# Label Smoothing Improves Neural Source Code Summarization

**Authors:** Sakib Haque, Aakash Bansal, Collin McMillan  
**arXiv:** [2303.16178](https://arxiv.org/abs/2303.16178)  
**Date:** 2023-03-28  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2303.16178v1/

## Key points

- Authors explore the use of label smoothing as a regularization technique for neural networks in source code summarization
- Label smoothing has been effective in language generation tasks but optimal parameters have not been reported for source code summarization
- Authors conduct experiments to determine good parameters for label smoothing and make recommendations for its use
- Different levels of smoothing affect model performance and moderate levels generally yield the best results
- Incorporating label smoothing can help prevent overconfidence in neural models and improve their ability to generate accurate natural language descriptions of source code behavior.

## Summary

In their paper titled "Label Smoothing Improves Neural Source Code Summarization," authors Sakib Haque, Aakash Bansal, and Collin McMillan explore the use of label smoothing as a regularization technique for neural networks in the context of source code summarization. They note that while label smoothing has been shown to be effective in language generation tasks, its optimal parameters have not yet been reported for source code summarization. To address this gap, the authors conduct experiments to determine good parameters for label smoothing and make recommendations for its use. Specifically, they demonstrate how different levels of smoothing affect model performance and show that moderate levels of smoothing generally yield the best results. Their findings suggest that incorporating label smoothing can help prevent overconfidence in neural models and improve their ability to generate accurate natural language descriptions of source code behavior.