Enriching a Model's Notion of Belief using a Persistent Memory

TL;DR

This paper introduces a memory-augmented approach to enhance the consistency and accuracy of pretrained language models by recording and reasoning over their beliefs using a BeliefBank and SAT-based mechanisms.

Contribution

It proposes a novel memory component and reasoning mechanisms to improve the global consistency of language model beliefs, a first step towards models with evolving world knowledge.

Findings

Improved model accuracy in controlled experiments

Enhanced consistency among model beliefs

Demonstrated potential for more coherent world understanding

Abstract

Although pretrained language models (PTLMs) have been shown to contain significant amounts of world knowledge, they can still produce inconsistent answers to questions when probed, even after using specialized training techniques to reduce inconsistency. As a result, it can be hard to identify what the model actually "believes" about the world. Our goal is to reduce this problem, so systems are more globally consistent and accurate in their answers. Our approach is to add a memory component -- a BeliefBank -- that records a model's answers, and two mechanisms that use it to improve consistency among beliefs. First, a reasoning component -- a weighted SAT solver -- improves consistency by flipping answers that significantly clash with others. Second, a feedback component re-queries the model but using known beliefs as context. We show that, in a controlled experimental setting, these two mechanisms improve both accuracy and consistency. This is significant as it is a first step towards endowing models with an evolving memory, allowing them to construct a more coherent picture of the world.