Embeddings

A bag of words gives every type its own axis. Two filings that never share a string have inner product zero, even when one says “impairment” and the other says “write-down” in the same local context. An embedding is a dense vector for the word, learned so that words which keep similar company land near each other. For the equity-event feature, the measurement decision is whether financial tone lives in exact list hits or in contextual similarity. The implementation is a set of four design choices. The validation deliverable is performance on the tone task, not a high score on a generic analogy test.

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