As discussed earlier, we can consider the problem of named entity recognition a sequence problem. A common deep learning approach, followed by most NLP systems, is to use RNNs. However, before we decide on the architecture for our RNN model, we need to think about how we will be providing our input and processing the output from our model. 

As our input data is in the form of words, we need a dense vector representation for each word, . Such a representation is also known as a word embedding; commonly available, pretrained word embeddings include Word2vec, GloVe, and fastText. Such pretrained word vectors provide very good semantic representations for each word, without the need to train them for every task. As there are quite a few entities that do not have pretrained word vectors, we can extract character-level vector representations, to take occurrences like hyphenated words or words that begin with a capital letter into consideration. This provides the model with valuable information about the entity discussed in the current context.

As we have decided to use a variant of RNNs, such as LSTM, to obtain a semantic representation of our input that is present in its context, we will obtain a vector representation for each word, in order to make a prediction about the entity.