BeliefBank: Adding Memory to a Pre-Trained Language Model for a Systematic Notion of Belief

Abstract

Although pretrained language models (PTLMs) contain significant amounts of world knowledge, they can still produce inconsistent answers to questions when probed, even after specialized training. As a result, it can be hard to identify what the model actually "believes" about the world, making it susceptible to inconsistent behavior and simple errors. Our goal is to reduce these problems. Our approach is to embed a PTLM in a broader system that also includes an evolving, symbolic memory of beliefs -a BeliefBank -that records but then may modify the raw PTLM answers. We describe two mechanisms to improve belief consistency in the overall system. First, a reasoning component -a weighted MaxSAT solver -revises beliefs that significantly clash with others. Second, a feedback component issues future queries to the PTLM using known beliefs as context. We show that, in a controlled experimental setting, these two mechanisms result in more consistent beliefs in the overall system, improving both the accuracy and consistency of its answers over time. This is significant as it is a first step towards PTLM-based architectures with a systematic notion of belief, enabling them to construct a more coherent picture of the world, and improve over time without model retraining.

Exposing and Addressing Cross-Task Inconsistency in Unified Vision-Language Models

“…[7] evaluate and improve factual consistency of pre-trained LMs across paraphrasings of factual statements. [15] consider the responses of a pre-trained LM to a stream of questions, and evaluate and improve the consistency and accuracy of its answers over time. [16] collect counterfactual instances to evaluate the overreliance of NLP models on spurious attributes.…”

“…In computer vision, cross-task consistency has been of some interest for classical tasks [26], while in natural language processing past work has studied consistency between tasks like question-answering [15]. However, in vision-and-language research, much work has focused on…”

GLUECons: A Generic Benchmark for Learning under Constraints

“…Selected Tasks GLUECons contains tasks ranging over five different types of problems categorized based on the type of available knowledge. This includes 1) Classification with label dependencies: Mutual exclusivity in multiclass classification using MNIST (LeCun et al 1998) and Hierarchical image classification using CIFAR 100 (Krizhevsky and Hinton 2009), 2) Self-Consistency in decisions: What-If Question Answering (Tandon et al 2019), Natural Language Inference (Bowman et al 2015), BeliefBank (Kassner et al 2021), 3) Consistency with external knowledge: Entity and Relation Extraction using CONLL2003 (Sang and De Meulder 2003), 4) Structural Consistency: BIO Tagging, 5) Constraints in (un/semi)supervised setting: MNIST Arithmetic and Sudoku. These tasks either use existing datasets or are extensions of existing tasks, reformulated so that the usage of knowledge is applicable to them.…”

“…We use off-the-shelf ILP tools that perform an efficient search and offer a natural way to integrate constraints. However, constraints should be converted to a linear form to be able to exploit these tools Kordjamshidi, Roth, and Wu 2015;Kordjamshidi et al 2016).…”

A Review on Language Models as Knowledge Bases

“…Similar to the efforts done for overcoming lack of consistency under different paraphrases, Hase et al (2021) add another loss term to their objective function to minimize the error across entailed data. On the other hand, Kassner et al (2021) use a feedback mechanism that issues relevant information from a symbolic memory of beliefs as input to the system during test-time in order to improve consistency under entailment.…”