# Generated Knowledge Prompting for Commonsense Reasoning

**Authors:** Jiacheng Liu, Alisa Liu, Ximing Lu, Sean Welleck, Peter West, Ronan Le Bras, Yejin Choi, Hannaneh Hajishirzi  
**arXiv:** [2110.08387](https://arxiv.org/abs/2110.08387)  
**Date:** 2021-10-15  
**Summary page:** https://www.summarizepaper.com/en/paper/2110.08387v3/

## Key points

- Incorporation of external knowledge in commonsense reasoning has been a topic of debate
- Generated knowledge prompting is a novel approach that enhances pretrained sequence models by integrating external knowledge
- This method involves generating knowledge from a language model and providing it as additional input during the question-answering process
- It does not require task-specific supervision or access to structured knowledge bases, unlike traditional methods
- Generated knowledge prompting has shown significant improvements in performance across four commonsense reasoning tasks
- Achieved state-of-the-art results in NumerSense, CommonsenseQA 2.0, and QASC benchmarks
- Existing commonsense reasoning datasets often rely on pre-existing knowledge bases like ConceptNet and ATOMIC
- Domain mismatches can pose challenges when dealing with datasets lacking suitable knowledge bases or queries that do not align with existing domains
- Generated text augmentation during inference has shown promise for enhancing model performance on commonsense reasoning tasks
- Generated knowledge prompting stands out for its simplicity and flexibility compared to other augmentation approaches
- Elicits human-written demonstrations in the style of the task to seamlessly integrate external knowledge without extensive model finetuning
- Sets new benchmarks in commonsense reasoning tasks and highlights the potential of large-scale language models as sources of high-quality external knowledge

## Summary

In the realm of commonsense reasoning, the incorporation of external knowledge has long been a topic of debate. The question remains: can we enhance the capabilities of pretrained sequence models by integrating external knowledge while still maintaining their flexibility? To address this query, a novel approach known as generated knowledge prompting has been developed. This method involves generating knowledge from a language model and providing it as additional input during the question-answering process. Unlike traditional methods that rely on task-specific supervision or access to structured knowledge bases, generated knowledge prompting does not require such constraints. Remarkably, this technique has shown significant improvements in the performance of large-scale pretrained models across four commonsense reasoning tasks. Notably, it has achieved state-of-the-art results in numerical commonsense (NumerSense), general commonsense (CommonsenseQA 2.0), and scientific commonsense (QASC) benchmarks. It is worth noting that many existing commonsense reasoning datasets are derived from pre-existing knowledge bases such as ConceptNet and ATOMIC. While leveraging related knowledge from these sources often leads to performance gains, domain mismatches can pose challenges when dealing with datasets lacking suitable knowledge bases or queries that do not align with existing domains. Generated text augmentation during inference has emerged as a promising strategy for enhancing model performance on commonsense reasoning tasks. Various approaches like self-talk and contrastive explanations have shown success in augmenting questions with model-generated text to improve inference accuracy. However, these methods typically rely on task-specific templates or finetuned generators, limiting their adaptability to new tasks. In contrast, generated knowledge prompting stands out for its simplicity and flexibility. By eliciting human-written demonstrations in the style of the task, this method seamlessly integrates external knowledge without the need for extensive model finetuning. This approach not only sets new benchmarks in commonsense reasoning tasks but also underscores the potential of large-scale language models as valuable sources of high-quality external knowledge. Overall, generated knowledge prompting represents a significant advancement in leveraging language models for enhancing commonsense reasoning capabilities. Its success across multiple datasets highlights the versatility and effectiveness of this approach in improving inference accuracy and setting new standards in the field of natural language processing.