Comprehensive Survey Maps Semantic Role Labeling in the Age of Large Language Models

Researchers have released a comprehensive survey of semantic role labeling (SRL)—a foundational natural language processing task that identifies which words in a sentence play which grammatical roles—examining how the field has evolved alongside pretrained language models.

The survey, submitted to arXiv, organizes nearly 200 papers published between 2000 and 2025 using a four-dimensional taxonomy covering:

• Model architectures used for SRL systems
• Syntactic feature modeling approaches
• Application scenarios across different domains
• Multimodal extensions including vision, video, and speech

Key Findings#

The authors provide critical analysis of when syntax actually helps. Rather than assuming syntactic features always improve SRL, they identify specific conditions where syntax-aided approaches consistently outperform syntax-free alternatives—and where they don't.

The survey also offers the first systematic treatment of how large language models interact with dedicated SRL systems. Rather than declaring one approach superior, the researchers explore complementary roles and potential hybrid strategies that combine the strengths of both.

Beyond traditional text-based SRL, the review extends into multimodal territory, analyzing how SRL performs across visual, video, and speech domains and examining structural differences in evaluation across these modalities.

Scope and Method#

The literature review drew from systematic searches across the ACL Anthology, IEEE Xplore, ACM Digital Library, and Google Scholar, applying explicit inclusion and exclusion criteria to ensure coverage and coherence.

The survey discusses SRL benchmarks, evaluation metrics, modeling paradigms, and practical applications before identifying future research directions as the field navigates the expanding role of large language models in NLP.