An Approach for Evaluating Topic Models for Knowledge Management

Over time, organizations develop an informal history of what works and what does not. Developing a formal history of documented problems is known as a lesson learned. A lesson learned is the process of successfully learning from past mistakes by applying knowledge (gained from experiences) in the present and future. Lessons learned is a subset of knowledge management, which is the process of applying a systematic approach to the gathering, documenting, managing, and communicating of knowledge throughout an organization. A sustained lessons learned implementation process includes the capturing, structuring, management, and dissemination of knowledge. Organizations that implement a sustained knowledge management system or a lessons learned database allowing staff access to the previous lessons documented have an advantage in learning from previous mistakes and avoiding reinventing the wheel. This paper describes an approach using topic models to address the lack of an automatic knowledge management/lessons learned system/database option that removes the need for users to perform manual searching while providing highly relevant results. Topic models are machine learning algorithms created to detect the underlying semantic structures of corpora using Bayesian hierarchical modeling. This research describes the process of evaluating topic models in the information retrieval space to improve relevancy in lessons learned query searches within knowledge management/lessons learned systems/databases. Widely used topic models such as the Latent Semantic Indexing (LSI) and Latent Dirichlet Allocation (LDA) in addition to topic models applied to short text are investigated.