论文标题
多语言模型中句法协议神经元的因果分析
Causal Analysis of Syntactic Agreement Neurons in Multilingual Language Models
论文作者
论文摘要
结构性探测工作已经找到了预训练的语言模型中潜在句法信息的证据。但是，这些分析的大部分集中在单语模型上，并且对多语言模型的分析采用了相关方法，这些方法被探测任务的选择混淆了。在这项研究中，我们有因果探究多语言语言模型（XGLM和多语言BERT）以及各种语言的单语BERT模型；我们通过对神经元激活进行反扰动并观察对模型主体 - 动词一致性概率的影响来做到这一点。我们观察到在模型中以及在何种程度上用每种语言编码句法一致性。我们发现自动回归多语言语言模型中的语言之间的显着神经元重叠，但没有掩盖的语言模型。我们还发现了两个不同的层面效应模式和两个不同的神经元集用于句法一致性的集合，具体取决于主题和动词是否被其他令牌分开。最后，我们发现语言模型的行为分析很可能低估了态度信息对句法信息的敏感性。
Structural probing work has found evidence for latent syntactic information in pre-trained language models. However, much of this analysis has focused on monolingual models, and analyses of multilingual models have employed correlational methods that are confounded by the choice of probing tasks. In this study, we causally probe multilingual language models (XGLM and multilingual BERT) as well as monolingual BERT-based models across various languages; we do this by performing counterfactual perturbations on neuron activations and observing the effect on models' subject-verb agreement probabilities. We observe where in the model and to what extent syntactic agreement is encoded in each language. We find significant neuron overlap across languages in autoregressive multilingual language models, but not masked language models. We also find two distinct layer-wise effect patterns and two distinct sets of neurons used for syntactic agreement, depending on whether the subject and verb are separated by other tokens. Finally, we find that behavioral analyses of language models are likely underestimating how sensitive masked language models are to syntactic information.
