The last decade has seen knowledge-seeking and data exploration workflows transition from classical machine learning methods to neural network based methods and, most recently, to pretrained large language models (LLMs). Yet evidence from task-oriented and user evaluations of the effectiveness and usefulness of the new methods remains mixed. We focus on topic models—once a popular data analysis tool in the pre-LLM era—for large-scale corpus analysis, helping researchers quickly visualize and identify the main themes within large collections of documents, and evaluating their effectiveness through real-world, human-in-the-loop validation. There is a also transition from using classic topic models to neural topic models and LLMs for knowledge seeking in the filed of computer science and social science. While NLP researchers keep advancing in techniques and algorithms to try to produce more interpretable and readable knowledge summaries (topics, themes, summaries) and visualization of the large data corpus, there is a significant gap between the method developer community and the user community who use the tools for practical applications. There has been a line of work that develops NLP methods for knowledge exploration, and there has been another line of work that aims to evaluate the effectiveness and usefulness of the tools in practical user applications. Studies show that the automatic evaluations do not necessarily reflect what users prefer in their use cases. In my work, I focus on adding human-in-the-loop for knowledge seeking process to use these tools and develop and evaluate them