Visual Classifier Training for Text Document Retrieval

Florian Heimerl, Steffen Koch, Harald Bosch, Thomas Ertl
Performing exhaustive searches over a large number of text documents can be tedious, since it is very hard to formulate search queries or define filter criteria that capture an analyst's information need adequately. Classification through machine learning has the potential to improve search and filter tasks encompassing either complex or very specific information needs, individually. Unfortunately, analysts who are knowledgeable in their field are typically not machine learning specialists. Most classification methods, however, require a certain expertise regarding their parametrization to achieve good results. Supervised machine learning algorithms, in contrast, rely on labeled data, which can be provided by analysts. However, the effort for labeling can be very high, which shifts the problem from composing complex queries or defining accurate filters to another laborious task, in addition to the need for judging the trained classifier's quality. We therefore compare three approaches for interactive classifier training in a user study. All of the approaches are potential candidates for the integration into a larger retrieval system. They incorporate active learning to various degrees in order to reduce the labeling effort as well as to increase effectiveness. Two of them encompass interactive visualization for letting users explore the status of the classifier in context of the labeled documents, as well as for judging the quality of the classifier in iterative feedback loops. We see our work as a step towards introducing user controlled classification methods in addition to text search and filtering for increasing recall in analytics scenarios involving large corpora.