Machine Learning-based software systems (ML-based systems) have been increasingly developed for various domains, such as medicine, smart cities, and the automotive sector. These systems have posed significant challenges for Requirements Engineering (RE), including for requirements specification, while ensuring critical qualities such as reliability. The main problem addressed in this work is the difficulty practitioners have faced in specifying the requirements for these systems. Additionally, effective approaches (e.g., methods, techniques, processes, or guidelines) for this still lack consensus. The main objective of this work is to present and discuss current trends in approaches to requirements specification for ML-based systems. We observe this research field is very new, with weak collaboration with industrial environments. Novel approaches or adaptations of existing ones were proposed, but several possibilities for future research still exist; hence, we also discuss some of the main ones.