Interactive Topic Modeling Using GDM

This paper proposes an interactive topic modeling system based on GDM (Geometric Dirichlet Means). Topic modeling is a kind of unsupervised learning and aims to extract topics from a set of documents. The problem is that it is not guaranteed that the obtained results will always satisfy the analyst's intention. To mitigate this problem, the concept of human-in-the-loop can be applied to control the modeling process with the feedback from analysts. The previous study has proposed the concept of Human-in-the-loop topic modeling and introduced seven operations for modifying topic models to support users who are unfamiliar with LDA. Although the result of the qualitative evaluation shows its effectiveness, we suppose that operating the probability space of LDA is difficult to imagine for novices. Aiming to provide analysts with more options for interactively applying various topic models, this paper employs GDM, which is another topic modeling method based on document clustering: we suppose that manipulating document space is easier to imagine than manipulating probability space. The proposed system implements the same seven operations as the existing study based on LDA. In addition, add document operation is newly introduced, taking advantage of the GDM's high affinity with document clustering. Furthermore, a prototype interface is implemented using multiple views, such as a scatter plot, parallel coordinates, and bar charts, to provide quantitative feedback of the topic-document and topic-word relations. A qualitative evaluation is conducted with the implemented prototype interface, and the result shows that test participants felt that they could get the intended and satisfying results in many cases. It is also observed that the coherence of topics tended to be improved with the feedback from the test participants, especially when using the add document operation.