A Property Induction Framework for Neural Language Models

TL;DR

The paper introduces a two-stage framework to study property induction in neural language models, first training models to judge concept–property statements and then examining their generalization to novel properties via domain adaptation. Using a refined property-norm dataset and nonce properties, the authors show that large LMs acquire and deploy generalized property knowledge, achieving strong property-judgment performance. They further show that LM representations exhibit a taxonomic bias, generalizing novel properties more strongly within taxonomic categories than outside, an effect only partially explained by property-overlap statistics. The findings illuminate how conceptual knowledge and inductive biases arise in text-trained models and offer a flexible approach to diagnose and explore them beyond traditional syntactic or surface-level analyses.

Abstract

To what extent can experience from language contribute to our conceptual knowledge? Computational explorations of this question have shed light on the ability of powerful neural language models (LMs) -- informed solely through text input -- to encode and elicit information about concepts and properties. To extend this line of research, we present a framework that uses neural-network language models (LMs) to perform property induction -- a task in which humans generalize novel property knowledge (has sesamoid bones) from one or more concepts (robins) to others (sparrows, canaries). Patterns of property induction observed in humans have shed considerable light on the nature and organization of human conceptual knowledge. Inspired by this insight, we use our framework to explore the property inductions of LMs, and find that they show an inductive preference to generalize novel properties on the basis of category membership, suggesting the presence of a taxonomic bias in their representations.

Figures (4)

• Figure 1: (A) Property Judgment Stage describing the training of the property judgment model $\phi$ to make judgments of truth on sentences expressing concept-property assertions. Sentences created using the concept ($\mathcal{C}$) and property ($\mathcal{P}$) data collected by devereux2014centre; (B) Depiction of the Induction Stage, in this case, for testing the generalization from robin to all birds. Here, $\mathcal{A} = \{\textsc{robin}\}$, $\mathcal{G} = \{\textsc{albatross}, ..., \textsc{ostrich}\}$, and the novel property being generalized is "can dax."
• Figure 2: (A): Results from Taxonomic Generalization experiment showing generalization scores ($\mathbb{G}$) between "Within-category" and "Outside-category" generalization sets across various number of adaptation concepts. (B): Same depiction as (A) but with the the effect of property-overlaps between adaptation and generalization concepts removed. Cases where the difference between $\mathbb{G}$-scores of within versus outside category sets are not significant under $\alpha = .05$ are marked with a '$\blacktriangle$'
• Figure 3: Caption
• Figure 4: Caption