Implicit Semantic Role Labeling

This direction is split into other directions: implicit argument resolution. Its last version stays readable and citable here.

Version 1 of 129 papers34 conceptsfrozen 2026-10-04

Standard Natural Language Processing often focuses on the "who did what to whom" within the narrow confines of a single sentence. However, human communication is rarely that explicit. We frequently omit essential participants in an event because they were mentioned earlier, are obvious from the context, or are part of shared world knowledge. These missing pieces are known as implicit arguments or null instantiations.

Implicit Semantic Role Labeling (iSRL) is the task of identifying these "open slots" and linking them to the correct entities in the broader discourse. If standard semantic role labeling is the detective work of a single sentence, iSRL is the work of connecting clues across an entire story to build a cohesive document-level understanding.

The Problem of Missing Participants

In a sentence like "The sale was finalized," a human reader intuitively knows there is a buyer and a seller, even though they aren't mentioned. Daniel Hershcovich and Ruixiang Cui describe these as the "dark matter" of linguistics—essential for meaning but invisible in the surface text.

Traditional systems struggle with this because they are designed to find arguments that are syntactically linked to a verb (e.g., "John sold the car"). When those links are broken or absent, the semantic representation becomes fragmented. This is particularly evident in:

  • Nominalizations: Nouns derived from verbs, like "failure" or "purchase," often omit the agent or object. Deborah A. Dahl and Martha Palmer noted early on that resolving these missing arguments is crucial for processing technical documents.
  • Zero Anaphora: In procedural texts like recipes, objects are frequently dropped (e.g., "Bake [it] for 30 minutes"). Yiwei Jiang and colleagues found that while models can often detect that something is missing, they struggle to accurately describe the omitted object.

Linking Roles as Anaphora Resolution

One of the most powerful intuitions in this field is treating implicit roles as "invisible pronouns." Carina Silberer reimagined iSRL as an anaphora resolution task. Just as the word "he" refers back to a previously mentioned person, an unfilled "buyer" role in a later sentence can be seen as an anaphoric link to a person mentioned paragraphs ago.

This perspective allows researchers to use discourse-level mechanisms to bridge gaps. Greg Whittemore and Greg Carlson proposed using "open" thematic roles as active anchors that wait for future information to fill them. By treating these roles as placeholders, a system can incrementally build an event description as a conversation progresses.

From Local Parsers to Reading Comprehension

As the field moved toward neural models, the framing of the problem shifted from structural linguistics to information retrieval. Pengxiang Cheng 0001 and Katrin Erk reframed implicit argument prediction as a reading comprehension task. In this setup, the missing role (e.g., "Who is the buyer?") acts as a query, and the model searches the entire document context to find the most likely answer.

This approach bypasses the need for complex syntactic rules and instead leverages the model's ability to reason over long-range dependencies. Similarly, Quynh Ngoc Thi Do and Marie-Francine Moens demonstrated that training Recurrent Neural Networks on millions of sentences allows models to learn the typical "sequences" of semantic frames, helping them predict missing participants based on statistical patterns of how events usually unfold.

Building Document-Level Graphs

The ultimate goal of iSRL is to move beyond isolated sentences toward a unified document-level semantic graph. Frameworks like Abstract Meaning Representation (AMR) and Universal Conceptual Cognitive Annotation (UCCA) provide the scaffolding for this.

Qiankun Fu and Yue Zhang developed models for AMR coreference resolution, which treat the task as "missing-link prediction" over a graph. By connecting nodes across different sentence-level graphs, they create a single, cohesive map of the entire text's meaning. More recently, Jon Cai and Martha Palmer introduced tools like LiDARR to scale this process, using AI "copilots" to help human annotators identify and link these latent roles more efficiently.

The Challenge of Data Sparsity

Despite these advances, iSRL remains difficult because implicit roles are, by definition, not there. This makes them hard to find and even harder to annotate. Paul Roit and Gabriel Stanovsky addressed this by developing "controlled crowdsourcing" for QA-SRL, which uses natural language questions to coax annotators into identifying roles that traditional frameworks might overlook.

Tatjana Moor and Michael Roth 0001 also found that "data sparsity" is a major hurdle; however, focusing on specific, high-frequency predicates allows models to learn the nuances of how certain verbs (like "give" or "leave") behave when their arguments go missing, leading to significant performance gains.

Papers in this version

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_implicit_semantic_role_labeling_62843181,
  title = {Implicit Semantic Role Labeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/implicit-semantic-role-labeling/rdr_62843181cfac453996a260d4876ce868},
  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_semantic_role_labeling_62843181_v1,
  title = {Implicit Semantic Role Labeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/implicit-semantic-role-labeling/rdr_62843181cfac453996a260d4876ce868/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}