When a recipe says, "Cut the celery into sticks. Add to the bowl," a human reader instantly understands that "Add" refers to the newly created celery sticks, not the original whole stalk. This mental leap—inferring the existence and properties of objects that are not explicitly named—is known as accommodation.
In natural language processing, tracking these "hidden" state changes is a fundamental challenge. Procedural texts, such as scientific protocols, medical records, and recipes, are notoriously underspecified. They rely on the reader’s world knowledge to bridge the gap between literal words and the physical reality they describe.
The Problem of Implicit Arguments
The primary hurdle in understanding instructions is that the most important information is often left unsaid. Edward Kim and Xiao Liu have highlighted that models struggle when the subject of an operation is omitted. If a text doesn't explicitly mention "the mixture" after a stirring step, a model may lose track of the entity entirely.
This isn't just a matter of missing words; it's a matter of evolving states. Aïssatou Diallo notes that current models often default to listing original ingredients rather than recognizing the emergence of a new, transformed state, such as a dough or a chemical complex. To solve this, systems must perform "accommodation," a term explored by Barbara Di Eugenio and Bonnie Webber, where the system infers necessary environmental states and new entities (like "the chopped celery") to maintain a coherent model of the world.
From Text Parsing to World Simulation
To move beyond simple pattern matching, researchers have proposed a shift from "text-centric" to "world-centric" understanding. Instead of just analyzing the syntax of a sentence, these models attempt to simulate the causal effects of actions.
Antoine Bosselut introduced Neural Process Networks (NPN), which treat actions as functional operators that transform entity states. In this framework, an action like "cut" isn't just a verb; it's a function that changes a "whole" state to a "separated" state. Similarly, Peter Clark argues for "machine reading as model construction," where the system builds a predictive simulation—a mental "participant grid"—to track how entity properties like location and temperature change over time, even when those changes are never explicitly stated.
Verbs as Anchors for State Change
If the world state is a sequence of frames, verbs are the transitions between them. Early work by Lynette Hirschman recognized that verbs signifying a change of state act as linguistic anchors, allowing systems to reconstruct a chronological timeline.
Modern approaches like Lexis, developed by Ghazaleh Kazeminejad and Martha Palmer, leverage the structured logic of VerbNet to automatically trigger these updates. By decoding the underlying semantic predicates of a verb—for instance, knowing that "absorb" implies a change in location and existence—systems can identify state changes zero-shot, without needing massive amounts of labeled data.
Injecting Constraints and Commonsense
Without a grounding in physical reality, neural models often make nonsensical predictions, such as an object being moved before it has been created. To prevent this, Niket Tandon developed ProStruct, which injects "hard" physics-based constraints and "soft" commonsense priors into the tracking process.
This grounding is essential for "open-vocabulary" tracking, where the possible states are not pre-defined. Heng Ji and colleagues have shown that integrating external knowledge bases is necessary to bridge the gap between literal text and the implicit physical reasoning humans take for granted. By using entity memory and conditioned prediction, these models can maintain a consistent history of an entity's status across long, complex instructions.