Semantic Role Labeling (SRL)

The first critical sub-task: detecting which constituents in a sentence are potential arguments of a given predicate. This step prunes the search space by filtering out irrelevant phrases before classification.

• Span detection models enumerate all possible text spans as candidate arguments
• BIO tagging treats identification as a token-level sequence labeling problem
• Pruning heuristics use syntactic distance to discard unlikely constituents
• Performance here directly bounds the recall ceiling of the entire SRL pipeline

Once arguments are identified, this sub-task assigns specific semantic role labels (e.g., ARG0 for Agent, ARG1 for Patient) to each argument. Modern systems use contextualized embeddings from models like BERT to capture the nuanced relationship between a predicate and its arguments.

• Requires predicate sense disambiguation to select the correct frameset
• Labels follow PropBank conventions: numbered core arguments (ARG0-ARG5) and modifier adjuncts (ARGM-TMP, ARGM-LOC)
• Biaffine attention mechanisms score predicate-argument pairs efficiently

The process of determining the exact sense of a predicate in context. A verb like 'run' evokes different semantic frames depending on usage: 'run a marathon' vs. 'run a company' vs. 'run water.'

• Links predicates to specific PropBank framesets or FrameNet lexical units
• Critical because different senses license different argument structures
• Often performed jointly with argument classification in neural models
• Errors cascade: a wrong sense leads to incorrect role assignments downstream

Two dominant architectural paradigms for structuring the SRL task:

• Dependency-based SRL operates on syntactic dependency trees, identifying semantic roles based on head-dependent relations. It leverages existing parser outputs but inherits parser errors.
• Span-based SRL enumerates arbitrary text spans as candidate arguments, removing reliance on pre-computed parse trees. This end-to-end neural approach achieves state-of-the-art results by learning latent syntactic structure implicitly.

Modern systems increasingly favor span-based methods with BERT-based encoders.

SRL distinguishes between arguments essential to a predicate's meaning and optional circumstantial modifiers:

• Core arguments (ARG0-ARG5) are required by the verb's semantic frame. For 'give,' ARG0 (giver), ARG1 (thing given), and ARG2 (recipient) are core.
• Non-core adjuncts (ARGM-*) add circumstantial information: temporal (ARGM-TMP), locative (ARGM-LOC), manner (ARGM-MNR), purpose (ARGM-PRP), and more.
• Null arguments are semantically understood but syntactically unrealized participants, common in pro-drop languages.

SRL benefits significantly from joint training with related NLP tasks through shared representations:

• Co-training with syntactic parsing improves argument boundary detection
• Joint modeling with coreference resolution enables cross-sentence role consistency
• Shared encoders with named entity recognition enhance selectional preference learning
• The CoNLL-2012 Shared Task on OntoNotes established the standard for evaluating end-to-end systems that combine SRL with coreference