GLML: Annotating Argument Selection and Coercion

James Pustejovsky, Jessica Moszkowicz, Olga Batiukova, Anna RumshiskyIWCS 2009 · OpenReview

Abstract

James Pustejovsky, Jessica Moszkowicz, Olga Batiukova, Anna Rumshisky. Proceedings of the Eight International Conference on Computational Semantics.

2009.

Introduction

In standard Natural Language Processing (NLP), annotation schemes like PropBank or FrameNet are typically used to identify surface-level roles—the "who did what to whom." While effective for straightforward sentences, these systems often struggle with semantic shifts such as metonymy or type coercion. For instance, in the sentence "The Boston office called," a computer might identify "office" as a location, whereas a human instinctively understands it refers to the people within that office.

The authors identify a critical gap: current tools do not record the "compositional history" of a sentence. They fail to capture the "semantic gymnastics" required to make a sentence make sense. Without this data, machine learning models struggle to generalize how verbs select their arguments across different contexts.

Methodology and Specification

To address this, the authors introduce Generative Lexicon Markup Language (GLML), an XML-based framework designed to capture how language shifts and stretches word meanings. Based on Generative Lexicon (GL) theory, GLML tracks three specific types of semantic interactions:

  1. Argument Selection: Distinguishing between Pure Selection (where an argument fits the verb's requirements perfectly) and Coercion (where an argument's type must be shifted, such as treating a location as a human agent).
  2. Qualia Selection: Using "Qualia Structure" to explain how adjectives modify nouns (e.g., a "sharp knife" refers to its physical form, while a "hunting knife" refers to its purpose).
  3. Dot Objects: Handling complex types like "book," which can be both a physical object and information, by annotating which "facet" is being used.

The system architecture involves a two-phase process: a construction phase for building sense inventories and an annotation phase where humans use a specialized interface to answer targeted questions about the text.

The GLML annotation interface facilitates the categorization of arguments and semantic types, as seen here for the verb "deny."

Key Insights and Results

A central contribution of GLML is the focus on compositional history over simple surface labeling. Recording the shift from a LOCATION to a HUMAN provides a roadmap for how language actually works. To keep the task manageable, the authors implemented a "shallow" type system (including categories like HUMAN, ARTIFACT, EVENT, and LOCATION). This flat list allows annotators to work quickly without navigating complex ontological hierarchies.

The primary result is the GLML specification itself, which demonstrates that complex linguistic phenomena can be successfully represented in a machine-readable format. The framework successfully distinguishes between different "facets" of dot objects, such as identifying that in "a delicious lunch," the modifier applies to the FOOD component, whereas in "a long lunch," it applies to the EVENT component.

Conclusion and Future Work

While GLML provides a robust methodology for capturing selection and coercion, the authors acknowledge that they have only scratched the surface. The current version is restricted to a specific set of relations, and a more fine-grained typology is needed for the future. Additionally, while the framework is currently focused on English, future work will involve expanding to other languages to identify language-specific constructions. Scaling this manual annotation process to massive datasets also remains a key challenge for future research.

Summary updated

BibTeX
@inproceedings{pustejovsky2009glml,
  author = {James Pustejovsky and Jessica Moszkowicz and Olga Batiukova and Anna Rumshisky},
  title = {{GLML: Annotating Argument Selection and Coercion}},
  booktitle = {IWCS},
  year = {2009},
  url = {https://openreview.net/forum?id=vC8aO6rjJ88},
}
Figure