Dependency-Based SRL

Unlike span-based methods that enumerate arbitrary text spans, dependency-based SRL identifies arguments by locating the syntactic head of each semantic role. The algorithm traverses the dependency tree from the predicate outward, selecting dependents that satisfy specific path constraints.

• Direct dependency: Arguments are often immediate dependents of the predicate
• Path-based heuristics: Non-local arguments are identified via dependency path patterns
• Syntactic scaffolding: The parse tree provides a natural pruning mechanism, reducing the search space compared to span enumeration

Dependency-based SRL excels with non-projective constructions—sentences where syntactic dependencies cross, common in languages with free word order such as Czech, Hindi, and German. Because dependency trees represent grammatical relations directly rather than through constituency brackets, crossing arcs are natively supported.

• Handles discontinuous arguments without special mechanisms
• Avoids the spurious ambiguity of phrase-structure trees
• Provides a unified framework for both projective and non-projective languages

The core modeling signal comes from the dependency path between a predicate and its candidate argument. These paths encode syntactic relationships as sequences of dependency labels and direction markers, providing a compact, interpretable feature representation.

• Path length: Shorter paths typically indicate core arguments; longer paths suggest adjuncts
• Label sequences: Specific dependency labels (e.g., nsubj, dobj) strongly correlate with thematic roles
• Tree kernels: Graph kernels over dependency paths enable similarity-based learning without explicit feature engineering

Dependency-based SRL naturally supports multi-task learning where syntactic parsing and semantic role labeling are trained jointly. Shared representations between syntax and semantics improve generalization, particularly in low-resource scenarios.

• Shared encoders: A single BiLSTM or Transformer encodes the sentence for both tasks
• Biaffine attention: The same scoring mechanism used for dependency arc prediction can be adapted for predicate-argument scoring
• Error propagation reduction: Joint models are less vulnerable to cascading errors from a separate parser

Because dependency grammar abstracts away from language-specific word order into universal grammatical relations, dependency-based SRL transfers more readily across languages. The Universal Dependencies framework provides a standardized annotation scheme that enables zero-shot and few-shot cross-lingual role labeling.

• UD v2: Standardized dependency relations across 100+ languages
• Delexicalized transfer: Models trained on English dependency-SRL can project labels onto target language parses
• Typological generalization: Dependency structures capture shared grammatical functions even when surface order differs radically

Dependency-based SRL extends naturally into semantic dependency parsing, where the output is a directed graph of semantic relations between words rather than a flat set of predicate-argument pairs. This unified representation captures not only 'who did what to whom' but also negation scope, modality, and discourse relations.

• Beyond trees: Semantic graphs allow words to have multiple semantic heads
• Enhanced Dependencies: Augmented representations that add semantic relations to basic syntactic trees
• End-to-end graph parsing: Neural models that predict semantic edges directly, integrating SRL with broader semantic analysis

The syntactic backbone of this SRL approach. A dependency parser produces a directed graph where nodes are words and edges represent binary grammatical relations like nsubj (nominal subject) or dobj (direct object). Dependency-based SRL exploits the fact that semantic arguments often correspond to specific syntactic subtrees governed by the predicate. Key algorithms include transition-based (MaltParser) and graph-based (Eisner's algorithm) methods. Modern neural parsers using biaffine attention achieve over 95% UAS on standard benchmarks.

A scoring mechanism central to neural dependency-based SRL. Instead of classifying roles independently, a biaffine layer computes pairwise scores between a predicate representation and every candidate argument token. The mechanism uses a low-rank bilinear transformation: s(i,j) = r_i^T U r_j + W(r_i ⊕ r_j) + b. This captures asymmetric relationships—critical because the role of 'eat' relative to 'apple' differs from 'apple' relative to 'eat'. Introduced by Dozat and Manning (2017), it remains the dominant architecture for both syntactic and semantic dependency parsing.

An alternative paradigm that enumerates arbitrary text spans as candidate arguments rather than relying on syntactic heads. While dependency-based SRL assigns roles to syntactic heads, span-based methods score every possible contiguous subsequence. This eliminates error propagation from imperfect parses but increases computational complexity. Modern implementations use span pruning heuristics and are often evaluated against dependency-based baselines. The debate centers on whether explicit syntax remains necessary when large pre-trained models implicitly encode grammatical structure.

A modeling paradigm that explicitly injects syntactic parse trees as structural priors. Techniques include: - Graph convolutional networks operating over dependency edges - Tree-structured LSTMs that propagate information along syntactic paths - Syntax-guided self-attention that biases transformer heads toward grammatical relations Dependency-based SRL is inherently syntax-aware, as it operates directly on the parse. The key insight: semantic roles rarely cross syntactic boundaries, making the dependency tree a powerful inductive bias that improves sample efficiency and generalization to rare argument structures.

The landmark evaluation campaign that established dependency-based SRL as a core NLP benchmark. Unlike CoNLL-2012 (which used phrase-structure parses), the 2009 task required systems to predict syntactic dependencies and semantic roles jointly across seven languages. Winning systems demonstrated that joint modeling of syntax and semantics outperforms pipeline approaches. The labeled semantic F1 metric from this task remains the standard evaluation for dependency-based SRL, with modern BERT-based systems achieving scores above 90% on in-domain English text.