ABSTRACT
You can use deep learning to build classification models to predict labels or categorical variables from a data set, including data sets that include text. Dense neural networks are the most straightforward network architecture that can be used to fit classification models for text features and are a good bridge for understanding the more complex model architectures that are used more often in practice for text modeling. These models have many parameters compared to the models we trained in earlier chapters, and require different preprocessing than those models. We can tokenize and create features for modeling that capture the order of the tokens in the original text. Doing this can allow a model to learn from patterns in sequences and order, something not possible in the models we saw in 6 and Chapter 7. We gave up some of the fine control over feature engineering, such as hand-crafting features using domain knowledge, in the hope that the network could learn important features on its own. However, feature engineering is not completely out of our hands as practitioners, since we still make decisions about tokenization and normalization before the tokens are passed into the network.
