Enhanced Dependencies

Captures predicates that are elided or implied in the surface syntax but essential for semantic interpretation. This includes:

• Gapping constructions: 'Kim likes apples and Pat pears' — recovers the missing 'likes' predicate for Pat
• Stripping: 'She went home, and he too' — reconstructs the elided verb phrase
• VP ellipsis: 'She will read it, and he will too' — links the empty VP to its antecedent

This augmentation transforms a syntactically sparse tree into a semantically complete graph where all logical predicates are explicitly represented.

Explicitly links matrix verbs to the embedded predicates they semantically govern through control or raising constructions:

• Subject control: 'She promised to go' — adds an arc from 'promised' to 'go' with the controller 'She'
• Object control: 'They persuaded her to stay' — links 'persuaded' to 'stay' via the controlled subject 'her'
• Raising: 'He seems to be happy' — connects 'seems' directly to 'happy' with 'He' as the raised argument

These arcs resolve the argument structure that basic dependencies leave implicit, enabling accurate semantic role assignment.

Adds arcs that make the shared argument in relative clauses explicit, connecting the head noun directly to its role inside the relative clause:

• 'The book that I read' — adds an arc from 'read' to 'book' with the obj relation
• 'The person who called' — links 'called' to 'person' with the nsubj relation
• 'The city where she was born' — connects 'born' to 'city' with the obl relation

This augmentation eliminates the gap-tracing problem that complicates information extraction from relative clause constructions.

Propagates shared dependents across conjoined structures to make the full scope of modification explicit:

• Shared subjects: 'Kim and Pat arrived' — propagates the nsubj relation to both 'Kim' and 'Pat'
• Shared modifiers: 'red cars and trucks' — propagates the amod relation from 'red' to both 'cars' and 'trucks'
• Shared prepositions: 'She is interested in and committed to the project' — propagates the obl relation to both conjuncts

This propagation resolves the scope ambiguity inherent in basic dependency trees, clarifying exactly which elements each modifier governs.

Introduces explicit null tokens into the dependency graph to represent elided material that has no surface realization but is required for complete semantic interpretation:

• Ellipsis resolution: 'She bought apples, and he oranges' — inserts a null node for the elided 'bought'
• Comparative deletion: 'She is taller than he' — adds a null node for the missing predicate in the comparative clause
• Sluicing: 'Someone called, but I don't know who' — inserts a null node for the elided clause 'called'

Null nodes transform the dependency graph into a fully specified logical form suitable for downstream reasoning systems.

Provides the critical bridge between surface syntax and deep semantics by making argument structure explicit:

• Passive-to-active mapping: 'The cake was eaten by the child' — adds a direct nsubj arc from 'eaten' to 'child' and obj to 'cake'
• Dative alternation: 'She gave him the book' and 'She gave the book to him' — normalizes both to the same underlying argument structure
• Causative decomposition: 'She made him laugh' — links the causer and causee to their semantic roles

This grounding enables accurate semantic role labeling and downstream tasks like question answering and relation extraction.