Extracting Author Meta-Data from Web Using Visual Features

Abstract

Enriching digital library's author meta-data can lead to valuable services and applications. This paper addresses the problem of extracting authors' information from their homepages. This problem is actually a multiclass classification problem. A homepage can be treated as a group of information pieces which need to be classified to different fields, e.g., Name, Title, Affiliation, Email, etc. In this problem, not only each information piece can be viewed as a point in a feature space, but also certain patterns can be observed among different fields on a page. To improve the extraction accuracy, this paper argues that visual features of information pieces on a homepage should be sufficiently utilized. In addition, this paper also proposes an inter-fields probability model to capture the relation among different fields. This model can be combined with feature- space based classification. Experimental results demonstrate that utilizing visual features and applying the inter- fields probability model can significantly improve the extraction accuracy.