Text generation with lstm

I am getting predicted next word as an output from my last dense layer which size is equals to vocabulary size. It works fine however i do not understand how it works. I do not use softmax. I am just getting the max value out of it.