Semi-supervised Classification of Twitter Messages for Organization Name Disambiguation

International Joint Conference on Natural Language Processing, pages 869–873, Nagoya, Japan, 14-18 October 2013. Semi-supervised Classification of Twitter Messages for Organization Name Disambiguation Shu Zhang1, Jianwei Wu2, Dequan Zheng2, Yao Meng1 and Hao Yu1 1 Fujitsu Research and Development Center, Beijing, China {zhangshu, mengyao, yu}@cn.fujitsu.com 2 School of Computer Science and Technology, Harbin Institute of Technology, Harbin, China {jwwu, dqzheng}@mtlab.hit.edu.cn Abstract In this paper, we probe the problem of organi- zation name disambiguation on twitter mes- sages. This task is challenging due to the fact of lacking sufficient information in a tweet message. Instead of conventional methods based on mining external information from web sources to enrich information about or- ganization, we propose to mine the relation- ship among tweets in data set to utilize context information for disambiguation. With a small scale of labeled tweets, we propose LP-based and TSVM-based semi-supervised methods to classify tweets. We aim to mine both related and non-related information for a given organ- ization. The experiments on WePS-3 show that proposed methods are effective. 1 Introduction Twitter is an online social networking and microblogging service, which rapidly gained worldwide popularity. How to retrieval, analyze and monitor Twitter information has been receiv- ing a lot of attention in natural language pro- cessing and information retrieval research com- munity (Kwak, et al., 2010; Boyd, et al.,2010; Tsagkias, et al., 2011). One of the essential things of these researches is first to get the in- formation which is related to the studied entity. This is caused by the ambiguity of entities. For example, the name of company “Apple” has a separate meaning referring to one kind of fruit. The word “Amazon” also could refer to river or company. In this paper, we focus on finding related tweets for a given organization, which can be treated as a binary classification problem. As- suming that tweets are retrieved by a query, such as “apple”, the task is to classify whether each retrieved tweet is relevant to the target organiza- tion (“Apple Inc.”) or not. However, constructing such a classifier is a challenging task, as tweets are short and informal. Additionally, the infor- mation about a given organization is limited, which is difficult to cover the word occurrences in the given organization related tweets. Different from previous work on mining ex- ternal information from web sources to enrich information about the given organization, we propose to mine the relationship among retrieved tweets in data set. With a small scale of labeled tweets, we propose semi-supervised methods to mine the relationships between labeled and unla- beled tweets for the given organization. The remainder of the paper is organized as fol- lows: Section 2 describes the related work on name disambiguation. Section 3 gives problem description and an overview of our approach. Section 4 and Section 5 present LP-based and TSVM-based semi-supervised methods to classi- fy tweets. Section 6 gives the experiments and results. Finally section 7 summarizes this paper. 2 Related Work Twitter contains little information in each tweet, with no …

Content truncated to 500 words