Extracting Semantic Roles from a Model of Eventualities

Sylvie RattéACL 1991 · DOI

Abstract

This paper proposes a formal method for defining semantic roles by modeling events as dynamic processes using Petri net theory. By representing participants as tokens moving through state transitions, the author provides a solution to the linguistic "expansion/contraction dilemma," offering a minimal yet expressive framework for event representation.

Introduction

In linguistic theory, semantic roles (e.g., Agent, Theme, Goal) describe the relationship between a participant and a verb. However, researchers have long faced a dilemma: defining roles too precisely leads to an unmanageable explosion of categories, while defining them too broadly results in a loss of predictive power. This paper argues that the issue stems from a focus on static properties rather than the dynamic mechanics of events. By adopting a "localist" approach rooted in Petri net theory, the author proposes defining roles based on how participants trigger or are affected by changes in state.

Methodology: Petri Nets for Eventualities

The proposed model represents an event (an "eventuality") as a Petri net consisting of:

  • Places (S): Representing locations or states (e.g., "the crate is not on the boat").
  • Transitions (T): Representing the actions or changes that move participants between states.
  • Tokens: Representing the participants (entities or properties) involved.
  • Flow Relations (F): Defining the connections between states and transitions.

In this framework, a semantic role is not just a label; it is defined by the participant's interaction with the "firing" of the net. For instance, a Theme is modeled as a token moving between places, while a Cause is identified as the entity responsible for triggering a transition.

Case Study: The Locative Alternation

The model's utility is demonstrated through the "locative alternation," a classic linguistic puzzle involving verbs like load. Consider the following sentences:

  1. Max loads bricks into the truck.
  2. Max loads the truck with bricks.

The Petri net formalism captures the subtle semantic difference between these two forms. In the first case, the focus is on the change of state of the bricks, specifically the quantity of bricks moving to a final state.

Petri net representation focusing on the quantity of bricks moving into the truck.

In the second case, the focus shifts to the change of state of the truck, emphasizing its capacity as it moves to a "loaded" state.

Petri net representation focusing on the truck's capacity as it becomes loaded.

By varying which participant's state transition is central to the net, the model explains why different syntactic structures are chosen to describe the same physical event.

Key Insights and Results

The paper demonstrates that a small set of primitives (Places, Transitions, Tokens) can represent complex semantic distinctions. Key results include:

  • Operationalizing Conceptual Semantics: The model provides a formal, operational interpretation of Ray Jackendoff’s conceptual semantics.
  • Causality as a Trigger: Causality is modeled by identifying the "first entity" responsible for a transition, allowing for recursive event definitions.
  • Role Clusters: The framework naturally accounts for "role clusters," where a single participant holds multiple roles, as a byproduct of the model's structural simplicity.

Conclusion and Future Work

While the current model is abstract, it successfully schematizes roles like Source and Goal as initial and final states. Future research will focus on adding more semantic content to handle non-physical predicates and testing whether the model can accurately predict syntactic forms—explaining why a language chooses a specific subject or object for a given event.

Summary updated

BibTeX
@inproceedings{ratte1991extracting,
  author = {Sylvie Ratté},
  title = {{Extracting Semantic Roles from a Model of Eventualities}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1991},
  doi = {10.3115/981344.981390},
}
Figure