A team is tagging word senses in complete, stored sentences. Nothing arrives live — the whole text is on disk before tagging begins. One sentence is
The bank, after a dispute that dragged on through three separate committee hearings and two unsuccessful appeals from the original owners, was finally closed for urgent repairs.
Read left to right, the model arrives at bank having seen only The. The words that settle its meaning — closed for urgent repairs — begin 22 tokens later.
Their first model was a left-to-right RNN with plain, ungated cells, and it tagged bank as the financial sense. The fix they propose is to make it bidirectional: a second RNN with the same plain ungated cell, reading right to left, with the two hidden states joined together at every position. Everything else stays as it was — the cell, the hidden state size, and a training set of thousands of similarly shaped sentences.
What should the team expect at the position of bank?
Select all that apply.