Deep Learning Architectures:

$($a$)$ Given a task to develop a language model for French speaking, Peter plans to train a bidirectional LSTM model for it based on the corpus. Before he starts, what suggestions do you have for him about the design? $(2$ points$)$

$($b$)$ In attention mechanism, why are the word embeddings from input not directly used for key, query, and values? $(1$ point$)$

$($c$)$ after pre$-$training a model $($e$.$g$.,$ BERT$),$ sometimes we only use the encoder part and sometimes we only use the decoder part. Please explain briefly in what scenario$($s$)$ we use encoder and in what scenario$($s$)$ we use decoder. $(1$ point$)$