COMPS: Conceptual Minimal Pair Sentences for testing Robust Property Knowledge and its Inheritance in Pre-trained Language Models

Abstract

A characteristic feature of human semantic memory is its ability to not only store and retrieve the properties of concepts observed through experience, but to also facilitate the inheritance of properties (can breathe) from superordinate concepts (ANIMAL) to their subordinates (DOG)-i.e. demonstrate property inheritance. In this paper, we present COMPS, a collection of minimal pair sentences that jointly tests pre-trained language models (PLMs) on their ability to attribute properties to concepts and their ability to demonstrate property inheritance behavior. Analyses of 22 different PLMs on COMPS reveal that they can easily distinguish between concepts on the basis of a property when they are trivially different, but find it relatively difficult when concepts are related on the basis of nuanced knowledge representations. Furthermore, we find that PLMs can demonstrate behavior consistent with property inheritance to a great extent, but fail in the presence of distracting information, which decreases the performance of many models, sometimes even below chance. This lack of robustness in demonstrating simple reasoning raises important questions about PLMs' capacity to make correct inferences even when they appear to possess the prerequisite knowledge.

Dissociating language and thought in large language models: a cognitive perspective

“…For instance, when prompted with "Seinfeld premiered on ____" and "Seinfeld originally aired on ____", they might provide names of different TV networks despite the fact that the prompts have the same semantic content Elazar et al [see also Ravichander et al 2020Ravichander et al 2020. Misra et al [2022] show another interesting failure of even advanced models like GPT-3 on a task where they are asked to reason about properties of novel objects. Models generally learn properties of objects, like that robins can fly and penguins cannot.…”

“…That is, given the example in the preceding sentence as input, the model should be able to know that "I went to a blick" is more probable than "I went to a dax" since blick was used as a noun. They conclude that BERT succeeds partially at this task: it does learn to generalize, but only after repeated examples [but see , Misra et al, 2022 for ways in which the word itself affects compositional ability]. More recent models, such as GPT-3, seem to be able to use a novel word appropriately right away, at least if prompted correctly , McCoy et al, 2021b.…”