Semantic Role Labeling (SRL) is traditionally described as the task of answering "who did what to whom" in a sentence. For decades, natural language processing frameworks operationalized this by focusing almost entirely on verbs. If a sentence contained the verb destroy, an SRL system would identify the destroyer (the Agent) and the entity being destroyed (the Patient).
However, human language frequently packages events, actions, and states into other parts of speech—namely nouns, adjectives, and prepositions. Consider the phrase: "The FBI's swift investigation of the company." The core event and its participants are identical to "The FBI swiftly investigated the company," yet the event trigger here is the noun investigation.
As early as 1991, Fuliang Weng in Case Revisited: In the Shadow of Automatic Processing of Machine-Readable Dictionaries warned that focusing exclusively on verbal predicates leaves nouns, adjectives, and prepositions unaddressed, creating an incomplete semantic representation. Overcoming this verb-centric bias through non-verbal SRL is essential for comprehensive text understanding.
The Cost of the Verb-Centric Bias
When machine learning models are trained to extract events or semantic roles solely from verbs, they develop a massive blind spot. This limitation manifests in two primary ways:
- Incomplete Event Coverage: Nouns account for a massive share of how information is structured. In CEO: Corpus-based Open-Domain Event Ontology Induction, the authors note that nouns make up approximately 30% to 40% of event mentions in standard text corpora. Relying strictly on verbal triggers means an event extraction system or ontology induction pipeline will miss nearly half of the actionable data, a vulnerability also highlighted by Bing Qin et al. in Building Chinese Event Type Paradigm Based on Trigger Clustering.
- Cross-Linguistic Inadequacy: The assumption that predicates are inherently verbal is an artifact of Western-centric language modeling. As Thai-Hoang Pham et al. demonstrate in Building a Semantic Role Labelling System for Vietnamese, many languages routinely employ nouns, adjectives, and prepositions to function as core predicates that govern subjects and objects.
Without expanding SRL to non-verbal predicates, downstream applications like incremental parsing (Ioannis Konstas et al., Incremental Semantic Role Labeling with Tree Adjoining Grammar) and coreference resolution (Altaf Rahman et al., Resolving Complex Cases of Definite Pronouns: The Winograd Schema Challenge) hit a performance ceiling because they cannot parse relationships anchored by non-verbal tokens.
Why Non-Verbal SRL is Hard
We cannot simply take a model trained on verbs and apply it to nouns or adjectives. Non-verbal predicates operate in entirely different syntactic environments, introducing unique structural challenges.
Syntactic Realization and Sparsity
Verbs have highly predictable structural arguments (e.g., subject-verb-object). Nouns, or nominalizations (nouns derived from verbs, like distribution from distribute), express arguments through highly variable noun-phrase syntax. Arguments might appear as genitive modifiers ("the company's distribution") or within prepositional phrases ("distribution of wealth"). To capture these, Sameer Pradhan et al. in Parsing Arguments of Nominalizations in English and Chinese showed that models must incorporate specialized features like "light" support verbs and noun-specific structural markers to locate arguments accurately.
Complex Predicates
Events are frequently split across a semantically empty verb and a meaningful noun—known as a Light Verb Construction (LVC), such as "take a walk" or "give a lecture". In PropBank Annotation of Multilingual Light Verb Constructions, Jena D. Hwang et al. demonstrated that treating the verb and noun as separate entities breaks semantic parsers. Instead, frameworks must treat the verb-noun pair as a single unified semantic predicate to correctly assign roles to the surrounding arguments.
Implicit Arguments
In verbal sentences, omitting a core argument often results in ungrammaticality ("The chef cooked." implies something was cooked, but "The chef devoured." feels incomplete). Nominal predicates, however, routinely drop arguments without violating syntax. In the sentence, "The investigation concluded," the investigator and the entity being investigated are completely implicit. Jennifer Sikos and Yannick Versley in Implicit Semantic Roles in a Multilingual Setting established that resolving these missing participants requires models to look beyond immediate syntactic dependencies, often leveraging cross-lingual alignment to recover hidden roles.
Frameworks for Broad-Coverage SRL
To resolve these challenges, researchers have shifted away from isolated verb-labeling toward unified, broad-coverage semantic frameworks.
Resource Unification
A major bottleneck for non-verbal SRL has been data scarcity. While early efforts like the Finnish PropBank (Katri Haverinen et al., The Finnish Proposition Bank) and the Revised Arabic PropBank (Wajdi Zaghouani et al., The Revised Arabic PropBank) painstakingly annotated verbal nouns and gerunds manually, modern approaches seek to unify predicates globally. For example, Roberto Navigli et al. introduced NounAtlas: Filling the Gap in Nominal Semantic Role Labeling, a large-scale resource designed to map nominal and verbal actions into a shared semantic space, allowing models to share knowledge between verbs and their noun counterparts.
Question-Answer Driven SRL (QA-SRL)
Traditional SRL relies on expert-defined linguistic ontologies (like PropBank or FrameNet), which become incredibly complex when applied to non-verbal parts of speech. To bypass this, Luheng He et al. proposed QA-SRL in Question-Answer Driven Semantic Role Labeling: Using Natural Language to Annotate Natural Language, replacing abstract role labels with natural language question-answer pairs (e.g., “What was distributed?” $\rightarrow$ “Wealth”).
Traditional SRL: [The FBI's]Arg0 [investigation]Predicate [of the company]Arg1
QA-SRL Style: Who investigated? -> The FBI | Who was investigated? -> The company
This paradigm was adapted specifically for nouns by Ayal Klein et al. in QANom: Question-Answer driven SRL for Nominalizations. By leveraging crowdsourced natural language questions, the authors proved that models can learn the intricate argument structures of nominalizations without requiring deep syntactic expertise.
Joint Syntactic-Semantic Modeling
Because non-verbal predicates are deeply intertwined with complex noun-phrase syntax, traditional pipeline systems (which parse syntax first, then label roles) suffer from severe error propagation. If the syntactic parser misidentifies a genitive modifier, the semantic role labeler fails downstream.
To fix this, Junhui Li et al. developed a framework in Joint Syntactic and Semantic Parsing of Chinese that optimizes both tasks simultaneously. By allowing nominal and verbal semantic role constraints to guide the syntactic tree-building process, the model achieves higher accuracy on both fronts, demonstrating that syntax and non-verbal semantics are best understood as a unified problem.
Key Foundational Work
Case Revisited: In the Shadow of Automatic Processing of Machine-Readable Dictionaries (1991) — contributed 1 concept elements to this direction.