Semi-supervised learning models for document classification: A systematic review and meta-analysis

The continuous increase of digital documents on the web creates the need to search for information patterns that allow the categorization of organizational documents to generate knowledge in an institution. An Artificial Intelligence technique for this purpose is text classification, it for its application uses labels (previously categorized documents) with supervised (with labels) or unsupervised (without labels) training models. Both traditional models with their advantages and disadvantages have been joined into semi-supervised models that extract the best qualities of each one, however, the labeling process involves resources and time that try to be optimized to improve classification accuracy. An analysis of the different semi-supervised models would show us the advantages of their training and the way how the structure of each of them affects the accuracy of their classification. In the present study, a classification structure of the semi-supervised models in the classification of documents is proposed to analyze their qualities and categorization process, through an SLR (Revision of systematic literature) that extracts performance metrics from the identified studies to perform a meta-analysis through forest plots. To define the search strategy for studies, the PICOC (Population, Intervention, Comparison, Outputs, Context) method has been used, it is supported by the research question defines a search string, which has allowed the collection of 228 research, these are filtered with the PRISMA declaration method and the determination of exclusion criteria, in this way 35 researches are selected for the present study. The analysis of the selected studies identifies a structure for the different semi-supervised learning models, and a scheme of their work process is obtained, it has been used to extract advantages, disadvantages, and performance metrics. Through a meta-analysis with forest diagrams, the classification accuracy performance of the researches in each learning model is evaluated, determining as results that regardless of the characteristics of its process, active learning (0.89) and assembled learning (0.83) present the best performance levels.