Assumption-based Argumentation (ABA) provides a structured formalism for modeling human arguments, where contraries play a crucial role in determining attacks and conflicts among assumptions. While recent advances in argumentation mining have begun to bridge the gap between natural language texts and structured argumentation frameworks, the tasks for mining ABA structure are still underexplored, especially the contrary mining task—identifying contraries of assumptions from text. This paper extends prior work on ABA-based argument mining by introducing the first dataset on hotel reviews. Indeed, we construct a novel, fine-grained annotation scheme for contraries within hotel reviews from Booking.com, enriching the existing ABA-annotated corpus. Furthermore, we evaluate the in-context learning abilities of large language models (LLMs), specifically GPT-4o, on the contrary mining task under few-shot settings. Our findings highlight both the challenges and potential of LLMs in learning and generalizing contrariness relations from natural language, marking an important step toward fully automating ABA knowledge construction from text.