Information Locality as an Inductive Bias for Neural Language Models

TL;DR

This paper introduces a new information-theoretic measure called m-local entropy to analyze how neural language models utilize local context, revealing their sensitivity to local statistical structures similar to human language processing.

Contribution

The paper proposes a quantitative framework and a novel measure, m-local entropy, to investigate the inductive biases of neural language models regarding local language structure.

Findings

Languages with higher m-local entropy are harder for neural LMs to learn.

Neural LMs are highly sensitive to local statistical structures.

The framework enables controlled studies of language model biases.

Abstract

Inductive biases are inherent in every machine learning system, shaping how models generalize from finite data. In the case of neural language models (LMs), debates persist as to whether these biases align with or diverge from human processing constraints. To address this issue, we propose a quantitative framework that allows for controlled investigations into the nature of these biases. Within our framework, we introduce $m$-local entropy$\unicode{x2013}$an information-theoretic measure derived from average lossy-context surprisal$\unicode{x2013}$that captures the local uncertainty of a language by quantifying how effectively the $m-1$ preceding symbols disambiguate the next symbol. In experiments on both perturbed natural language corpora and languages defined by probabilistic finite-state automata (PFSAs), we show that languages with higher $m$-local entropy are more difficult for Transformer and LSTM LMs to learn. These results suggest that neural LMs, much like humans, are highly sensitive to the local statistical structure of a language.