重新思考上下文化：通过重新解读注意力头通道

中文摘要 AI 辅助

上下文化是语言建模的核心操作，它在单词之间传递信息以构建特定于句子的单词表示。以往的研究主要从个体单词和注意力头的角度研究上下文化，将其视为一个不断增长的离散字典，缺乏对其整体行为的全局视角。因此，我们提出一个一般性原则：从全局来看，我们发现并估计不同单词携带不同数量的信息，而信息量较少的单词倾向于吸收更多的上下文信息。具体而言，这些低信息单词并非均匀地吸收上下文单词，更细粒度的选择性能够实现更精确的路由，以促进匹配单词之间的信息传递。此外，为了找出导致这种处理的机制，我们将注意力头重新解释为由其奇异向量门控的通道，并发现：（1）这些奇异向量指向信息量更大的单词的隐藏状态，使得这些单词能够更强烈地将自身信息写入其他单词，从而充当信息源，反之亦然；（2）这些奇异向量同样可以被视为隐藏状态特征，使得能够超越先前启发式的头发现方法，对注意力头进行自动解释，同时将头嵌入连续空间，而非将其视为离散、独立的字典条目。

英文摘要

Contextualization, the core operation of language modeling, transmits information across words to build sentence-specific word representations. Prior works mainly study contextualization, focusing on individual words and attention heads as a growing discrete dictionary, lacking a global view of their general behavior. Therefore, we propose a general principle: Globally, we find and estimate that different words carry different amounts of information, and less-informative words tend to absorb more contextual information. Specifically, these low-information words do not absorb contextual words uniformly, and finer-grained selectivity enables more precise routing to promote information transmission between matched words. Moreover, to find what mechanism causes such processing, we reinterpret attention heads as channels gated by their singular vectors and find that: (1) these singular vectors point to the hidden states of more informative words, allowing such words to write their information to others more strongly to act as information sources, and vice versa; and (2) these singular vectors can be viewed equally as hidden state features, enabling automated interpretation of attention heads beyond prior heuristic head discovery, also embedding heads into a continuous space rather than treating them as discrete, independent dictionary entries.

发表机构

• RIKEN（理化学研究所）
• Tohoku University（东北大学）
• New York University（纽约大学）
• JAIST（北陆先端科学技术大学院大学）
• MBZUAI（穆罕默德·本·扎耶德人工智能大学）

机构由 AI 辅助整理，请以论文原文为准。

补充信息

备注

47 pages, 81 figures, 4 tables