Recite, Reconstruct, Recollect: Memorization in LMs as a Multifaceted Phenomenon

TL;DR

This paper proposes a nuanced taxonomy of memorization in language models, categorizing it into recitation, reconstruction, and recollection, and demonstrates how different factors influence memorization likelihood across these categories.

Contribution

It introduces a detailed taxonomy of memorization in language models and develops a predictive model to analyze how various factors affect memorization.

Findings

Different factors influence memorization depending on the category.

A predictive model can effectively classify types of memorization.

Analysis reveals distinct dependencies for each memorization category.

Abstract

Memorization in language models is typically treated as a homogenous phenomenon, neglecting the specifics of the memorized data. We instead model memorization as the effect of a set of complex factors that describe each sample and relate it to the model and corpus. To build intuition around these factors, we break memorization down into a taxonomy: recitation of highly duplicated sequences, reconstruction of inherently predictable sequences, and recollection of sequences that are neither. We demonstrate the usefulness of our taxonomy by using it to construct a predictive model for memorization. By analyzing dependencies and inspecting the weights of the predictive model, we find that different factors influence the likelihood of memorization differently depending on the taxonomic category.