Modeling Procedural State Changes over Time with Probabilistic Soft Logic

Robust natural language understanding involves the automatic extraction and representation of entities, events, and states from unstructured text. However, a significant portion of the knowledge required for human-level understanding is implicit in the text and can only be accessed through inference. In this work, we employ Probabilistic Soft Logic (PSL) as a framework for leveraging common-sense knowledge to support natural language understanding over procedural texts. Under this framework, we combine logical consistency constraints with succinct representations of commonsense knowledge to probabilistically model entity-centric stative information over time. We demonstrate the feasibility of using PSL to represent procedural stative knowledge through a scalability assessment over an in-house, multi-domain, synthetic dataset.