Implicit Argument Resolution in Natural Language Discourse

32 concepts23 papers1 proposal

Standard Semantic Role Labeling (SRL) excels at answering the fundamental question of "who did what to whom" within a single sentence. By analyzing local syntactic structures, traditional SRL models map verbs to their explicit arguments, such as the agent performing an action or the patient receiving it.

However, human language is inherently telegraphic and efficient. In real-world discourse, we routinely omit essential participants when they are already clear from context. Consider a simple narrative: “The company launched a new software product. The first sale was finalized yesterday.” A standard sentence-level SRL system analyzing the second sentence will identify "The first sale" as the event, but it will fail to identify who sold the product or who bought it. The local syntax provides no clues.

As noted by Moor et al. and Thi et al., this sentence-boundary limitation leaves significant information gaps, resulting in a fragmented and shallow machine understanding of narrative flow. To bridge this gap, researchers developed Implicit Argument Resolution—the task of recovering missing event participants that are syntactically absent from the local clause but inferable from the broader discourse or background knowledge.

The Anatomy of Missing Information

Before a system can resolve an implicit argument, it must understand how and why information disappears from the surface text. Omissions generally fall into three linguistic categories:

1. Null Instantiations

Sometimes, a verbal predicate simply leaves an essential role unexpressed. Thompson et al. modeled this behavior statistically by allowing semantic roles to produce "null emissions," representing participants understood by context but absent from the surface text.

Li et al. emphasize that systems must distinguish between two types of null instantiations:

  • Indefinite omissions: The missing argument refers to a generic or unknown entity (e.g., "He ate [something] at noon"), meaning the system does not need to look for a specific referent.
  • Definite omissions: The missing argument refers to a specific entity mentioned previously in the text, requiring the system to actively search the preceding context to bind the role.

2. Nominalizations

When a verb is transformed into a noun (e.g., "fail" becomes "failure", "distribute" becomes "distribution"), it frequently sheds its structural requirements for explicit arguments. Dahl et al. demonstrated that a unified semantic system can reuse the same underlying definitions for both verbs and their related nouns to capture these missing pieces. More recently, the QANom framework introduced by Klein et al. highlighted that nominalizations routinely omit the agents of an action, making them a primary driver of semantic underspecification.

3. Verb Phrase Ellipsis

In sentences containing ellipsis (e.g., "Cats chase dogs, and children do too"), entire segments of meaning are dropped to avoid redundancy. Wijnholds et al. showed that standard linear distributional models fail to capture the semantics of these sentences. Instead, models must employ non-linear composition mechanisms to explicitly "copy" and integrate the elided components back into the sentence representation.

Exploiting Discourse Coherence

How can a machine logically deduce what isn't explicitly written? The foundational intuition relies on the structural continuity of text.

Early rule-based systems like PUNDIT, developed by Palmer et al., separated semantic roles into essential and non-essential categories, treating unfulfilled essential roles as active triggers that force the system to query discourse context or domain knowledge bases. To maintain cohesion as a narrative progresses, Whittemore et al. proposed treating these open thematic roles as "active anchors" that remain open to capture and bind new entities introduced in subsequent sentences.

A powerful heuristic for resolving these anchors is argument coherence: the tendency of a writer to keep the same participants across a sequence of events. Laparra et al. exploited this principle via a deterministic algorithm (ImpAr) that utilizes a stateful default filler mechanism. By assuming participant continuity across a discourse, the algorithm successfully recovers nearly half of all missing semantic roles without requiring massive, predicate-specific training data.

[Sentence 1]  The company launched a new software product.
                                 │
                         (Argument Coherence)
                                 ▼
[Sentence 2]  The first sale [by the company] was finalized yesterday.

When simple continuity heuristics fall short, the problem can be framed as a specialized coreference resolution task. Laparra et al. adapted classical linguistic constraints like Centering Theory to resolve implicit links. Moving beyond linear text, Fu et al. transitioned this idea to document-level Abstract Meaning Representation (AMR) graphs. By treating coreference as a "missing-link prediction" problem over semantic graphs, their model uses Graph Recurrent Networks to dynamically wire missing arguments to their cross-sentence referents.

Interestingly, the signal for an implicit role can also be found by looking across languages. Sikos et al. demonstrated that when aligning parallel corpora (such as English and German), information that is explicit in one language often becomes implicit or completely dropped in translation, providing a natural supervisory signal for detecting implicit roles.

From Parsing to Semantic Reasoning

As natural language processing shifted toward deep learning, the methodology for implicit argument resolution underwent a paradigm shift. Instead of treating the problem as an extension of syntactic parsing, modern architectures re-frame it as a semantic query and reasoning task.

Question-Answer Frameworks

Traditional syntax-driven annotation schemes often struggle to capture implicit arguments because human annotators cannot easily anchor them to specific syntactic constituents. Roit et al. addressed this by using a natural language Question-Answer (QA-SRL) framework. By asking simple questions like "Who was something sold to?", the framework allows crowd-workers to intuitively identify implicit arguments that are semantically present but syntactically absent.

Reading Comprehension

Cheng et al. formalized this query-based intuition by re-framing implicit argument prediction as a reading comprehension task.

Context:  "The company launched a new software product. The first sale was finalized."
Query:    "Who finalized the sale?"
Target:   [The company]

By treating the missing argument as a cloze-style query and the entire document as context, their Pointer Attentive Reader model can look across sentence boundaries, reasoning through complex narratives to select the exact span corresponding to the missing participant.

Textual Entailment

Taking the query paradigm to its logical conclusion, Roit et al. cast document-level argument detection as a Natural Language Inference (NLI) task. Rather than training a complex, fully supervised sequence-labeling model on scarce document-level data, the authors generate simple synthetic sentences to "ask" the document if a specific entity performs a specific role.

For instance, to check if "the company" is the seller, the system evaluates whether the document entails the statement "The company is the seller." This formulation leverages the extensive, pre-existing reasoning capabilities of large NLI models, bypassing the need for expensive document-level annotations and outperforming fully supervised alternatives.

Key Foundational Work

On the Types of Semantic Compression of Text (1982) — contributed 2 concept elements to this direction.
Recovering Implicit Information (1986) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_implicit_argument_resolution_b98df6a1,
  title = {implicit argument resolution},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/implicit-argument-resolution/rdr_b98df6a148104aa58993dee224adaf5e},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_implicit_argument_resolution_b98df6a1_v1,
  title = {implicit argument resolution},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/implicit-argument-resolution/rdr_b98df6a148104aa58993dee224adaf5e/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake