Argument Identification

The process of identifying syntactic constituents—such as noun phrases, prepositional phrases, and clauses—that serve as candidate arguments for a given predicate. This step relies on parse tree traversal to locate nodes that are siblings of the predicate in the syntactic structure. Key aspects include:

• Phrase boundaries: Determining where an argument begins and ends in the token sequence
• Syntactic filtering: Excluding constituents that cannot logically function as arguments (e.g., punctuation, interjections)
• Multi-word expressions: Handling idiomatic phrases where the predicate and argument boundaries blur

In the sentence "John gave Mary a book in the library," the system must identify four distinct constituents: John, Mary, a book, and in the library as potential arguments of the predicate gave.

Argument identification leverages syntactic proximity heuristics to narrow the search space. Constituents that are syntactically closer to the predicate in the parse tree are more likely to be true arguments than distant modifiers. Common proximity patterns include:

• Sister nodes: Constituents that share the same parent node as the predicate in a phrase-structure tree
• Immediate dominance: Arguments that are directly dominated by the predicate's governing verb phrase
• C-command relationships: Syntactic configurations where a node's sibling or any descendant of that sibling is a candidate

This proximity principle helps distinguish core arguments from adjuncts early in the pipeline, reducing computational overhead for the subsequent argument classification stage.

A central challenge in argument identification is distinguishing core arguments from peripheral adjuncts. Core arguments are semantically required by the predicate to form a complete proposition, while adjuncts provide optional circumstantial information. Key discriminators include:

• Obligatoriness: Removing a core argument renders the sentence ungrammatical or incomplete
• Verb-specificity: Core arguments are tied to the predicate's lexical entry in resources like PropBank
• Iterability: Adjuncts can typically be stacked (multiple time expressions), while core arguments cannot

For example, in "She ate the sandwich," the sandwich is a core argument; in "She ate the sandwich at noon," at noon is an adjunct. Early identification of this distinction prevents false positives in the SRL pipeline.

Traditional argument identification is tightly coupled to syntactic parsing quality. The accuracy of constituent detection directly depends on the correctness of the underlying parse tree. Critical dependencies include:

• Phrase-structure parses: Constituent boundaries from Penn Treebank-style parses define candidate argument spans
• Dependency parses: Head-dependent relations identify which tokens are governed by the predicate
• Parse errors propagate: A single misattached prepositional phrase can cause both false positives and false negatives

Modern span-based neural SRL architectures have reduced this dependency by enumerating all possible text spans directly, bypassing the need for pre-computed parse trees and mitigating cascading errors from syntactic parsers.

Argument pruning is a critical efficiency technique that reduces the candidate search space before classification. Since a sentence with n words has O(n²) possible spans, exhaustive enumeration is computationally prohibitive. Effective pruning methods include:

• Syntactic heuristics: Only considering constituents that are siblings of the predicate in the parse tree
• Span length constraints: Filtering out excessively long spans that are unlikely to be single arguments
• Learned pruning: Training a lightweight classifier to score and filter candidate spans based on contextual embeddings
• Predicate-argument distance thresholds: Limiting candidates to spans within a certain token distance from the predicate

High-quality pruning achieves a recall of over 95% while eliminating more than 80% of negative candidates, dramatically accelerating the downstream argument classification stage.

Argument identification must account for null arguments—semantically required participants that are syntactically unrealized. These implicit arguments are common in:

• Pro-drop languages: Languages like Spanish, Italian, and Japanese that routinely omit subject pronouns
• Imperative constructions: Commands where the subject you is understood but not expressed
• Passive voice: Constructions where the agent is demoted and optionally omitted
• Control structures: Infinitival clauses where the subject is co-referential with a matrix clause argument

Advanced SRL systems detect null arguments by examining the predicate's PropBank frameset, which specifies the full set of expected roles, and flagging roles with no corresponding syntactic realization for special handling.