Shortcut Learning of Large Language Models in Natural Language Understanding

Natural Language Understanding (NLU) is a subfield of artificial intelligence that requires computer software to comprehend input in the form of sentences. Representative NLU tasks include natural language inference (NLI), question answering (QA), and reading comprehension, among others. Furthermore, NLU has several real-world applications, including Alexa, Siri, and Google Assistant. The major characteristic of NLU tasks is they are difficult and typically require world knowledge and commonsense reasoning. Recently, large language models (LLMs), such as BERT,8 RoBERTa,19 T5,29 GPT-3,4 have been reported to achieve state-of-the-art performance in a series of high-level NLU tasks.