The typical SEC's reaction to a crisis is strengthening disclosure requirements mandating firms to inform investors about their assessment of future contingencies. This should enable investors to monitor risks a firm is facing. However lengthy and complex disclosures -- mostly for dozens or hundreds of firms in an investor's portfolio -- can hardly be processed by a human. Additionally, it is unclear if investors follow regulatory requirements or disclosures are merely boilerplates giving the investor a limited view. It would be informative for investors to see the firm's risk assessment regarding these kinds of risks. To cope with the flood of information, we propose using an unsupervised machine learning algorithm to identify and quantify the risk factor topics discussed in the SEC's 10-K filing. We apply this algorithm (Structural Topic Model, STM) to the Item 1A and Item 7A of the US REITs' 10-K's filings between 2005 and 2019. Our results suggest, that STM is advantageous over the traditional method since it finds clearer and consequently more meaningful risk factor topics beyond the investment foci of REITs. Furthermore, we investigate whether and how the identified topics affect the risk perception of investors after the filing date. We find all three kinds of topics: uninformative topics with no impact (null argument), increasing risk perception topics (divergence argument), and decreasing risk perception topics (convergence argument) -- the majority. Overall, our results suggest that REIT managers use risk disclosures to reveal previously unknown information that has not yet been incorporated into market prices in the short run; but they diminish in the long run.