Labeled Interactive Neural Topic Models: No Longer Take It or Leave It
Abstract: Topic models are essential for understanding document collections but often fail to identify all relevant topics. While classical probabilistic and anchor-based models offer interactive features for user guidance, such capabilities are missing in neural topic models. To address this, we introduce a user-friendly interaction for neural topic models, allowing users to assign word labels to topics. This interaction updates the model, aligning topic words with the given label. Our approach covers two types of neural topic models: those with trainable topic embeddings that evolve during training, and those with embeddings integrated post-training. We develop an interactive interface for user engagement and re-labeling of topics. A human study shows that user labeling improves document rank scores on average by at least 30\% and helps users find more relevant documents compared to no user labeling.
Paper Type: Long
Research Area: Information Retrieval and Text Mining
Research Area Keywords: topic model, interactive, neural topic models
Contribution Types: Publicly available software and/or pre-trained models
Languages Studied: english
Submission Number: 1195
Loading