Schematic Example of a Sentence-Pair Classification Pipeline

In a text-pair classification pipeline, two texts are formatted into a single input sequence, which includes a special token at the beginning designated for classification. After a Transformer encoder processes this sequence and generates a contextualized hidden state for every token, a single vector must be selected to represent the entire text pair for the final prediction. Which of the following best explains the standard method for selecting this vector and the reasoning behind it?

A language model is tasked with determining if two sentences are semantically equivalent. Arrange the following steps to correctly represent the end-to-end computational pipeline, from preparing the input to generating the final prediction.

Applying the Text-Pair Classification Pipeline