Moving the Eiffel Tower to ROME: Tracing and Editing Facts in GPT
Moving the Eiffel Tower to ROME: Tracing and Editing Facts in GPT
复制标题
将埃菲尔铁塔移至罗马：在 GPT 中跟踪和编辑事实
DOI：
--
复制
发表时间：
2021
期刊：
影响因子：
--
通讯作者：
Kenton Lee
中科院分区：
文献类型：
--
作者：
David Bau;Steven Liu;Tongzhou Wang;Jun;Tom Brown;Benjamin Mann;Nick Ryder;Jared D Subbiah;Prafulla Kaplan;A. Dhariwal;P. Neelakantan;Girish Shyam;Amanda Sastry;Sandhini Askell;Ariel Agarwal;Herbert;Gretchen Krueger;T. Henighan;R. Child;Aditya Ramesh;Daniel M. Ziegler;Jeffrey Wu;Clemens Winter;Chris Hesse;Mark Chen;Eric Sigler;Mateusz Litwin;S. Gray;B. Chess;Christopher Clark;Sam Berner;Alec McCandlish;Ilya Radford;Sutskever Dario;Amodei;Damai Dai;Li Dong;Y. Hao;Zhifang Sui;Nicola De Cao;Wilker Aziz;Ivan Titov. 2021;Edit;J. Devlin;Ming;Kenton Lee
我们研究了自回归trans-002前语言模型中face -001知识回忆的机制。为此，我们开发了一种方法来识别能够改变模型的实际预测的神经元激活。在GPT-2中，这揭示了006有两组不同的神经元，我们假设这两组神经元的大小分别对应于知道一个抽象事实和008说一个具体的词。基于这一见解，我们提出了ROME，这是一种简单有效的一级模型编辑方法，用于重写自回归语言模型中的抽象事实。为了验证，我们介绍了013c counter F ACT，这是一个超过20,000个014和可重写事实的数据集，以及用于fa-015的工具，有助于对编辑质量进行敏感测量。与先前发表的知识编辑方法相比，ROME实现了更好的gen-018的实体化和特异性。019
We investigate the mechanisms underlying fac-001 tual knowledge recall in auto-regressive trans-002 former language models. To this end, we de-003 velop a method for identifying neuron activa-004 tions that are capable of altering a model’s fac-005 tual predictions. Within GPT-2, this reveals 006 two distinct sets of neurons that we hypothe-007 size correspond to knowing an abstract fact and 008 saying a concrete word, respectively. Based 009 on this insight, we propose ROME, a simple 010 and efficient rank-one model editing method 011 for rewriting abstract facts in auto-regressive 012 language models. For validation, we introduce 013 C OUNTER F ACT , a dataset of over twenty thou-014 sand rewritable facts, as well as tools to fa-015 cilitate sensitive measurements of edit quality. 016 Compared to previously-published knowledge 017 editing methods, ROME achieves superior gen-018 eralization and specificity. 019
DOI：
10.1162/tacl_a_00324
发表时间：
2020-01-01
影响因子：
10.9
作者：
Jiang, Zhengbao;Xu, Frank F.;Neubig, Graham
通讯作者：
Neubig, Graham