Abstract
Historical analogies, which compare known past events with contemporary butunfamiliar events, are important abilities that help people make decisions andunderstand the world. However, research in applied history suggests that peoplehave difficulty finding appropriate analogies. And previous studies in the AIcommunity have also overlooked historical analogies. To fill this gap, in thispaper, we focus on the historical analogy acquisition task, which aims toacquire analogous historical events for a given event. We explore retrieval andgeneration methods for acquiring historical analogies based on different largelanguage models (LLMs). Furthermore, we propose a self-reflection method tomitigate hallucinations and stereotypes when LLMs generate historicalanalogies. Through human evaluations and our specially designed automaticmulti-dimensional assessment, we find that LLMs generally have a good potentialfor historical analogies. And the performance of the models can be furtherimproved by using our self-reflection method.