As general perception over training and validation accuracy is that if training accuracy is high and validation accuracy is marginally low, then it is most probably over fitting.
Consider a case of simple LSTM model in which training accuracy after 200 Epocs is 0.91 but validation accuracy is 0.08.
What is possible very wrong? The training data is not good for that problem ? or model is weak and I should add more hidden layer. The model picture is attached below. 
