This paper seems to be officially "arXiv 2014" since the authors literally ask us to cite someone else's work. Interesting ... but the Tensorflow Tutorial for RNNs appears to refer to this paper, so I'll be reading it anyway.
This is a natural language processing paper. This includes language modeling, meaning that the model (i.e. the RNN) has to be able to determine the probability of the next word in a sentence given the previous word. This is the job of the LSTM. There are also speech recognition, machine translation, and image caption generation experiments (wow!).
Their main contribution seems to be to show how to correctly apply dropout to LSTMs. I guess it was tricky before?
The dataset for language modeling comes from the Penn Treebank Corpus.