FeatureForge: A Novel Tool for Visually Supported Feature Engineering and Corpus Revision
Florian Heimerl, Charles Jochim, Steffen Koch, Thomas Ertl · International Conference on Computational Linguistics · 2012
In many fields of NLP, supervised machine learning methods reach the best performance results. Apart from creating new classification models, there are two possibilities to improve classification performance: (i) improve the comprehensiveness of feature representations of linguistic instances, and (ii) improve the quality of the training gold standard. While researchers in some fields can rely on standard corpora and feature sets, others have to create their own domain specific corpus and feature representations. The same is true for practitioners developing NLP-based applications. We present a software prototype that uses interactive visualization to support researchers and practitioners in two aspects: (i) spot problems with the feature set and define new features to improve classification performance, and (ii) find groups of instances hard to label or that get systematically mislabeled by annotators to revise the annotation guidelines.