LABAS-TS A system for assisting labeling of training sets for text classification

Most text classification techniques rely on the existence of training data sets that are required to build models. However, in many text classification projects, the availability of previously labeled texts is not frequent due to differences in language (e.g. Spanish), domain (e.g. healthcare) and regional or institutional written culture (e.g. specific hospital). In order to contribute to dealing with this problem, this paper presents LABAS-TS, a web-enabled system for assisting the open, collaborative labeling of training sets for text classification. LABAS-TS is framed within a named entity recognition approach that identifies important entities from a domain-specific corpus, based on gazetteers, and uses a language specific sentence analyzer that extracts the portions of text that should be annotated. LABAS-TS was evaluated in the generation of training data sets to classify whether an electronic health record text contains a diagnosis, a test or a procedure, and demonstrated its utility in reducing the required time for building a reliable training set, with an average of eleven seconds between two labels.