In many application scenarios such as target tracking., one is confronted with the problem of suitably fusing both hard data (e.g., radar, acoustic) and soft data (e.g., text from blogs, or witness statements). Complicating this challenge are the imperfections inherent within soft text data which possess semantic uncertainty, in addition to uncertainty regarding the source reliability and statement credibility. While natural language processing can convert text into uncertainty-augmented first order logic, the question of how to incorporate such soft data models into estimation and tracking remains an important fundamental issue. In this paper, we develop a Bayesian filtering and tracking framework for incorporating both hard and soft data, demonstrating how the probability posterior can be decomposed into a product of combining functions over subsets of the state and measurement variables. This combining function approach offers a framework for the development and incorporation of more sophisticated uncertainty modeling and tracking/estimation models.