand the weight of each keyword and its impor- tance to users since the idf value and occurrence in short sentences are difficult to be both well discriminative. In this paper, we propose a graph-based diffusion model to generate conceptual level keyword cloud. We utilize the RDF-based Wikipedia word relation and apply in the Chinese news titles from different news sources. The result shows that our visualization can easily capture the importance concept revealed in a set of news titles. Keywords: Keyword extraction, Document analysis, Document visualization, Graph- based ranking algorithm 1. INTRODUCTION As of today, information is ubiquitously generated as the advance of Internet and mobile device. News, movies or books are digitalized to reach more people around the world. For example, Google undertook a Google Books Library Project1 to scan some old and hard- copy books into its search database, allowing people to query the content in the convenient 1http://www.google.com.tw/googlebooks/library/ The 2016 Conference on Computational Linguistics and Speech Processing ROCLING 2016, pp. 402-411  The Association for Computational Linguistics and Chinese Language Processing 402 manner. Information from anywhere and at anytime can be reached in seconds. However, the amount of information needed for each person is significantly small comparing to the all available information. The previous keyword extraction algorithms [5] have been developed for providing concise yet accurate keywords about an article or a piece of news. As such, people can easily acquire sufficient information by merely viewing the keywords. On the other hand, people can consume less time to grasp the concept of a piece of information simply from the extracted keywords. For example, if we know a piece of news has ”AlphaGo”, ”Lee Sedol” and ”Go match” as its keywords, it is easy to know that the news is about the Go match2 between the 18-time world champion, Lee Sedol, and a computer Go program, AlphaGo. Likewise, if we extract the keywords from a complex system, we can help people understand the system without diving into large amount of information within the system. Take another scenario as an example. To understand the current status in the US, we simply collect the news title of the past few weeks or months and extracted the keywords from those titles. The extracted keywords from the titles may contains ”Hillary”, ”Trump” and ”President”. It is easy to know that this year is the election year for the US president. Users can easily conclude that …

Content truncated to 500 words