Learning Only from Relevant Keywords and Unlabeled Documents

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 3993–4002, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 3993 Learning Only from Relevant Keywords and Unlabeled Documents Nontawat Charoenphakdee1,3 Jongyeong Lee1,3 Yiping Jin2 Dittaya Wanvarie2 Masashi Sugiyama3,4,1 1Department of Computer Science, The University of Tokyo 2Department of Mathematics and Computer Science, Faculty of Science, Chulalongkorn University 3RIKEN Center for Advanced Intelligence Project 4Department of Complexity Science and Engineering, The University of Tokyo {nontawat, lee}@ms.k.u-tokyo.ac.jp jinyiping@knorex.com. Dittaya.W@chula.ac.th sugi@k.u-tokyo.ac.jp Abstract We consider a document classification prob- lem where document labels are absent but only relevant keywords of a target class and unla- beled documents are given. Although heuristic methods based on pseudo-labeling have been considered, theoretical understanding of this problem has still been limited. Moreover, pre- vious methods cannot easily incorporate well- developed techniques in supervised text clas- sification. In this paper, we propose a theo- retically guaranteed learning framework that is simple to implement and has flexible choices of models, e.g., linear models or neural net- works. We demonstrate how to optimize the area under the receiver operating characteristic curve (AUC) effectively and also discuss how to adjust it to optimize other well-known eval- uation metrics such as the accuracy and F1- measure. Finally, we show the effectiveness of our framework using benchmark datasets. 1 Introduction Supervised text classification is a traditional prob- lem in natural language processing that has been studied extensively (Nigam et al., 2000; Sebas- tiani, 2002; Forman, 2003; Joulin et al., 2016). In this problem, we are given a set of labeled text documents and the goal is to construct a classifier that can classify an unseen document effectively. There are many useful applications for text clas- sification, e.g., sentiment classification (Liu and Zhang, 2012; Medhat et al., 2014), biomedical text mining (Cohen and Hersh, 2005; Huang and Lu, 2015), and social media monitoring (Zeng et al., 2010; Hu and Liu, 2012). In the real-world, it is impractical to expect that labeled data can always be obtained abun- dantly. For example, given big data of unlabeled texts, it can be very time-consuming and costly for the labeling task so that we can apply a super- vised text classification method. Another exam- ple is when the label information is protected due to privacy concerns. These bottlenecks motivate researches on weakly-supervised learning (Zhou, 2017), which focuses on devising a machine learn- ing method that can learn effectively although a lot of labeled data are not accessible. An alternative is to provide keywords as a hint for classifying a document to a target class. Intu- itively, this approach can be much cheaper when the number of documents is huge since the num- ber of keywords does not necessarily grow linearly with the number of documents. Dataless classifi- cation is a text classification problem where we are given keywords for each class and unlabeled doc- uments (Chang et al., 2008; Song and Roth, 2014; …

Content truncated to 500 words