Surprisal Scoring

Overview

This study is currently calibrating. Content will document:

1. Reference networks — Small World of Words (SWOW) and LLM World of Words (LWOW) association databases
2. Transition probability — computing association strength from frequency data
3. Surprisal calculation — negative log-probability of observed transitions
4. Interpretation — distinguishing creative (high surprisal, high relevance) from random (high surprisal, low relevance)

Definition

$$\text{surprisal}(w_1 \to w_2) = -\log P(w_2 | w_1)$$

Where $P(w_2 | w_1)$ is the conditional probability of associating $w_2$ given $w_1$, estimated from reference network frequencies.

Reference Networks

Network	Source	Coverage
SWOW-EN	De Deyne et al. (2019)	Human association norms, ~12K cues
LWOW	Abramski et al. (2025)	LLM association norms for comparison

Theoretical Foundation

Surprisal distinguishes:

• Low surprisal + high relevance = conventional, expected associations
• High surprisal + high relevance = creative, unexpected but connected
• High surprisal + low relevance = random, noise

---

This methodology is under development. Check back for updates.