Hierarchical topic map generation for exploratory browsing
Shengliang Dai · Illinois Digital Environment for Access to Learning and Scholarship (University of Illinois at Urbana-Champaign) · 2017
This thesis proposes a novel model for automatically generate topic map for a document corpus with no supervision. We extend a previous approach to discovery of lexical relations from text data to construct a hierarchy of topics. Given a collection of documents, we will generate a set of topics on the fly which will help the user to efficiently navigate through the corpus space and finally land upon the desired document. We use Latent Dirichlet Allocation to generate the top level topics and then leverage paradigmatic and syntagmatic relations between words to construct the hierarchy. We characterize each topic in the hierarchy by a single phrase. Our topic map captures the requirements of user while he/she navigates through the corpus space. Instead of a rigid tree structure, we define links on topic map such that they take user to next desired finer level/related topic based on the history of already visited nodes in map/regions in the corpus. Experiments on DBLP titles datasets show that our topic map can be used very effectively and intuitively by the user to reach to the desired document.