# Why categorical cross entropy loss is not correlated with NLP scores?

I'm training a deep network for image captioning which is consist of one CNN and three GRUs. During training epoch by epoch model loss (categorical cross entropy) decreases but when I'm measuring bleu,METEOR,ROUGE,CIDEr and SPICE scores,I get best ones in the first epoch that has worst loss. I don't get why this is happening? And if categorical cross entropy is not a suitable loss function for autoencoder then what should I use instead?