Chapter 7

Smoothing and Backoff

Add-one smoothing, interpolation, backoff, Good-Turing and Kneser-Ney as historical names, rare events, and sparse-data overconfidence.

What this chapter does

Smoothing teaches how simple models handle missing evidence without pretending unseen means impossible.

Lessons

Read these in order.

Start with the chapter introduction, then move through the topic lessons. The order is chosen so each page can reuse ideas from the pages before it.

  1. 01
    Introduction

    Smoothing and backoff as ways to handle missing evidence in count models.

  2. 02
    The Zero-Probability Problem

    Why unseen continuations should not always be treated as impossible.

  3. 03
    Add-One Smoothing

    Adding one count to each vocabulary item before normalizing.

  4. 04
    Interpolation

    Mixing estimates from several context lengths.

  5. 05
    Backoff

    Using a shorter context when the longer context lacks evidence.

  6. 06
    Good-Turing and Kneser-Ney

    Recognizing historical smoothing ideas for rare and unseen events.

Before moving on

  • Explain why unseen phrases should not always receive zero probability.
  • Compare smoothed and unsmoothed probabilities.
  • Recognize smoothing as evidence management, not a hack.

Where this leads

  • Evaluating Language Models
  • Neural Language Models

Chapter progress