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: composition, graph convolution, and decomposition. We evaluate SpanGCN against alternatives, including a model using GCNs over dependency trees, and show its effectiveness on standard CoNLL-2005, CoNLL-2012, and FrameNet benchmarks.
Introduction
Semantic Role Labeling (SRL) aims to answer the question "who did what to whom" by identifying predicates and their associated semantic arguments. For instance, in a sentence like "Investors appeal to the CEO not to limit their access to sales data," an SRL system must identify "appeal" as the predicate, "Investors" as the agent (ARG0), and "the CEO" as the entity being appealed to (ARG1).
Historically, SRL formalisms like PropBank and FrameNet were designed around constituency syntax, where arguments correspond to syntactic phrases (constituents). However, modern neural systems have largely shifted toward dependency-based syntax or syntax-agnostic deep learning models. This paper introduces SpanGCN, an architecture that re-integrates constituency syntax into neural SRL using Graph Convolutional Networks (GCNs).
The SpanGCN Architecture
The core innovation of SpanGCN is treating syntactic constituents as nodes in a graph. The model processes information in three distinct stages to create syntactically-informed word representations.
1. Constituent Composition
The process begins with a sentence processed by a BiLSTM. To represent a constituent (e.g., a Noun Phrase), the model "composes" a node representation by looking at the BiLSTM hidden states of the first and last words of that span. These composed representations serve as the initial nodes in the constituency graph.
2. Constituent GCN
The constituency tree is treated as a graph where nodes represent constituents. GCN layers facilitate message passing between nodes and their parents or children. This allows a high-level node (like a Verb Phrase) to incorporate information from its children (Noun Phrases) and its parent (the Sentence node), effectively encoding the hierarchical structure of the sentence into the constituent representations.
3. Constituent Decomposition
To return to the word level for final classification, the information is "decomposed." Constituent nodes send their updated representations back to the start and end tokens of their spans. A final BiLSTM layer then propagates this syntactic context across all words within each constituent, ensuring the entire span is informed by the global syntactic structure.
Experimental Results
The authors evaluated SpanGCN on three major benchmarks: CoNLL-2005 (WSJ and Brown), CoNLL-2012 (OntoNotes), and FrameNet 1.5.
Performance Gains
SpanGCN achieved state-of-the-art results for syntax-aware models. On the CoNLL-2005 WSJ test set, it reached an F1 score of 85.4, surpassing previous models like LISA (84.6). On CoNLL-2012, it achieved 84.4 F1, a new high for single-model performance.
Constituency vs. Dependency
A key finding is that SpanGCN outperforms models using GCNs over dependency trees (DepGCN). Because SRL is fundamentally a span-labeling task, the hierarchical nature of constituency trees provides a more natural inductive bias than word-to-word dependencies. Error analysis revealed that SpanGCN is significantly better at identifying precise argument boundaries, as the constituency tree provides a "scaffolding" that constrains where arguments can start and end.
The Role of Gold Syntax
When the model was tested using "gold" (perfect) syntactic trees instead of predicted ones, performance jumped to 90.58 F1. This suggests that the SpanGCN framework is highly effective at utilizing syntactic information, and its performance will continue to scale as standalone constituency parsers improve.
Conclusion
SpanGCN demonstrates that constituency syntax remains a powerful tool for Semantic Role Labeling. By explicitly modeling constituents as graph nodes, the model achieves superior boundary precision and overall accuracy compared to dependency-based or syntax-agnostic alternatives. Beyond SRL, the SpanGCN architecture holds promise for other span-based NLP tasks such as coreference resolution and entity-relation extraction.
Summary updated