Quality-biased Ranking of Short Texts in Microblogging Services

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 373–382, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Quality-biased Ranking of Short Texts in Microblogging Services Minlie Huang Dept. of Computer Science and Technology, Tsinghua University Beijing 100084, China aihuang@tsinghua.edu.cn Yi Yang School of Software Beihang University Beijing, China yangyiycc@gmail.com Xiaoyan Zhu Dept. of Computer Science and Technology, Tsinghua University Beijing 100084, China zxy dcs@tsinghua.edu.cn Abstract The abundance of user-generated content comes at a price: the quality of content may range from very high to very low. We pro- pose a regression approach that incorporates various features to recommend short-text doc- uments from Twitter, with a bias toward qual- ity perspective. The approach is built on top of a linear regression model which includes a regularization factor inspired from the con- tent conformity hypothesis - documents sim- ilar in content may have similar quality. We test the system on the Edinburgh Twitter cor- pus. Experimental results show that the regu- larization factor inspired from the hypothesis can improve the ranking performance and that using unlabeled data can make ranking perfor- mance better. Comparative results show that our method outperforms several baseline sys- tems. We also make systematic feature anal- ysis and find that content quality features are dominant in short-text ranking. 1 Introduction More and more user-generated data are emerging on personal blogs, microblogging services (e.g. Twit- ter), social and e-commerce websites. However, the abundance of user-generated content comes at a price: there may be high-quality content, but al- so much spam content such as advertisements, self- promotion, pointless babbles, or misleading infor- mation. Therefore, assessing the quality of infor- mation has become a challenging problem for many tasks such as information retrieval, review mining (Lu et al., 2010), and question answering (Agichtein et al., 2008). In this paper, we focus on predicting the quality of very short texts which are obtained from Twit- ter. Twitter is a free social networking and micro- blogging service that enables its users to send and read other users’ updates, known as ”Tweets”. Each tweet has up to 140 characters in length. With more than 200 million users (March 2011), Twitter has be- come one of the biggest mass media to broadcast and digest information for users. It has exhibited advan- tages over traditional news agencies in the success of reporting news more timely, for instance, in report- ing the Chilean earthquake of 2010 (Mendoza et al., 2010). A comparative study (Teevan et al., 2011) shows that queries issued to Twitter tend to seek more temporally relevant information than those to general web search engines. Due to the massive information broadcasted on Twitter, there are a huge amount of searches ev- ery day and Twitter has become an important source for seeking information. However, according to the Pear Analytics (2009) report on 2000 sample tweet- s, 40.5% of the tweets are pointless babbles, 37.5% are conversational tweets, and only 3.6% are news (which are most valuable for users who …

Content truncated to 500 words