Summary of On the Noise Robustness Of In-context Learning For Text Generation, by Hongfu Gao et al.
On the Noise Robustness of In-Context Learning for Text Generation
by Hongfu Gao, Feipeng Zhang, Wenyu Jiang, Jun Shu, Feng Zheng, Hongxin Wei
First submitted to arxiv on: 27 May 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Machine Learning (cs.LG)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary Large language models have excelled in downstream tasks through in-context learning (ICL), relying heavily on the quality of demonstrations. Recent studies claim that ICL is robust to noisy text classification examples. However, this work reveals that noisy annotations significantly impede performance on text generation tasks. To overcome this issue, we propose Local Perplexity Ranking (LPR), a simple yet effective approach that replaces noisy candidates with their nearest neighbors likely to be clean. Our method is motivated by analyzing perplexity deviation and decomposing it into inherent and matching components. By decoupling matching perplexity in semantic space, LPR prevents mismatched input-label pairs while preserving original selection methods’ effectiveness. Extensive experiments demonstrate LPR’s efficacy, improving EM scores up to 18.75 on common benchmarks with noisy annotations. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This research paper is about how big language models can learn new tasks by looking at examples of what they should do. Some people thought that even if the examples are a little messy, the model will still learn well. But this study shows that when the examples are really messy, the model actually gets worse! So, we came up with a new way to fix this problem called Local Perplexity Ranking (LPR). It works by picking better examples from those that might be good but not perfect. We tested LPR and it makes the model do much better on hard tasks. This is important because it means we can make language models more reliable. |
Keywords
» Artificial intelligence » Perplexity » Text classification » Text generation
