Measuring and Improving Consistency in Pretrained Language Models

2021/02/01 by Yanai Elazar, Elazar, Yanai, Nora Kassner +11 · 29 citations
Computer Science · #Computation and Language (cs.CL) #FOS: Computer and information sciences #Multimodal Machine Learning Applications #Natural Language Processing Techniques #Topic Modeling

paper · pdf · doi:10.48550/arxiv.2102.01017

openalex publication_date 2021/02/01 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28

Abstract

Consistency of a model -- that is, the invariance of its behavior under meaning-preserving alternations in its input -- is a highly desirable property in natural language processing. In this paper we study the question: Are Pretrained Language Models (PLMs) consistent with respect to factual knowledge? To this end, we create ParaRel, a high-quality resource of cloze-style query English paraphrases. It contains a total of 328 paraphrases for 38 relations. Using ParaRel, we show that the consistency of all PLMs we experiment with is poor -- though with high variance between relations. Our analysis of the representational spaces of PLMs suggests that they have a poor structure and are currently not suitable for representing knowledge robustly. Finally, we propose a method for improving model consistency and experimentally demonstrate its effectiveness.

Citations

• e-SNLI: Natural Language Inference with Natural Language Explanations
• VQA: Visual Question Answering
• Pretrained Encyclopedia: Weakly Supervised Knowledge-Pretrained Language Model
• ALBERT: A Lite BERT for Self-supervised Learning of Language Representations
• Language Models are Few-Shot Learners
• Enriching a Model's Notion of Belief using a Persistent Memory

Cited by

• Same Question, Different Answers: Evaluating LLM Reliability Beyond Accuracy
• FIBER: A Multilingual Evaluation Resource for Factual Inference Bias
• ART: Adaptive Response Tuning Framework -- A Multi-Agent Tournament-Based Approach to LLM Response Optimization
• CrossCheck-Bench: Diagnosing Compositional Failures in Multimodal Conflict Resolution
• On the Brittleness of LLMs: A Journey around Set Membership
• Metamorphic Testing of Large Language Models for Natural Language Processing
• PRISMM-Bench: A Benchmark of Peer-Review Grounded Multimodal Inconsistencies
• HACK: Hallucinations Along Certainty and Knowledge Axes
• Temporal Referential Consistency: Do LLMs Favor Sequences Over Absolute Time References?
• When Facts Change: Probing LLMs on Evolving Knowledge with evolveQA
• Annotation-Efficient Universal Honesty Alignment
• Measuring the Effect of Disfluency in Multilingual Knowledge Probing Benchmarks
• Harmonizing Diverse Models: A Layer-wise Merging Strategy for Consistent Generation
• Rewriting History: A Recipe for Interventional Analyses to Study Data Effects on Model Behavior
• Improving Consistency in Retrieval-Augmented Systems with Group Similarity Rewards
• Knowledge Editing with Subspace-Aware Key-Value Mappings
• Concept Unlearning in Large Language Models via Self-Constructed Knowledge Triplets
• Quantifying Self-Awareness of Knowledge in Large Language Models
• Self-Improvement of Language Models by Post-Training on Multi-Agent Debate
• Do All Autoregressive Transformers Remember Facts the Same Way? A Cross-Architecture Analysis of Recall Mechanisms
• Analyzing the Instability of Large Language Models in Automated Bug Injection and Correction
• Memorization ≠ Understanding: Do Large Language Models Have the Ability of Scenario Cognition?
• Scaling Laws for Task-Stratified Knowledge in Post-Training Quantized Large Language Models
• LM Agents May Fail to Act on Their Own Risk Knowledge
• Online Anti-sexist Speech: Identifying Resistance to Gender Bias in Political Discourse
• A Dual-Axis Taxonomy of Knowledge Editing for LLMs: From Mechanisms to Functions
• EMSEdit: Efficient Multi-Step Meta-Learning-based Model Editing
• Explainability Through Systematicity: The Hard Systematicity Challenge for Artificial Intelligence
• Integrating LLM in Agent-Based Social Simulation: Opportunities and Challenges

Related

• Accurate, yet inconsistent? Consistency Analysis on Language Understanding Models
• Measuring Reliability of Large Language Models through Semantic Consistency
• How often are errors in natural language reasoning due to paraphrastic variability?
• The Queen of England is not England's Queen: On the Lack of Factual Coherency in PLMs
• Dissecting Paraphrases: The Impact of Prompt Syntax and supplementary Information on Knowledge Retrieval from Pretrained Language Models