Summary of Source-aware Training Enables Knowledge Attribution in Language Models, by Muhammad Khalifa et al.
Source-Aware Training Enables Knowledge Attribution in Language Models
by Muhammad Khalifa, David Wadden, Emma Strubell, Honglak Lee, Lu Wang, Iz Beltagy, Hao Peng
First submitted to arxiv on: 1 Apr 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI)
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 The paper investigates the problem of intrinsic source citation in large language models (LLMs), where they learn to cite the pretraining source supporting a generated response. The authors propose a training recipe called source-aware training, which involves associating unique document identifiers with knowledge and then teaching the LLM to cite a supporting source when prompted. Experiments on synthetic data show that this approach enables faithful attribution to pretraining data without significantly impacting perplexity. The paper highlights the importance of pretraining data augmentation in achieving attribution. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper looks at how large language models can learn from lots of knowledge during training, but often don’t know where they got that knowledge from. It wants to help these models “cite” their sources so people can understand where they got their ideas from. To do this, the authors come up with a special way of training the model called source-aware training. This involves teaching the model to link specific pieces of knowledge to the documents it came from, and then telling it to say what document it got its ideas from when asked. The authors tested this approach on fake data and found that it works without making the model’s predictions worse. |
Keywords
» Artificial intelligence » Data augmentation » Perplexity » Pretraining » Synthetic data
