Two Passes, One Representation

A named-entity tagger reads a 25-token sentence with a bidirectional LSTM.

One pass runs left to right and one runs right to left, each carrying its own hidden state. At every position the two hidden states are joined into a single representation, and a small classifier reads that representation to choose the tag for that token.

Across the whole sentence, how many numbers does the classifier read in total?

Count the joined representations only — one per position — and give an exact whole number.