Chapter 7. Classification and Regression
The most common machine learning tasks performed on documents are classification and regression. From determining insurance billing codes for a clinical note (classification) to predicting the popularity of a social media post (regression), most document-level machine learning tasks fall into one of these categories, with classification being the much more common of the two.
When beginning a machine learning task, it is very informative to try and manually label some documents, even if there are already labels in the data set. This will help you understand what content in the language of the documents can be used in your task. When labeling, note what you look for. For example, particular words or phrases, certain sections of the document, and even document length can be useful.
In a chapter about classification and regression, you might expect most of the discussion to be about different modeling algorithms. With NLP, most of the work is in the featurization. Many of the general techniques for improving models will work with NLP, assuming you have created good features. We will go over some of the considerations for tuning modeling algorithms, but most of this chapter focuses on how to featurize text for classification and regression.
We’ll discuss the bag-of-words approach, regular expression-based features, and feature selection. After this, we will talk about how to iterate when building a model on text data.
Let’s load and process ...