A transformer-based text classifier and a transformer-based named-entity recogniser both sit on top of the same pre-trained encoder, yet they attach different prediction heads. Conceptually, what most fundamentally distinguishes the two task formulations at the output stage?
- AText classification requires a softmax output layer, whereas NER must always use a sigmoid output layer because entities can overlap within a sentence.
- BText classification reads only the final token's hidden state, whereas NER reads only the first token's hidden state.
- CText classification freezes the encoder while NER fine-tunes it, which is why their heads differ.
- DText classification assigns one label to the whole sequence using a single pooled representation, whereas NER assigns a label to every token using each token's own contextual representation. Correct
Why A is wrong: Tempting because activation choice does vary by task, but it is wrong: standard single-type NER uses a per-token softmax over the tag set, and sigmoid is tied to multi-label problems, not to NER inherently.
Why B is wrong: Tempting because both tasks do select specific positions, but it is wrong: classification conventionally pools or uses the prepended classification token, and NER uses every token's state, not just the first.
Why C is wrong: Tempting because freezing strategies exist, but it is wrong: the head difference comes from the granularity of the prediction, not from whether the encoder is frozen, and either task can freeze or fine-tune the encoder.
Why D is correct: Correct: classification is a sequence-level task that consumes one aggregated vector, while NER is a token-level (sequence-labelling) task that emits a label per token position, which is the defining structural difference.