Intermediary Semantic Representation through Proposition Structures

Gabriel Stanovsky, Jessica Ficler, Ido Dagan, Yoav Goldberg2014 · DOI

Abstract: We propose an intermediary-level semantic representation, providing a higher level of abstraction than syntactic parse trees, while not committing to decisions in cases such as quantification, grounding or verbspecific roles assignments. The proposal is centered around the proposition structure of the text, and includes also implicit propositions which can be inferred from the syntax but are not transparent in parse trees, such as copular relations introduced by appositive constructions. Other benefits over dependency-trees are explicit marking of logical relations between propositions, explicit marking of multiword predicate such as light-verbs, and a consistent representation for syntacticallydifferent but semantically-similar structures. The representation is meant to serve as a useful input layer for semanticoriented applications, as well as to provide a better starting point for further levels of semantic analysis such as semantic-rolelabeling and semantic-parsing.

Introduction

Modern semantic parsing often faces a significant hurdle: the attempt to map raw text directly to complex logical formalisms. This "one-step" approach forces parsers to simultaneously handle surface syntactic variations (like passive vs. active voice) and deep semantic challenges (like quantification or world-knowledge grounding), making it difficult to isolate and correct errors.

While dependency trees offer a syntactic alternative, they remain too close to the surface text. They often fail to capture implicit information—such as the fact that "The CEO, John Smith" implies "John Smith is the CEO." This paper proposes a "missing link" in the NLP pipeline: a Proposition Structure representation that abstracts away syntactic noise while remaining grounded in the sentence's core structure.

The Proposition Structure Framework

The proposed representation breaks every sentence down into a set of inter-linked propositions. Each proposition consists of a predicate and its associated arguments, defined by several key features:

  • Canonicalization: Predicates are reduced to base lemmas. Tense, negation, and modality are stored as features rather than separate nodes. This allows for the unification of passive and active voices; for instance, "door" is labeled as the Object in both "John opened the door" and "The door was opened by John."
  • Syntactic Role Inventory: To ensure robustness, the system uses a syntactic inventory (Subject, Object, Preposition_X) rather than complex thematic roles like "Agent" or "Patient," which are often difficult for automated systems to assign accurately.
  • Linking Propositions: The framework explicitly models logical relationships between propositions, such as conditionals or temporal sequences.
Figure 1: A proposition structure representing the conditional sentence 'If you leave the park, you will find the Peak Tram terminal,' illustrating linked predicates and argument assignments.

Making the Implicit Explicit

A core strength of this representation is its ability to extract "hidden" propositions from complex grammatical constructions:

  1. Appositions: A phrase like "Pierre Vinken, 61 years old" is transformed into the explicit proposition: [Pierre Vinken] is [61 years old].
  2. Nominalizations: "Google's acquisition of Waze" is represented as the action: [Google] acquired [Waze].
  3. Adjectives: "The high prevalence of illness" is captured as a state: [prevalence of illness] is [high].
  4. Multi-word Expressions: Light-verb constructions (e.g., "take a decision") are treated as single predicates, preventing core meanings from being fragmented by syntactic dependencies.

Conclusion and Future Work

Qualitative analysis using snippets from the Wall Street Journal (WSJ) demonstrates that this schema successfully unifies disparate ways of expressing the same meaning into a machine-readable format of feature-value pairs. By separating syntactic abstraction from deep semantic interpretation, the authors provide a more attainable starting point for applications like Question Answering and Information Extraction.

Future work aims to automate the production of these structures from raw text and expand the framework to handle nested infinitive structures and discourse-level propositions that span multiple sentences. Currently, the representation does not attempt to resolve word-sense ambiguity, leaving that for subsequent stages of semantic analysis.

Summary updated

Abstract

We propose an intermediary-level semantic representation, providing a higher level of abstraction than syntactic parse trees, while not committing to decisions in cases such as quantification, grounding or verbspecific roles assignments. The proposal is centered around the proposition structure of the text, and includes also implicit propositions which can be inferred from the syntax but are not transparent in parse trees, such as copular relations introduced by appositive constructions. Other benefits over dependency-trees are explicit marking of logical relations between propositions, explicit marking of multiword predicate such as light-verbs, and a consistent representation for syntacticallydifferent but semantically-similar structures. The representation is meant to serve as a useful input layer for semanticoriented applications, as well as to provide a better starting point for further levels of semantic analysis such as semantic-rolelabeling and semantic-parsing.
BibTeX
@misc{stanovsky2014intermediary,
  author = {Gabriel Stanovsky and Jessica Ficler and Ido Dagan and Yoav Goldberg},
  title = {{Intermediary Semantic Representation through Proposition Structures}},
  year = {2014},
  doi = {10.3115/v1/w14-2413},
}
Figure