The motivation behind training custom models
In this book we so far discussed the features and models that are available in Flair straight out of the box. However, if you work with NLP long enough, you will likely encounter a sequence labeling problem that is complex or specific enough that there will be no pre-trained models available out there. This can happen in either of the following situations:
- The problem you are solving is domain-specific: Sequence taggers such as Named Entity Recognition (NER) or Part-of-Speech (PoS) taggers are usually trained on large, generic corpora that are supposed to represent the general use of a language. But if our problem is domain-specific, it's likely that we will require a custom tagger with domain-specific labels trained on domain-specific corpora.
- A pre-trained model exists but doesn't perform well enough: Every model made available in Flair and the approaches used for training are usually reviewed by other contributors...