Using Content-level Structures for Summarizing Microblog Repost Trees

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 2168–2178, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. Using Content-level Structures for Summarizing Microblog Repost Trees ∗ Jing Li1,2, Wei Gao3, Zhongyu Wei4, Baolin Peng1,2 and Kam-Fai Wong1,2 1The Chinese University of Hong Kong, Shatin, N.T., Hong Kong 2MoE Key Laboratory of High Confidence Software Technologies, China 3Qatar Computing Research Institute, Hamad Bin Khalifa University, Doha, Qatar 4The University of Texas at Dallas, Richardson, Texas, USA {lijing,blpeng,kfwong}@se.cuhk.edu.hk1,2 wgao@qf.org.qa3, zywei@@hlt.utdallas.edu4 Abstract A microblog repost tree provides strong clues on how an event described therein develops. To help social media users capture the main clues of events on mi- croblogging sites, we propose a novel re- post tree summarization framework by ef- fectively differentiating two kinds of mes- sages on repost trees called leaders and followers, which are derived from content- level structure information, i.e., contents of messages and the reposting relations. To this end, Conditional Random Fields (CRF) model is used to detect leaders across repost tree paths. We then present a variant of random-walk-based summariza- tion model to rank and select salient mes- sages based on the result of leader detec- tion. To reduce the error propagation cas- caded from leader detection, we improve the framework by enhancing the random walk with adjustment steps for sampling from leader probabilities given all the re- posting messages. For evaluation, we construct two annotated corpora, one for leader detection, and the other for repost tree summarization. Experimental results confirm the effectiveness of our method. 1 Introduction Microblogging platforms have become the center for reporting, discussing, and disseminating real- life issues, on which users usually repost to share microblog messages with their following users. Also, users can repost with commentary for not only further broadcasting but also extending the ∗This work is partially supported by General Research Fund of Hong Kong (417112), RGC Direct Grant (417613), and Huawei Noah’s Ark Lab, Hong Kong. We would like to thank anonymous reviewers for the useful comments. original microblog post content. Because an in- dividual post is generally too short to cover the main clues of an event, microblogging users can- not easily capture the key information from re- ceived posts due to the lack of context. And re- posting messages, namely reposts, can provide valuable context information to the previous posts including their background, development, public opinions and so on. However, a popular post usu- ally attracts a large number of reposts. It is imprac- tical for users to read them all and fully understand their contents. The task of microblog context summarization aims to produce succinct summaries to help users better understand the main clues by extracting salient information among massive reposts of the original posts. An intuitive approach is to directly apply existing extractive summarizers based on the unstructured, plain microblog contents. But such short and informal reposts render the lack of structures in each individual message, and it is difficult for conventional extractive summarizersto identify salient messages. …

Content truncated to 500 words