In the era of the Social Web, crowdfunding has become an increasingly more important channel for entrepreneurs to raise funds from the crowd to support their startup projects. Previous studies examined various factors such as project goals, project durations, and categories of projects that might influence the outcomes of the fund raising campaigns. However, textual information of projects has rarely been studied for analyzing crowdfunding successes. The main contribution of our research work is the design of a novel text analytics-based framework that can extract latent semantics from the textual descriptions of projects to predict the fund raising outcomes of these projects. More specifically, we develop the Domain-Constraint Latent Dirichlet Allocation (DC-LDA) topic model for effective extraction of topical features from texts. Based on two real-world crowdfunding datasets, our experimental results reveal that the proposed framework outperforms a classical LDA-based method in predicting fund raising success by an average of 11% in terms of F1 score. The managerial implication of our research is that entrepreneurs can apply the proposed methodology to identify the most influential topical features embedded in project descriptions, and hence to better promote their projects and improving the chance of raising sufficient funds for their projects.