Abstract
Semantic role labeling (SRL) is the task of identifying predicates and labeling argument spans with semantic roles. Even though most semantic-role formalisms are built upon constituent syntax, and only syntactic constituents can be labeled as arguments (e.g., FrameNet and PropBank), all the recent work on syntax-aware SRL relies on dependency representations of syntax. In contrast, we show how graph convolutional networks (GCNs) can be used to encode constituent structures and inform an SRL system. Nodes in our SpanGCN correspond to constituents. The computation is done in 3 stages. First, initial node representations are produced by ‘composing’ word representations of the first and last words in the constituent. Second, graph convolutions relying on the constituent tree are performed, yielding syntactically-informed constituent representations. Finally, the constituent representations are ‘decomposed’ back into word representations, which are used as input to the SRL classifier. We evaluate SpanGCN against alternatives, including a model using GCNs over dependency trees, and show its effectiveness on standard English SRL benchmarks CoNLL-2005, CoNLL-2012, and FrameNet.
Introduction
Semantic Role Labeling (SRL) aims to identify "who did what to whom" in a sentence by mapping predicates to their respective arguments. Historically, SRL has been deeply intertwined with constituent syntax—the hierarchical grouping of words into phrases—because argument boundaries typically align with these syntactic spans.
Despite this linguistic foundation, modern neural SRL models have largely shifted toward either syntax-agnostic approaches (using LSTMs or Transformers) or dependency-based syntax. Dependency trees, while useful, do not explicitly represent spans, creating a mismatch for a task fundamentally defined by span labeling. This paper introduces SpanGCN, an architecture designed to bridge this gap by directly encoding constituent trees using Graph Convolutional Networks (GCNs).
Methodology: The SpanGCN Architecture
SpanGCN integrates hierarchical constituent information into a sequence-based neural network through a three-stage process:
- Constituent Composition: Starting with word-level representations from a BiLSTM, the model creates "nodes" for every constituent in a syntactic tree. Each node is formed by composing the representations of the first and last words of its corresponding span.
- Constituent GCN: A GCN performs message passing between these nodes. Information flows along the tree edges (parent-to-child and vice versa), allowing higher-level phrases (like Verb Phrases) to incorporate context from their sub-constituents (like Noun Phrases).
- Constituent Decomposition: The syntactically-enriched constituent representations are decomposed back into word-level signals. These enriched word vectors are then passed to a standard SRL classifier consisting of a bilinear scorer and a Conditional Random Field (CRF).
To handle potentially noisy inputs from automatic parsers, the model employs scalar gates to weight the importance of different edges during the convolution stage.
Experimental Results
The model was evaluated on three major benchmarks: CoNLL-2005 (WSJ and Brown corpora), CoNLL-2012 (OntoNotes 5.0), and FrameNet 1.5.
Key Performance Insights
- Superiority over Dependency Models: SpanGCN consistently outperformed GCNs built over dependency trees. On CoNLL-2005, it achieved an F1 of 84.37, compared to 83.56 for the dependency-based GCN and 83.18 for the syntax-agnostic baseline.
- Handling Long-Range Dependencies: A significant challenge in SRL is identifying arguments far from the predicate. SpanGCN's tree structure provides "shortcuts" for information flow, leading to better performance on longer sentences compared to sequence-only baselines.
- Robustness and Upper Bounds: The model showed particular strength on out-of-domain data (the Brown test set), suggesting that syntactic structure is a more reliable feature than surface patterns when vocabulary shifts. Furthermore, using "gold" (human-annotated) syntax pushed the F1 score to 90.58, indicating that SRL performance will continue to scale with improvements in syntactic parsing.
Discussion and Conclusion
Error analysis reveals that SpanGCN's primary advantage lies in its ability to accurately identify argument boundaries. By modeling constituents directly, the system gains an inductive bias that aligns with the underlying linguistic formalism of SRL.
While the model's performance is currently linked to the quality of external parsers, it demonstrates that explicit constituent syntax remains highly relevant even in the era of large-scale contextual embeddings like RoBERTa. Future work may explore applying SpanGCN to other span-based tasks, such as coreference resolution or entity-relation extraction.
Summary updated