Analogy

We introduce SCAN, a novel analogy dataset spanning dissimilar domains, and show that state-of-the-art pretrained language models struggle with complex analogical reasoning.