Open lan2720 opened 6 years ago
I am looking at the same issue. I believe it is because the word embedding is limited to the vocab size. These embeddings are updated during training, therefore the words in the article (fed into the encoder) must be derived from the embedding matrix. Whereas, the target words are not in the path of back propagation, and can point to the extended words, which are simply indices, with no vector representation.
Hi, I have a question about the data preprocessing. I mean this line: https://github.com/abisee/pointer-generator/blob/master/batcher.py#L71 Here you create
enc_input_extend_vocab
and modifytarget
, butdec_input
unchanged. Why didn't you replace the UNK in dec_input with oovs? Is there any consideration?