Trouble information extraction based on a bootstrap approach from Twitter

Trouble information extraction based on a bootstrap approach from Twitter Kohei Kurihara Kazutaka Shimada Department of Artificial Intelligence Kyushu Institute of Technology 680-4 Kawazu Iizuka Fukuoka Japan shimada@pluto.ai.kyutech.ac.jp Abstract In this paper, we propose a method for ex- tracting trouble information from Twitter. One useful approach is based on machine learn- ing techniques such as SVMs. However, trou- ble information is a fraction of a percent of all tweets on Twitter. In general, imbalanced distribution is not suitable for machine learn- ing techniques to generate a classifier. An- other approach is to extract trouble informa- tion by using handwritten rules. However, constructing high coverage rules by handwork is costly. First, we verify these problems in a preliminary experiment. Then, to solve these problems, we apply a bootstrapping method to our trouble information extraction task. We introduce three characteristics and a scoring method to the bootstrapping. As a result, the iteration process on the bootstrapping in- creased the number of tweets and patterns for trouble information dramatically. 1 Introduction The World Wide Web contains a huge number of on- line documents that are easily accessible. Analysis of the documents has an important role for natural language processing. One of the important informa- tion for business companies is trouble information of a product as the risk management. If they can mon- itor the information about products and the troubles from the Web automatically, they might be able to avoid critical damages by realizing the risk in ad- vance. Therefore trouble information extraction is a significant task in business. There are many stud- ies which handled news articles (Sakai et al., 2006), review documents (Ivanov and Tutubalina, 2014), financial documents (Leider and Schilder, 2010), daily reports (Kakimoto and Yamamoto, 2008), a failure database on the Web (Awano et al., 2012) and so on, as the target data. However, these in- formation sources are not usually instantaneous and exhaustive. To solve this problem, we focus on Twit- ter. It is one of the most famous microblogging ser- vices and text-based posts of up to 140 characters. The posted sentences are described as “tweets.” We suppose users on Twitter often post tweets with trou- ble information because they tend to post tweets as lifelog data in real time. Some researchers focused on the characteristic (Aramaki et al., 2011; Sakaki et al., 2010; Shimada et al., 2012). In this paper, we propose a method to extract trou- ble information from Twitter. One of the most com- mon approaches is to classify an input into trouble information and non-trouble information by using a machine learning technique. However, most of the tweets do not relate to trouble information. In other word, the ratio of trouble tweets and non-trouble tweets is biased. Such biased data generally gen- erate a unsuitable classifier. Another approach is to extract trouble information by using handwritten rules. However, constructing high coverage rules by handwork is usually a difficult task. In this paper, we investigate these problems through a preliminary experiment. On the basis of …

Content truncated to 500 words