Maximum and minimum values of the reconstructed signals by autoencoder are far from the original signal

Question

I want to use autoencoder for sentences. Words in sentences are converted to vectors using Bert embedding. Then I normalized data to [-1,1]. Then normalized embedded sentences a