Semantic Dependency Parsing

Unlike syntactic trees, SDP produces a directed graph where words can have multiple semantic heads. This allows a node to participate in several relations simultaneously, capturing phenomena like control structures and shared arguments that are invisible to surface syntax. The graph is typically acyclic, ensuring logical semantic flow from predicates to their dependents.

SDP unifies predicate-argument relations directly into the dependency graph. Each predicate (verb, nominal, or adjectival) anchors a subgraph where edges are labeled with deep roles:

• ARG0: Proto-Agent (doer, experiencer)
• ARG1: Proto-Patient (undergoer, theme)
• ARG2: Instrument, benefactive, or secondary object This eliminates the separation between syntactic parsing and semantic role labeling.

Beyond predicate arguments, SDP explicitly labels modifier and logical relations that alter truth conditions:

• Negation (NEG): Reverses the truth value of a predicate
• Modality (MOD): Encodes possibility, necessity, or obligation
• Quantification (QUANT): Links quantifiers to their scopal domain
• Tense (TENSE): Anchors events temporally These are first-class citizens in the graph, not secondary annotations.

SDP abstracts away from language-specific syntactic idiosyncrasies. The same semantic relation—such as Agent or Location—is represented identically across languages with radically different word orders or case systems. This makes SDP ideal for multilingual semantic search and cross-lingual knowledge extraction, where syntactic structures diverge but meaning remains constant.

Modern SDP parsers use deep biaffine attention to score every possible directed edge between word pairs. For a sentence of length n, the model computes an n×n matrix of relation scores, then applies maximum spanning tree algorithms or graph-based inference to construct the final semantic graph. This architecture, pioneered by Dozat & Manning (2017), achieves state-of-the-art accuracy on benchmarks like the SDP shared tasks.

SDP occupies a middle ground between surface syntax and full semantic abstraction:

• Abstract Meaning Representation (AMR): Collapses inflection and function words into concept nodes; SDP retains lexical units
• Universal Conceptual Cognitive Annotation (UCCA): Focuses on scene-level structures; SDP preserves word-level granularity SDP is often used as a precursor or scaffold for building these deeper representations.

The syntactic foundation upon which semantic dependencies are built. This task analyzes the grammatical structure of a sentence by establishing directed binary relations between a head word and its dependent words.

• Produces a tree structure where each word has exactly one syntactic head
• Labels relations such as nsubj (nominal subject) or dobj (direct object)
• Serves as the structural scaffold for dependency-based SRL
• Contrasts with constituency parsing, which builds phrase-structure trees

The task of identifying the predicate-argument structure of a sentence—answering 'who did what to whom, when, where, and how.' SRL assigns labels like Agent, Patient, and Instrument to constituents.

• Semantic Dependency Parsing subsumes SRL by representing roles as labeled edges in a unified graph
• Traditional SRL produces flat structures; semantic dependencies capture nested and non-local relations
• Key resources include PropBank and FrameNet for role inventories

A rooted, directed graph that encodes the meaning of a sentence independent of its syntactic form. AMR abstracts away from function words and word order to represent conceptual relations.

• Nodes represent concepts; edges represent semantic relations like :ARG0 or :location
• Semantic Dependency Parsing and AMR parsing share the goal of producing graph-structured meaning
• AMR focuses on canonicalization of semantic roles across syntactic alternations
• Used in machine translation and summarization as a language-neutral interlingua

The process of resolving the exact sense of a predicate in context. A verb like 'run' has distinct meanings in 'run a marathon' versus 'run a company,' each corresponding to a different frameset or lexical unit.

• Links predicates to specific entries in PropBank framesets or VerbNet classes
• Critical for assigning correct semantic role labels to arguments
• Neural models use contextualized embeddings from BERT to disambiguate senses without explicit word sense inventories

A deep learning scoring mechanism that computes pairwise scores between a head and a dependent using a low-rank bilinear transformation. It is the dominant architecture for both syntactic and semantic dependency parsing.

• Applied to score the likelihood of a directed edge between every pair of words
• Enables global inference over the full graph structure
• Used in state-of-the-art parsers like Dozat and Manning (2017) for dependency parsing
• Extends naturally to multi-label classification for semantic edges

The task of identifying all expressions in a text that refer to the same real-world entity. For example, linking 'John' and 'he' and 'the CEO' as coreferent mentions.

• Essential for building coherent semantic graphs across sentence boundaries
• Enables argument sharing across clauses in multi-sentence semantic dependency structures
• The OntoNotes corpus provides gold-standard annotations for both coreference and SRL
• Modern systems use span-based neural architectures with antecedent scoring