A Cost-Reducing Partial Labeling Estimator in Text Classification Problem

Abstract

The paper proposes a new approach to address the text classification problems when learning with partial labels is beneficial. Instead of offering each training sample a set of candidate labels, researchers assign negative-oriented labels to ambiguous training examples if they are unlikely falling into certain classes. Researchers construct two new maximum likelihood estimators with self-correction property, and prove that under some conditions, new estimators converge faster. Also the paper discusses the advantages of applying one of new estimators to a fully supervised learning problem. The proposed method has potential applicability in many areas, such as crowd-sourcing, natural language processing and medical image analysis.