Explain encoder-only, decoder-only, and encoder-decoder transformer architectures and typical NLP tasks for which each is used.

• A. Encoder-Only Like BERT For Comprehension/Embedding; Decoder-Only Like GPT For Text Generation; Encoder-Decoder Like T5 For Translation And Summarization Tasks
• B. Decoder-Only For Language Modeling; Encoder-Only For Translation
• C. Encoder-Decoder For Embeddings; Decoder-Only For Comprehension
• D. All Three Are Used Equally For All NLP Tasks

Answer: (A)

The key idea is matching the architecture to the kind of processing a task needs. An encoder-only model reads the input and builds rich, contextual representations, which makes it ideal for comprehension and embedding tasks—like understanding a sentence, answering questions, or labeling the text—where you care about the meaning encoded in the input. A decoder-only model generates text autoregressively, predicting the next token based on what came before, so it shines at text generation, continuation, and conversational responses. An encoder-decoder model combines both: the encoder turns the input into a contextual representation, and the decoder uses that representation to produce an output sequence, making it well suited for translation, summarization, and other text-to-text transformations.

This aligns with the statement that encoder-only models (like BERT) handle comprehension and embeddings, decoder-only models (like GPT) handle text generation, and encoder-decoder models (like T5) handle translation and summarization.

Other options misstate these roles: translation isn’t typically done with an encoder alone because generating the target text requires a decoding step; embeddings aren’t the primary purpose of an end-to-end encoder–decoder setup; and saying all three are used equally for all NLP tasks ignores the strengths and typical use cases of each architecture.