Summary

We have covered quite a lot of ground in this chapter. NER and its importance in the industry were explained. To build NER models, BiLSTMs and CRFs are needed. Using BiLSTMs, which we learned about in the previous chapter while building a sentiment classification model, we built a first version of a model that can label named entities. This model was further improved using CRFs. In the process of building these models, we covered the use of the TensorFlow DataSet API. We also built advanced models for CRF mode by building a custom Keras layer, a custom model, custom loss function, and a custom training loop.

Thus far, we have trained embeddings for tokens in the models. A considerable amount of lift can be achieved by using pre-trained embeddings. In the next chapter, we'll focus on the concept of transfer learning and the use of pre-trained embeddings like BERT.