LSTM Train accuracy decreases in training time during the epoch

I’m training an LSTM model for sentiment analysis on a text corpus. There’s a thing that I believe is not normal because I never have seen it in training my models. At the start of the epoch, the training accuracy is something around 80 and then during the epoch, it decreases to 40. And this is happening in every epoch. The model starts with high accuracy and then the accuracy decreases!

I never have seen such behaviour from a model while training. I believe if it’s something is wrong it must be the data because I’ve been using this exact model architecture for topic classification for another dataset and the training behaviour was normal.

My question is this behaviour of the model a normal thing? If it’s not, what is wrong with the data that causes this behaviour?