Summary of Has This Fact Been Edited? Detecting Knowledge Edits in Language Models, by Paul Youssef et al.
Has this Fact been Edited? Detecting Knowledge Edits in Language Models
by Paul Youssef, Zhixue Zhao, Christin Seifert, Jörg Schlötterer
First submitted to arxiv on: 4 May 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 A novel task is proposed to detect edited knowledge in language models, driven by the desire to increase users’ trust and provide transparency. The goal is to classify whether a generated output is based on edited or pre-trained knowledge. To instantiate this task, four KEs, two LLMs, and two datasets are used. Features such as hidden state representations and probability distributions are proposed for detection. A simple AdaBoost classifier establishes a strong baseline, requiring only limited data and maintaining performance in cross-domain settings. However, distinguishing edited from unedited but related knowledge proves challenging, highlighting the need for further research. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Language models can learn outdated or incorrect information during pre-training. To fix this, “knowledge editing methods” (KEs) can be used to update the model’s knowledge. But KEs can also be used to add misinformation and toxic content. To solve this problem, scientists propose a new task: detecting edited knowledge in language models. This means identifying whether an output is based on old information or new information added later. The researchers use four different ways to edit knowledge, two large language models (LLMs), and two datasets to test their approach. They suggest using the model’s internal representations and probabilities as clues to detect edited knowledge. Their results show that a simple computer program can be trained to do this task well, even when it’s applied to new, unseen data. |
Keywords
» Artificial intelligence » Probability
