Mallen et al. (2023)

When Not to Trust Language Models: Investigating Effectiveness of Parametric and Non-Parametric Memories

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 9802-9822

Find It

DOI

Abstract

Despite their impressive performance on diverse tasks, large language models (LMs) still struggle with tasks requiring rich world knowledge, implying the difficulty of encoding a wealth of world knowledge in their parameters. This paper aims to understand LMs’ strengths and limitations in memorizing factual knowledge, by conducting large-scale knowledge probing experiments on two open-domain entity-centric QA datasets: PopQA, our new dataset with 14k questions about long-tail entities, and EntityQuestions, a widely used open-domain QA dataset. We find that LMs struggle with less popular factual knowledge, and that retrieval augmentation helps significantly in these cases. Scaling, on the other hand, mainly improves memorization of popular knowledge, and fails to appreciably improve memorization of factual knowledge in the tail. Based on those findings, we devise a new method for retrieval-augmentation that improves performance and reduces inference costs by only retrieving non-parametric memories when necessary.

Reference

Mallen, A., Asai, A., Zhong, V., Das, R., Khashabi, D., & Hajishirzi, H. (2023, July). When not to trust language models: Investigating effectiveness of parametric and non-parametric memories. In Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: Long papers) (pp. 9802-9822).

What Rests on This

2 mechanisms cite this source. Each finding below is quoted as it appears on that mechanism's own page.

Semantic Knowledge Memory Systems

Measured in AI

Facts held in a model's training weights track how often they appear online: across 10 models and 14,000 questions, accuracy on the 4,000 least-popular facts stayed at 15 to 19% even as models grew, while retrieval, looking facts up in an external store and placing them in the conversation, let a far smaller model beat GPT-3 on those facts; no human baseline was run

Forgetting Memory Processes

Measured in AI

Using monthly Wikipedia page views as a stand-in for how often an entity comes up, ten language models were much less accurate on facts about less popular entities, and scaling up mainly improved recall of popular ones