In natural language processing, Semantic Role Labeling (SRL) answers the fundamental question: who did what to whom, when, and where? For decades, systems excelled at extracting these relationships when they were anchored by a verb. Consider the sentence:
"The company invested millions in technology."
Here, the verb invested acts as a structural beacon. It clearly dictates its core arguments: the subject (The company is the investor) and the direct object (millions is the asset).
But what happens when the exact same event is expressed through a noun?
"The company's investment of millions in technology..."
Suddenly, the clear grammatical scaffolding of the sentence dissolves. The event-evoking word is now a noun—a nominal predicate. Despite conveying the same semantic information, identifying the participants of nominal predicates is notoriously difficult. As shown by Sebastian Riedel 0001 et al. in Collective Semantic Role Labelling with Markov Logic and Mihai Surdeanu et al. in The CoNLL 2008 Shared Task on Joint Parsing of Syntactic and Semantic Dependencies, there is a persistent performance gap between verbal and nominal SRL.
Understanding why this gap exists—and how researchers have attempted to close it—requires looking closely at how nouns structurally distort the relationship between syntax and meaning.
The Structural Slippiness of Nouns
The foundational challenge of nominal SRL stems from a lack of rigid syntactic constraints. In early NLP work, Richard H. Granger observed in FOUL-UP: A Program that Figures Out Meanings of Words from Context that verbs inherently provide the structural framework of a sentence, whereas nouns typically serve as fillers for those slots. When a noun becomes the framework, standard parsing assumptions break down.
Verbs rely on highly predictable subject-object patterns. Nominal predicates, by contrast, express their arguments through a chaotic mix of possessive markers, prepositional phrases, and external modifiers. As Sameer Pradhan et al. demonstrated in Parsing Arguments of Nominalizations in English and Chinese, a nominalization like investigation might pack its arguments inside a single noun phrase ("the FBI's investigation of the crime") or distribute them across a sentence using a "light" support verb ("the FBI carried out an investigation").
Zheng Ping Jiang quantified this challenge in Semantic Role Labeling of NomBank: A Maximum Entropy Approach, noting that nominal arguments exhibit extreme syntactic flexibility. They are frequently optional, highly variable in positioning, and governed more by fluid semantic relationships than by strict grammatical rules.
Why Tree Pruning Fails for Nouns
To understand why automated systems struggle with this flexibility, consider how traditional semantic parsers locate arguments in a dependency tree.
For a verbal predicate, an algorithm can safely use localized pruning heuristics. Because a verb's arguments are almost always its immediate children or siblings in the syntax tree, the parser can discard distant nodes and focus its search locally.
[Tree Structure for a Verb]
[invested] (Predicate)
/ \
[company] [millions]
(Subject) (Object)
For nominal predicates, this spatial locality vanishes. Hongling Wang et al. detailed this breakdown in Dependency Tree-based SRL with Proper Pruning and Extensive Feature Engineering. Because noun arguments can appear as ancestors, distant cousins, or deep descendants in a dependency tree, standard verb-based pruning heuristics fail to capture them.
[Tree Structure for a Noun Predicate]
[carried] (Support Verb)
/ \
[company] [investigation] (Predicate)
(Ancestor) \
[millions] (Descendant)
Compounding this issue, Chang Liu et al. pointed out in Learning Predictive Structures for Semantic Role Labeling of NomBank that nominal predicates sometimes structurally overlap with their own arguments within a single word, completely breaking the distinct spatial boundaries that parsers rely on to separate actions from participants.
The Invisible Argument Problem
Verbs generally demand that their core arguments be explicitly present in the clause to be grammatically correct ("He gave..." feels incomplete). Nouns suffer from no such restrictions; they routinely leave their participants entirely implicit.
In Towards a Resource for Lexical Semantics: A Large German Corpus with Extensive Semantic Annotation, Katrin Erk et al. highlighted that nominal expressions frequently omit participants or bury them inside complex, non-standard prepositional structures.
When these arguments are expressed, they often cross sentence boundaries entirely. Matthew Gerber et al. tackled this in Beyond NomBank: A Study of Implicit Arguments for Nominal Predicates, revealing that standard sentence-bound SRL systems miss a massive portion of a text's meaning. For nominal predicates, up to 65% of total arguments are "implicit"—meaning they reside in previous sentences or broader discourse context, requiring coreference resolution and discourse parsing to uncover.
The Predicate Identification Bottleneck
Before a system can label arguments, it faces a stark binary hurdle: deciding whether a noun is acting as a predicate at all.
Verbs are almost always predicates. Nouns, however, are overwhelmingly static entities. Junhui Li et al. noted in Improving Nominal SRL in Chinese Language with Verbal SRL Information and Automatic Predicate Recognition that only a small fraction of nouns in a given text function as event triggers. This sparsity creates a severe data bottleneck.
Furthermore, a system must distinguish between a noun's eventive use and its static or idiomatic use. As Daniel Hershcovich et al. explained in Comparison by Conversion: Reverse-Engineering UCCA from Syntax and Lexical Semantics, shallow syntactic frameworks often treat "scene-evoking" nouns identically to simple entities. For instance, a parser must determine if the word construction refers to an active process involving an agent and a patient ("the construction of the bridge") or a static object ("the building is a concrete construction"). As Matthew Gerber et al. observed, when nominal predicates become trapped inside fixed collocations or compound idioms, they lose their eventive properties entirely, making precise disambiguation critical.
Representation in Modern Models
Given these fundamental differences, can deep learning models simply learn to bridge the gap? Evidence suggests it is not that simple.
Using probing techniques, Simone Conia et al. discovered in Probing for Predicate Argument Structures in Pretrained Language Models that large language models construct fundamentally different internal representations for nominal versus verbal predicate-argument structures. Because these structural "blueprints" diverge within the network's layers, the model cannot easily transfer its robust knowledge of verbal syntax over to nominal syntax.
This divergence has forced a shift toward alternative annotation paradigms. Rather than forcing nouns into rigid, verb-centric dependency frameworks, researchers have turned to question-answering templates to capture nominal roles intuitively. Frameworks like QANom, introduced by Ayal Klein et al. in QANom: Question-Answer driven SRL for Nominalizations, bypass complex syntax trees by using natural language questions (e.g., "What was invested?", "Who invested?") to map out the semantic space. By decoupling the underlying semantic roles from surface-level syntax, as further explored by Valentina Pyatkin et al. in Asking It All: Generating Contextualized Questions for any Semantic Role, modern systems can finally capture the elusive participants of nominal events, regardless of how deeply they are buried in the text.
Key Foundational Work
FOUL-UP: A Program that Figures Out Meanings of Words from Context (1977) — contributed 2 concept elements to this direction.
Case Structure Systems (1973) — contributed 1 concept elements to this direction.