Metadata Extraction Approach of PDF Documents Based on Measurement Fusion

摘要

To deal with the problems of low precision rate and weak adaptability in the existing metadata extraction methods, a novel metadata extraction approach is proposed based on measurement fusion rule in this paper. First, the features of the document header are extracted, the three statistical learning methods such as HMM, SVM and CRF are respectively employed to train the labeled data set, and corresponding metadata extraction models are constructed. Then, the results from three extraction models are fused by the sum rule so as to achieve the accurate metadata extraction of documents. Finally, we dynamically update the three extraction models to guarantee the effectiveness of the ensemble models by the time period statistics-based method. Experiments on different datasets are conducted and the comparative results of these extraction methods are presented; Experimental results show that the proposed approach not only improves the precision of metadata extraction, but also enhances the adaptability.