Preparing SNACS for Subjects and Objects

Adi Shalev, Jena D. Hwang, Nathan Schneider, Vivek Srikumar, Omri Abend, Ari Rappoport2019 · DOI

Abstract

Research on adpositions and possessives in multiple languages has led to a small inventory of general-purpose meaning classes that disambiguate tokens. Importantly, that work has argued for a principled separation of the semantic role in a scene from the function coded by morphosyntax. Here, we ask whether this approach can be generalized beyond adpositions and possessives to cover all scene participants—including subjects and objects—directly, without reference to a frame lexicon. We present new guidelines for English and the results of an interannotator agreement study.

Introduction

In Natural Language Processing (NLP), identifying "who did what to whom" is typically the domain of Semantic Role Labeling (SRL). However, traditional SRL systems like PropBank or FrameNet are heavily lexicon-driven, relying on extensive dictionaries of verbs and their specific arguments. This dependency introduces significant challenges, notably the Out-of-Vocabulary (OOV) problem and a conceptual gap between core arguments (subjects/objects) and peripheral arguments (prepositional phrases). This research explores an expansion of the Semantic Network of Adposition and Case Supersenses (SNACS) to create a unified, lexicon-independent framework for all sentence participants.

The SNACS Framework and Construal

The core innovation of this approach is the application of Construal, which decouples the deep semantic role from the surface grammatical function. Every participant is assigned two labels:

  1. Scene Role: The underlying semantic role in the event (e.g., RECIPIENT).
  2. Function: The role suggested by surface morphosyntax or grammatical coding (e.g., AGENT for a transitive subject).
Examples of Scene Role vs. Function labels across different phrases.

To adapt the SNACS hierarchy for core grammatical relations, the researchers introduced role duplication (allowing multiple participants to share a role), a thematic hierarchy to resolve labeling conflicts (e.g., AGENT > INSTRUMENT > THEME), and specific logic for copular constructions using IDENTITY and CHARACTERISTIC labels.

Experimental Setup

To validate this expansion, a pilot interannotator agreement (IAA) study was conducted using a sample of 100 scenes from the English UCCA-annotated Wiki corpus. The dataset comprised 31 subjects, 26 objects, 42 prepositional phrases (PPs), and 1 possessive. Four expert annotators labeled the participants according to the refined SNACS guidelines.

Key Results

The study demonstrated that lexicon-independent labeling is viable for core arguments. Surprisingly, agreement on the "deep" Scene Role was higher for subjects and objects than for prepositional phrases, suggesting that the underlying meaning of core participants is often more intuitive than that of prepositions.

Interannotator agreement (Cohen’s Kappa) for Subjects/Objects and PPs.

Quantitative results using Cohen’s Kappa ($\kappa$) showed:

  • Subjects/Objects: $\kappa = 0.66$ for Scene Role and $0.61$ for Function.
  • Prepositional Phrases: $\kappa = 0.64$ for Scene Role and $0.77$ for Function.

The higher Function agreement for PPs is expected, as prepositions provide an explicit signal, whereas the "function" of a subject is more abstract.

Discussion and Conclusion

The results indicate that a coarse-grained, universal hierarchy can effectively categorize sentence participants without verb-specific frames. This "Scene vs. Function" split elegantly handles metaphors and grammatical shifts. For instance, in "Jane bought the book," Jane's Function is AGENT, but her Scene Role is RECIPIENT because she ends up with the book.

Future work will address "agentivity" gray areas (e.g., perception verbs) and complex "caused-motion" constructions. Additionally, scaling the study beyond the 100-scene pilot and testing on non-English languages will be crucial to proving the language-neutral claims of the SNACS framework.

Summary updated

Abstract

Research on adpositions and possessives in multiple languages has led to a small inventory of general-purpose meaning classes that disambiguate tokens. Importantly, that work has argued for a principled separation of the semantic role in a scene from the function coded by morphosyntax. Here, we ask whether this approach can be generalized beyond adpositions and possessives to cover all scene participants-including subjects and objectsdirectly, without reference to a frame lexicon. We present new guidelines for English and the results of an interannotator agreement study.
BibTeX
@misc{shalev2019preparing,
  author = {Adi Shalev and Jena D. Hwang and Nathan Schneider and Vivek Srikumar and Omri Abend and Ari Rappoport},
  title = {{Preparing SNACS for Subjects and Objects}},
  year = {2019},
  doi = {10.18653/v1/w19-3316},
}
Figure