Intelligent Fusion of Evidence from Multiple Sources for Text Classification

Automatic text classification using current approaches is known to perform poorly when documents are noisy or when limited amounts of textual content is available. Yet, many users need access to such documents, which are found in large numbers in digital libraries and in the WWW. If documents are not classified, they are difficult to find when browsing. Further, searching precision suffers when categories cannot be checked, since many documents may be retrieved that would fail to meet category constraints. In this work, we study how different types of evidence from multiple sources can be intelligently fused to improve classification of text documents into predefined categories. We present a classification framework based on an inductive learning method -- Genetic Programming (GP) -- to fuse evidence from multiple sources. We show that good classification is possible with documents which are noisy or which have small amounts of text (e.g., short metadata records) -- if multiple sources of evidence are fused in an intelligent way. The framework is validated through experiments performed on documents in two testbeds. One is the ACM Digital Library (using a subset available in connection with CITIDEL, part of NSF's National Science Digital Library). The other is Web data, in particular that portion associated with the Cadê Web directory. Our studies have shown that improvement can be achieved relative to other machine learning approaches if genetic programming methods are combined with classifiers such as kNN. Extensive analysis was performed to study the results generated through the GP-based fusion approach and to understand key factors that promote good classification.