Argument Pruning

Applies linguistic rules to eliminate impossible arguments before classification. Common heuristics include:

• Syntactic path filtering: Discard constituents whose dependency path to the predicate exceeds a threshold distance
• Constituent type restrictions: Remove adverbial clauses when searching for core arguments like Agent or Patient
• Voice constraints: In passive constructions, prune the subject position as a potential Agent

These rules are fast, interpretable, and require no training data, making them ideal for bootstrapping SRL systems on low-resource languages.

Trains a binary classifier to predict whether a constituent is a likely argument before role classification. The pruner scores candidate spans using features such as:

• Predicate-argument embedding similarity from contextual models like BERT
• Path embeddings encoding the syntactic route between predicate and candidate
• Linear distance and constituent type as auxiliary signals

Learned pruning adapts to domain-specific argument patterns and can achieve higher recall than rigid heuristics, though at the cost of additional inference latency.

Limits the span proposal mechanism in span-based SRL architectures. Without pruning, a sentence of length n generates O(n²) possible spans. Pruning strategies include:

• Width constraints: Reject spans exceeding a maximum token length (e.g., >10 words)
• Score thresholding: Keep only spans whose unary mention score exceeds a learned cutoff
• Predicate-specific filtering: For each predicate, retain only spans within a syntactic locality window

These constraints are essential for making span-based SRL computationally tractable on long sentences.

Uses parse tree structure as a hard constraint on argument boundaries. Constituents that do not align with syntactic nodes are pruned outright. Key techniques:

• Constituent pruning: Only consider nodes from a phrase-structure parse as valid argument spans
• Head-driven filtering: Retain only constituents whose syntactic head falls within a dependency window of the predicate
• Clause boundary detection: Prune arguments that cross clause boundaries unless a discourse relation permits it

This approach guarantees that all surviving candidates are syntactically coherent, improving role classification accuracy.

Leverages semantic type constraints to filter arguments based on ontological compatibility. A predicate like eat expects an edible Theme and an animate Agent. Pruning mechanisms include:

• Type hierarchy lookup: Map candidate head words to WordNet or ontology types and reject mismatches
• Distributional preference models: Use corpus-derived selectional association scores to rank and prune unlikely fillers
• Predicate-specific role filters: For PropBank rolesets, maintain a whitelist of acceptable entity types per role

This semantic layer catches errors that purely syntactic pruning misses, such as rejecting the idea as the Theme of devour.

Integrates pruning directly into the neural SRL architecture rather than treating it as a separate preprocessing step. Modern approaches include:

• Attention-based gating: Use predicate-aware attention scores to softly mask irrelevant tokens before span scoring
• End-to-end span ranking: Train a single model to jointly score and prune spans, optimizing for both recall and precision
• Coarse-to-fine inference: First pass identifies candidate regions with lightweight scoring; second pass applies expensive role classification only to survivors

Joint optimization eliminates the error propagation that occurs when a hard pruner mistakenly discards a true argument.