# 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/2110.08387v1/

## Key points

- Study on leveraging self-generated knowledge prompts for commonsense reasoning
- Approach of generating knowledge statements directly from a language model with a generic prompt format
- Significantly improved performance of off-the-shelf and finetuned language models on commonsense reasoning tasks
- Improvements observed in numerical commonsense, general commonsense, and scientific commonsense benchmarks
- Importance of symbolic knowledge representation in neural reasoning processes highlighted
- Advancements in augmenting questions with model-generated text discussed to boost performance on commonsense reasoning tasks
- Techniques such as self-talk for eliciting clarifications and contrastive explanations for prompting inference models explored
- Promise shown in enhancing model performance by providing additional context and information during the reasoning process

## Summary

In the study "Generated Knowledge Prompting for Commonsense Reasoning," conducted by Jiacheng Liu, Alisa Liu, Ximing Lu, Sean Welleck, Peter West, Ronan Le Bras, Yejin Choi, and Hannaneh Hajishirzi, the researchers explore the effectiveness of leveraging knowledge elicited from language models themselves to enhance performance on commonsense reasoning tasks. By generating knowledge statements directly from a language model with a generic prompt format and selecting the knowledge that maximizes prediction probability, the researchers found that this approach significantly improves the performance of both off-the-shelf and finetuned language models on various commonsense reasoning tasks. The research demonstrates that incorporating self-generated knowledge can lead to improvements in numerical commonsense (NumerSense), general commonsense (CommonsenseQA 2.0), and scientific commonsense (QASC) benchmarks. The findings highlight the importance of symbolic knowledge representation in neural reasoning processes and showcase how a model's predictions can be enhanced when utilizing its own generated knowledge. Furthermore, the study discusses recent advancements in augmenting questions with model-generated text to boost performance on commonsense reasoning tasks. Techniques such as self-talk for eliciting clarifications to concepts in questions and contrastive explanations for prompting inference models with generated explanations are explored. These methods show promise in enhancing model performance by providing additional context and information during the reasoning process. Overall, the research underscores the significance of leveraging self-generated knowledge prompts and augmenting questions with generated text to improve performance on commonsense reasoning tasks across various domains. The findings contribute valuable insights into enhancing neural reasoning processes through symbolic knowledge representation and task-specific augmentation techniques.