How Language Models Process Subject and Object Control: On Promise as a Subject Control Verb

This squib examines how neural language models process subject and object control. Processing these two types of control requires an understanding of the structural dependency between an overt determiner phrase (DP) and a non-overt argument (PRO). By using surprisal as a complexity metric, the study tests whether two representative encoder models - BERT and RoBERTa - can distinguish subject control from object control. The results showed that, while the models succeeded in processing object control, they failed to process subject control. The surprisal estimates, which are expected to be higher for unlikely or unacceptable sentences, were elevated for subject inputs even in subject control sentences. This contrast between subject and object control is attributed to the statistical rarity of subject control in the corpus data, particularly regarding the subject control verb promise. Thus, model parsing relies more on statistical patterns than genuine syntactic analysis.