Combining Lightly-Supervised Text Classification Models for Accurate Contextual Advertising

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 545–554, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP Combining Lightly-Supervised Text Classification Models for Accurate Contextual Advertising Yiping Jin Dept. of Mathematics & Computer Science Chulalongkorn university Thailand 10300 Yiping.Ji@student .chula.ac.th Dittaya Wanvarie Dept. of Mathematics & Computer Science Chulalongkorn university Thailand 10300 Dittaya.W@chula.ac.th Phu T. V. Le Knorex Pte. Ltd. 2 Science Park Drive, Singapore 118222 le phu@knorex.com Abstract In this paper we propose a lightly- supervised framework to rapidly build text classifiers for contextual advertising. In contextual advertising, advertisers of- ten want to target to a specific class of webpages most relevant to their product, which may not be covered by a pre-trained classifier. Moreover, the advertisers are only interested in the target class. There- fore, it is more suitable to model as a one- class classification problem, in contrast to traditional classification problems where disjoint classes are defined a priori. We first apply two state-of-the-art lightly- supervised classification models, gener- alized expectation (GE) criteria (Druck et al., 2008) and multinomial na¨ıve Bayes (MNB) with priors (Settles, 2011) to one- class classification where the user only provides a small list of labeled words for the target class. We fuse the two mod- els together by using MNB to automati- cally enrich the constraints for GE train- ing. We also explore ensemble method to further improve the accuracy. On a corpus of real-time bidding requests, the proposed model achieves the highest aver- age F1 of 0.69 and closes half of the gap between previous state-of-the-art lightly- supervised models to a fully-supervised MaxEnt model. 1 Introduction Contextual advertising (or contextual targeting) is a technique to maximize the relevance of online advertisements by performing page analysis of the webpages. Contextual advertising is closely re- lated to text classification problem, which is well- known to the NLP community. Underlying, the system classifies webpages into predefined Ads verticals (categories). Based on the classification result, the real-time bidding (RTB) system will de- cide whether to bid for a particular page to dis- play their Ads. Successful contextual advertising leads to lower advertising cost and higher click- through and conversion rate (Chatterjee et al., 2003; Broder et al., 2007). Contextual targeting differs from traditional text classification in three ways. Firstly, with thou- sands of vastly different products coming to mar- ket every day, there are large number of categories (usually thousands of categories). Secondly, the advertisers often want to customize an existing category or create new categories to target a set of more relevant webpages. I.e. the classes are not static but evolving. It is therefore unrealistic to ask the advertisers to provide labeled webpages for each new or modified category they want to tar- get. Lastly, while the advertisers can provide prior knowledge for the class they want to target, they cannot accurately specify the irrelevant (or nega- tive) category because it likely covers broad topics in the wild. However, to train a classifier, usu- ally labeled instances …

Content truncated to 500 words