Overview

A lexicon writes the mapping from words to tone by hand. This topic learns a mapping, or borrows one from a pretrained model, and then asks what an annotation from a large language model is allowed to mean. The equity-event feature is still a measurement of disclosure tone. Classification, transformers, and model-assisted labels are three implementations. None of them is a demonstrated return.

By the end of this note you should be able to:

• define supervised classification as a learned map from a document representation to a researcher-specified label;
• state what a contextualised embedding adds to a fixed word vector, and what this extract does not specify about the transformer equations;
• place a zero-shot language-model label on a human annotation team, with agreement and accuracy checked against people.

1. Contents
2. The big picture