基于编码器-解码器的机器翻译

Machine Translation with Encoder-Decoders

The encoder-decoder with cross-attention

Beam search and evaluation

Beam search and chrF in NumPy

Which set of typological statements about translation divergences is correct?

In the decoder's cross-attention layer, where do the queries, keys, and values come from?

A decoder's first-token probabilities are fog $0.6$ and mist $0.4$. After fog, the next token is bank with $0.5$ then end with $1.0$; after mist, the next token is shoal with $0.95$ then end with $1.0$. Compute the probability of greedy's complete output and of the beam's best output, then give the beam path's probability.

Why do beam search implementations divide each hypothesis's log probability by its length before comparing finished candidates?

The reference is 'red fishing boat' and a candidate is 'red fishing boot'. With $k = 2$, chrF ignores spaces, averages unigram and bigram precision into chrP and recall into chrR, then computes $\text{chrF}_2 = \frac{5 \cdot P \cdot R}{4P + R}$. Compute chrP, rounding to three decimals.