Semantic Role Labeling for News Tweets

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 698–706, Beijing, August 2010 Semantic Role Labeling for News Tweets 1,2Xiaohua Liu, 3Kuan Li*, 4Bo Han*, 2Ming Zhou, 2Long Jiang, 3Zhongyang Xiong and 2Changning Huang 1School of Computer Science and Technology Harbin Institute of Technology 2Microsoft Research Asia 3College of Computer Science Chongqing University 4School of Software Dalian University of Technology {xiaoliu, v-kuli, v-bohan, mingzhou, longj} @microsoft.com zyxiong@cqu.edu.cn v-cnh@microsoft.com Abstract News tweets that report what is happen- ing have become an important real-time information source. We raise the prob- lem of Semantic Role Labeling (SRL) for news tweets, which is meaningful for fine grained information extraction and retrieval. We present a self-supervised learning approach to train a domain spe- cific SRL system to resolve the problem. A large volume of training data is auto- matically labeled, by leveraging the ex- isting SRL system on news domain and content similarity between news and news tweets. On a human annotated test set, our system achieves state-of-the-art performance, outperforming the SRL system trained on news. 1 Introduction Tweets are text messages up to 140 characters. Every day, more than 50 million tweets are gen- erated by millions of Twitter users. According to the investigation by Pear Analytics (2009), about 4% tweets are related to news1. * This work has been done while the author was visiting Microsoft Research Asia. 1 http://blog.twitter.com/2010/02/measuring-tweets.html We divide news related tweets into two cate- gories: those excerpted from news articles and those not. The former kind of tweets, hereafter called news excerpt, is formally written while the latter, hereafter called news tweet, varies in style and often is not grammatically correct. To understand the proportion of news tweets, we randomly selected 1000 tweets related to news, and got 865 news tweets. Following is an exam- ple of anews tweet, containing oh, yea, which usually appear in spoken language, and :-(, an emoticon. oh yea and Chile earthquake the earth off it's axis according to NASA and shorten the day by a wee second :-( (S1) News tweets arean important information source because they keep reporting what is hap- pening in real time. For example, the earthquake near Los Angeles that happened on Tuesday, July 29, 2008 was first reported through news tweets only seconds later than the outbreak of the quake. Official news did not emerge about this event until four minutes later. By then, "Earthquake" was trending on Twitter Search with thousands of updates2. However, it is a daunting task for people to find out information they are interested in from such a huge number of news tweets, thus moti- vating us to conduct some kind of information 2 http://blog.twitter.com/2008/07/twitter-as-news-wire.html 698 extraction such as event mining, where SRL plays a crucial role (Surdeanu et al., 2003). Considering Sentence 1, suppose the agent earthquake and the patient day for the predicate shorten are identified. Then it is straightforward to output the event Chile earthquake shorten the day, which captures the essential information encoded in this tweet. …

Content truncated to 500 words