#decomposable-attention

Question 1: Single choice

mediumAI-checkedNLP Tasks

The decomposable attention model for NLI aligns tokens with $e_{ij} = f(\mathbf{a}_i)^\top f(\mathbf{b}_j)$ , compares each token with its soft-aligned counterpart, and then sums the comparison vectors over each sentence. Premise: "the dog bit the man". Hypothesis: "the man bit the dog". What goes wrong?

No step sees word order, so the pair looks like two identical sentences and is likely called entailment1
The $m \times n$ attention scores use too much memory for sentences this short2
The GloVe vectors of "dog" and "man" are too similar for the model to align them3
The model compares the premise with the hypothesis, but it never compares the hypothesis with the premise4

#decomposable-attention

Timed test