Subevent Structure as a Predictor of Entity Identity Change in Procedural Text
Proceedings of the Workshop on Structured Linguistic Data and Evaluation (SLiDE)
Abstract
We test whether the subevent structure encoded in VerbNet-GL predicts entity identity change in procedural text, using only the verb’s lexical specification and no training data. From each VN verb class’s SEMANTICS block we extract an aspectual classification and an I/O count, yielding a predicted dynamic event topology (DET). On the observation side, ten large language models (LLMs) annotate per-entity dynamic object mode (DOM) labels over ∼3100 OpenPI steps, from which we derive observed DET for comparison. The VN-only predictor achieves 67.4% precision (F1 = 0.35) for transformation, showing that formal subevent structure carries genuine predictive signal only for the most common topology, but low performance for other topologies. For events VN predicts as having no result state, only 24% are confirmed as no-change by silver, indicating that the remaining outcomes arise from the argument side of the composition. These results provide empirical evidence that event semantics is distributed across predicate and argument: the VN supplies the subeventual skeleton, but is not sufficient to determine the final outcome.