The transformer architecture, introduced in "Attention is All You Need" (Vaswani et al., 2017), is particularly effective for language modeling due to its ability to handle long sequences. Unlike RNNs, which struggle with long-term dependencies due to sequential processing, transformers use self-attention mechanisms to process all tokens in a sequence simultaneously, capturing relationships across long distances. NVIDIA's NeMo documentation emphasizes that transformers excel in tasks like language modeling because their attention mechanisms scale well with sequence length, especially with optimizations like sparse attention or efficient attention variants. Option B (embeddings) is a component, not a unique strength. Option C (class tokens) is specific to certain models like BERT, not a general transformer feature. Option D (translations) is an application, not a structural advantage.
References:
Vaswani, A., et al. (2017). "Attention is All You Need."
NVIDIA NeMo Documentation: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/nlp/intro.html