Minimal Pair

Minimal pairs, sentence pairs differing by a single linguistic feature (e.g., a phoneme, morpheme, or word order), are a powerful tool for probing language models' grammatical and semantic understanding. Current research focuses on using minimal pairs to evaluate large language models (LLMs), particularly transformer-based architectures, across various languages and linguistic phenomena, employing diverse evaluation methods ranging from probability comparisons to attention map analysis. These studies reveal insights into LLMs' internal representations, highlighting strengths and weaknesses in their grasp of different linguistic structures and their sensitivity to factors like context and data bias, ultimately informing the development of more robust and linguistically accurate models.