# Knowledge Infused Decoding

**Authors:** Ruibo Liu, Guoqing Zheng, Shashank Gupta, Radhika Gaonkar, Chongyang Gao, Soroush Vosoughi, Milad Shokouhi, Ahmed Hassan Awadallah  
**arXiv:** [2204.03084](https://arxiv.org/abs/2204.03084)  
**Date:** 2022-04-06  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2204.03084v1/

## Key points

- Authors: Ruibo Liu, Guoqing Zheng, Shashank Gupta, Radhika Gaonkar, Chongyang Gao, Soroush Vosoughi, Milad Shokouhi, Ahmed Hassan Awadallah
- Introduces Knowledge Infused Decoding (KID) algorithm for generative language models
- Addresses limitations of pre-trained LMs in recalling factually correct knowledge within specific contexts
- KID interacts with externally created knowledge trie and is continuously updated using reinforcement learning
- Evaluated on six diverse knowledge-intensive NLG tasks with strong performance in few-shot scenarios
- Human evaluation confirms KID enhances generation of more relevant and factual language compared to baseline models
- Code for implementing KID available on GitHub at https://github.com/microsoft/KID
- Presented at ICLR 2022 and contributes insights into improving generative LMs through dynamic external knowledge infusion during decoding

## Summary

In their paper titled "Knowledge Infused Decoding," authors Ruibo Liu, Guoqing Zheng, Shashank Gupta, Radhika Gaonkar, Chongyang Gao, Soroush Vosoughi, Milad Shokouhi, and Ahmed Hassan Awadallah introduce a novel decoding algorithm called Knowledge Infused Decoding (KID) for generative language models (LMs). The study addresses the limitations of pre-trained LMs in recalling factually correct knowledge within specific contexts. This often leads to counterfactual or hallucinatory generation in knowledge-intensive natural language generation tasks. Existing solutions typically involve modifying pre-training or task fine-tuning objectives to incorporate knowledge. However, these methods require additional training or architectural adjustments. <br>
 This memory interacts with an externally created knowledge trie and is continuously updated as a knowledge-aware constraint using reinforcement learning. The effectiveness of KID was evaluated on six diverse knowledge-intensive NLG tasks. In these tasks, Particularly strong performance was observed in few-shot scenarios compared to seven related knowledge-infusion techniques. Human evaluation confirmed that KID enhances the generation of more relevant and factual language based on input context when compared to multiple baseline models. Additionally,

The code for implementing KID is openly available on GitHub at https://github.com/microsoft/KID. This research was presented at ICLR 2022 and contributes valuable insights into improving the performance of generative LMs through dynamic external knowledge infusion during decoding processes.