Automatic Training Corpora Acquisition through Web Mining
Chien‐Chung Huang, Kuan-Ming Lin, Lee‐Feng Chien · 2005
Text classification is a task having been extensively studied for decades. However, most previous work pre-assumes the existence of explicitly labeled corpora. In this study, we focus on the issue of automatic corpora acquisition. We propose a Web-based mining approach to collect necessary corpora, which can be greatly useful to both common users and system designers. Moreover, the proposed technique can also be incorporated with existing classification techniques to further boost classifier performance. It has been shown that the concept of the class can be captured by the class name and its associated terms (Huang et al., 2004). In this work, we aim at analyzing Web-retrieved documents to discover the associated terms, which are further utilized to collect more training corpora. Working iteratively, the proposed approach can acquire training corpora of high quality. We give empirical evidence that the classifiers thus created have promising accuracy. In sum, the convenience and efficiency of the proposed approach, along with the new perspective on the issue of corpora acquisition, are the primary contributions of this work.