Context-based communication services analyze user data and offer new and novel services that enhance end user unified communication experience. These services rely on data analysis and machine learning techniques to predict user behavior. In this paper we look at topic modeling as an unsupervised learning tool to categorize user communication data for retrieval. However, modeling topics based on user communication data, such as emails, meetings, invites, etc, poses several interesting challenges. One challenge is that user communication, even for a single topic, varies with the current context of the participating users. Other challenges include low lexical content and high contextual data in communication corpus. Hence, relying primarily on lexical analysis could result in inferior topic models. In this paper, we look at this problem of modeling topics for documents based on user communication. First, we use Latent Dirichlet Allocation (LDA) for extracting topics. LDA models documents as a mixture of latent topics where each topic consists of a probabilistic distribution over words. Then we use a technique that overlays a user-relational model over the lexical topic model generated by LDA. In this paper, we present our work and discuss our results.