Event Semantic Knowledge in Procedural Text Understanding

Proceedings of the The 12th Joint Conference on Lexical and Computational Semantics (*SEM 2023), pages 388–398 July 13-14, 2023 ©2023 Association for Computational Linguistics Event Semantic Knowledge in Procedural Text Understanding Ghazaleh Kazeminejad University of Colorado Boulder ghazaleh.kazeminejad@colorado.edu Martha Palmer University of Colorado Boulder martha.palmer@colorado.edu Abstract The task of entity state tracking aims to auto- matically analyze procedural texts – texts that describe a step-by-step process (e.g. a baking recipe). Specifically, the goal is to track vari- ous states of the entities participating in a given process. Some of the challenges for this NLP task include annotated data scarcity and anno- tators’ reliance on commonsense knowledge to annotate implicit state information. Zhang et al. (2021) successfully incorporated com- monsense entity-centric knowledge from Con- ceptNet into their BERT-based neurosymbolic architecture. Since English mostly encodes state change information in verbs, we attempted to test whether injecting semantic knowledge of events (retrieved from the state-of-the-art Verb- Net parser) into a neural model can also im- prove the performance on this task. To achieve this, we adapt the methodology introduced by Zhang et al. (2021) for incorporating symbolic entity information from ConceptNet to the in- corporation of VerbNet event semantics. We evaluate the performance of our model on the ProPara dataset (Mishra et al., 2018). In ad- dition, we introduce LEXIS, our purely sym- bolic model for entity state tracking that uses a simple set of case statements, and is informed mostly by linguistic knowledge retrieved from various computational lexical resources. Our approach is inherently domain-agnostic, and our model is explainable and achieves state-of- the-art results on the Recipes dataset (Bosselut et al., 2017). 1 Introduction Language understanding in humans requires at least the knowledge of the semantics of events and entities. One needs to know the sequences of subevents that together make up a ‘throwing’ event, as well as the causal and temporal relationships between the subevents that distinguish a ‘throw- ing’ event from a ‘pouring’ event, or a ‘running’ event. Furthermore, reasoning about entities that are participating in these events requires a deep understanding of the properties of an entity. It is the distinction between such entity properties that enables us, for example, to distinguish between ‘throwing a ball’ vs. ‘throwing a Molotov cock- tail’. In contrast to humans, many high-performing NLP models do not depend on explicit knowledge of events and entities to process natural language; rather, they rely on the surface forms and patterns of word co-occurances in colossal amounts of lan- guage data to learn the mechanics of language as well as the interpretation of linguistic forms. Since human knowledge and reasoning capabilities bene- fit from knowledge of events and entities, we sug- gest that a neural model may also benefit from such explicit symbolic knowledge. This requires suc- cessful incorporation of such symbolic knowledge into a subsymbolic system. Explicit semantic knowledge, such as entity knowledge extracted from ontologies, has often been used in the field of natural language ground- ing, where the connection between natural lan- guage and …

Content truncated to 500 words