On the Acquisition of Lexical Entries: The Perceptual Origin of Thematic Relations

James PustejovskyACL 1987 · DOI

Introduction: The Mapping Problem

A central puzzle in language acquisition is the "mapping problem": how does a learner determine which part of a sentence (syntax) corresponds to which participant in an event (semantics)? When a child hears "Mary hit Bill," they must identify that Mary is the one performing the action (the Agent) and Bill is the one affected (the Theme).

Existing models often struggle to explain how children transition from concrete perceptual observations—such as seeing an object move—to abstract linguistic expressions involving mental states or metaphorical movement. This paper proposes that the meanings of thematic roles like Agent, Theme, and Goal are not arbitrary but are grounded in a structured perceptual hierarchy.

Figure 1: Mapping thematic roles (Agent, Theme) to syntactic positions (Subject, Object).

Theoretical Framework

The authors introduce a computational model grounded in two primary theories: an Extended Aspect Calculus and a Markedness Theory of thematic relations.

Extended Aspect Calculus

This framework uses first-order logic to decompose events into three semantic fields:

  1. Causal: Predicates for causes ($C_1$), effects ($C_2$), and instruments ($I$).
  2. Spatial: Predicates for location ($L$) and themes ($Th$).
  3. Aspectual: A temporal logic that divides events into beginning ($t_1$), middle ($t_2$), and end ($t_3$) intervals.

As shown in the formal representations below, a verb like run is decomposed into a causal agent ($C_1$) and a theme ($Th$) moving through time, while give involves a more complex interaction between a causer, a theme, and a location (the recipient).

Figure 6: Formal decomposition of the verbs 'run' and 'give' using the Extended Aspect Calculus notation.

Markedness Theory

Rather than treating thematic roles as a flat list, the authors propose a partial ordering. The "Theme" is considered the most basic (unmarked) role. Other roles, such as "Goal" or "Agent," are more "marked" (complex) and their presence typically depends on the existence of a Theme. This hierarchy provides a learning bias that constrains the infinite possibilities of word meanings to only those that are "human-like."

The TULLY System

The theories are implemented in a system called TULLY (written in ZETALISP). TULLY takes two inputs: a simulated event sequence (representing output from a vision system) and a sentence describing that scene.

TULLY employs two key learning techniques:

  • Constraint Propagation: The system generalizes a verb's meaning by relaxing specific features, such as transitioning from a "tangible" object to an "abstract" one.
  • Thematic Decoupling: Initially, a learner might "bake" specific information into a verb (e.g., assuming the verb cut requires a knife). As the learner encounters more diverse examples, TULLY "decouples" these specific requirements to allow for broader usage.

Key Insights and Results

The paper provides a qualitative proof-of-concept demonstrating that abstract thematic relations (like "experiencer") originate from tangible perceptual predicates. For instance, we understand the abstract concept of "having an idea" because we first mastered the spatial concept of "having a book."

TULLY successfully replicates stages of child language acquisition by following the Subset Principle: starting with the most specific interpretation possible and only generalizing when encountering new positive evidence. The model demonstrates how a system can move from a concrete spatial verb like have (physical possession) to an abstract one (possessing a "problem") by simply toggling an [±abstract] feature on the Theme argument.

Limitations and Future Work

While TULLY offers a compelling model for semantic mapping, it relies on highly idealized, hand-coded visual input. Integrating this framework with raw computer vision remains a significant challenge. Furthermore, while the system handles positive evidence effectively, the authors acknowledge that the innate linguistic knowledge required to make this work is substantial. Future research aims to scale this vocabulary and integrate the semantic model with a robust syntactic parser to create a more holistic language-learning agent.

Summary updated

BibTeX
@inproceedings{pustejovsky1987acquisition,
  author = {James Pustejovsky},
  title = {{On the Acquisition of Lexical Entries: The Perceptual Origin of Thematic Relations}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1987},
  doi = {10.3115/981175.981199},
}
Figure