Determining the Fitness of a Document Model by Using Conflict Instances

Documents cannot be automatically classi ed unless they have been represented as a collection of computable features. A model is a representation of a document with computable features. However, a model may not be su cient to express a document, especially when two documents have the same features, they might not be necessarily classi ed into the same category. We propose a method for determining the tness of a document model by using con ict instances. Con ict instances are instances with exactly same features, but with di erent category labels given by human expert in an interactive document labelling process for training of the classi er. In our paper, we do not treat con ict instances as noises, but as the evidences that can reveal a distribution of positive instances. We develop an approach to the representation of this distribution information as a hyperplane, namely distribution hyperplane. Then the tness problem becomes a problem of computing the distribution hyperplane. Besides determining the tness of a model, distribution hyperplane can also be used for: 1) acting as classi er itself; and 2) being a membership function of fuzzy sets. In this paper, we also propose the selection criteria of e ectiveness measuring for a model in a process of tness computations.