Abstract. At the very beginning of system development, typically only natural-language requirements are documented. As an informal source of information, however, natural-language specifications may be ambiguous and incomplete; this can be hard to detect by means of manual inspection. In this work, we present a formal model, named Data-Flow Reactive System (DFRS), which can be automatically obtained from natural-language requirements that describe functional, reactive and temporal properties. A DFRS can also be used to assess whether the requirements are consistent and complete. We define two variations of DFRS: a symbolic and an expanded version. A symbolic DFRS (s-DFRS) is a concise representation that inherently avoids an explicit representation of (possibly infinite) sets of states and, thus, the state spaceexplosion problem. We use s-DFRS as part of a technique for test-case generation from natural-language requirements. In our approach, an expanded DFRS (e-DFRS) is built dynamically from a symbolic one, possibly limited to some bound; in this way, bounded analysis (e.g., reachability, determinism, completeness) can be performed. We adopt the s-DFRS as an intermediary representation from which models, for instance, SCR and CSP, are obtained for the purpose of test generation. An e-DFRS can also be viewed as the semantics of the s-DFRS from which it is generated. In order to connect such a semantic representation to established ones in the literature, we show that an e-DFRS can be encoded as a TIOTS: an alternative timed model based on the widely used IOLTS and ioco. To validate our overall approach, we consider two toy examples and two examples from the aerospace and automotive industry. Test cases are independently created and we verify that they are all compatible with the corresponding e-DFRS models generated from symbolic ones. This verification is performed mechanically with the aid of the NAT2TEST tool, which supports the manipulation of such models.