arXivDaily arXiv每日学术速递 周一至周五更新
北美退伍军人节(Veterans Day),12 日无更新
arXiv 2407.10488cs.CLcs.AI
收藏

CLIP如何以及在何处处理否定?

How and where does CLIP process negation?

Utrecht University(乌得勒支大学)

Vincent Quantmeyer, Pablo Mosteiro, Albert Gatt

AI总结:

本研究针对多模态模型否定理解的内部机制问题,借助可解释性方法分析CLIP的文本编码器,定位否定处理模块并分析注意力头作用,同时验证方法可迁移性、揭示VALSE基准的固有局限。

摘要:

目前已有多种基准被提出用于测试预训练视觉与语言(VL)模型的语言理解能力。本研究基于VALSE基准(Parcalabescu等人,2022)中的存在任务开展研究,用该任务测试模型对否定的理解能力——这是多模态模型中一个尤为值得关注的问题。然而,这类VL基准虽可用于衡量模型性能,却无法揭示这些模型在完成此类视觉-语言任务时生成输出的内部过程。我们从日益增多的模型可解释性相关文献中汲取灵感,以解释VL模型在否定理解任务上的行为。具体而言,我们通过深入分析极具影响力的VL模型CLIP(Radford等人,2021)的文本编码器来探究这些问题。我们定位了编码器中处理否定的部分,并分析了注意力头在该任务中的作用。本研究的贡献有三点:我们证明了语言模型可解释性文献中的方法(如因果追踪)可迁移应用于多模态模型与任务;我们针对CLIP在VALSE存在任务上如何处理否定提供了具体洞见;我们还指出了VALSE数据集作为语言理解基准存在的固有局限性。

英文摘要:

Various benchmarks have been proposed to test linguistic understanding in pre-trained vision \& language (VL) models. Here we build on the existence task from the VALSE benchmark (Parcalabescu et al, 2022) which we use to test models' understanding of negation, a particularly interesting issue for multimodal models. However, while such VL benchmarks are useful for measuring model performance, they do not reveal anything about the internal processes through which these models arrive at their outputs in such visio-linguistic tasks. We take inspiration from the growing literature on model interpretability to explain the behaviour of VL models on the understanding of negation. Specifically, we approach these questions through an in-depth analysis of the text encoder in CLIP (Radford et al, 2021), a highly influential VL model. We localise parts of the encoder that process negation and analyse the role of attention heads in this task. Our contributions are threefold. We demonstrate how methods from the language model interpretability literature (such as causal tracing) can be translated to multimodal models and tasks; we provide concrete insights into how CLIP processes negation on the VALSE existence task; and we highlight inherent limitations in the VALSE dataset as a benchmark for linguistic understanding.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑