Labeled Interactive Topic Models

Anonymous

Abstract: Topic models help users understand large document collections;however, topic models do not always find the ``right'' topics.While classical probabilistic and anchor-based topic models haveinteractive variants to guide models toward better topics, suchinteractions are not available for neural topic models such as theembedded topic model (ETM).We correct this lacuna by adding an intuitive interactionto ETM: users can label a topic with a word, and topics areupdated so that the topic words are close to the label. This allowsuser to refine topics based on their information need.We evaluate our method through a human study, where users can relabeltopics to find relevant documents. We find that using our method, user labeling improves document rankscores, helping to find more relevant documents to a givenquery when compared to no user labeling.

Paper Type: long

Research Area: Dialogue and Interactive Systems