Shallow Semantic Parsing

The core output of shallow semantic parsing is a flat structure identifying who did what to whom, when, where, and how. It maps a predicate (typically a verb) to its associated arguments (Agent, Patient, Instrument) and adjuncts (Location, Time). This provides a 'who-did-what' abstraction layer over raw syntax, enabling machines to grasp the basic event semantics without deep logical form.

Classic systems decompose the task into two sequential sub-tasks:

• Argument Identification: Detecting which constituents in a parse tree are potential arguments of a given predicate.
• Argument Classification: Assigning a specific semantic role label (e.g., ARG0 for Agent, ARG1 for Patient) to each identified argument. Modern neural approaches often solve both stages jointly using span-based architectures.

The 'shallow' distinction is critical. Unlike deep semantic parsing (e.g., Abstract Meaning Representation), shallow parsing does not resolve:

• Quantifier scope (e.g., 'Every boy loves a dog').
• Complex compositional semantics or logical entailment.
• Coreference across sentences. It focuses strictly on local, sentence-level predicate-argument relations, making it computationally cheaper and more robust for noisy text.

Shallow semantic parsing relies heavily on annotated lexical databases:

• PropBank: Annotates verbs with verb-specific numbered roles (ARG0–ARG5) and modifiers (ARGM-*).
• FrameNet: Defines semantic frames (e.g., Commerce_buy) with frame elements (Buyer, Seller, Goods).
• VerbNet: Maps syntactic frames to thematic roles and selectional preferences. These resources provide the label inventory for supervised training.

State-of-the-art systems use Transformer-based encoders (e.g., BERT) fine-tuned for SRL. Key architectural patterns include:

• Span-Based SRL: Enumerates and scores arbitrary text spans as potential arguments, removing reliance on external syntactic parsers.
• BIO Tagging: Treats SRL as a token-level sequence labeling problem (Beginning, Inside, Outside of an argument).
• Biaffine Attention: Computes pairwise scores between predicate and argument representations for high-precision role assignment.

The standard benchmark is the CoNLL-2012 Shared Task using the OntoNotes 5.0 corpus. Evaluation metrics include:

• F1 Score for argument identification and classification.
• Labeled F1: Requires exact match of argument span and role label.
• Unlabeled F1: Requires only correct argument span identification. OntoNotes provides multi-genre data (news, conversations, web text) with gold syntactic and semantic annotations.

The linguistic framework representing a sentence as a predicate (typically a verb) and its associated arguments (such as agent, patient, and instrument). SSP maps flat structures like 'who did what to whom' without nesting or scope resolution.

• Predicate: The event or relation (e.g., 'broke')
• Arguments: Participants in the event (e.g., 'the programmer', 'the build')
• Flat representation: No hierarchical embedding of propositions

A set of generalized semantic categories describing the relationship a noun phrase has with the main verb. SSP assigns these roles without deep compositional semantics.

• Agent: Deliberate initiator of action
• Patient: Entity undergoing the effect
• Theme: Entity moving or located
• Instrument: Tool used to perform action
• Goal: Destination or recipient

The first sub-task of SSP: detecting which constituents in a parse tree are potential arguments of a given predicate. This is a binary classification step.

• Operates on syntactic constituents or spans
• Prunes non-argument material (adjuncts, interjections)
• Modern systems use span enumeration with neural scoring
• Critical for reducing the search space before role classification

The second sub-task: assigning a specific semantic role label (e.g., Agent, Patient, Instrument) to a previously identified argument. This is a multi-class classification problem.

• Labels come from PropBank framesets or FrameNet frame elements
• Requires predicate sense disambiguation first
• BERT-based models achieve >90% F1 on standard benchmarks
• Contextualized embeddings capture role-specific cues

A corpus annotated with predicate-argument relations and verb-specific semantic roles, serving as the foundational training resource for supervised SSP systems.

• Defines numbered roles: Arg0 (proto-Agent), Arg1 (proto-Patient), Arg2-Arg5 (verb-specific)
• Includes adjunct labels: ArgM-TMP (temporal), ArgM-LOC (locative)
• Covers Wall Street Journal and other genres
• Framesets disambiguate verb senses

A sequence labeling encoding that transforms SSP into a token classification problem, removing the need for pre-computed syntactic parse trees.

• B-ARG: Beginning of an argument span
• I-ARG: Inside an argument span
• O: Outside any argument
• Enables end-to-end neural models without pipeline dependencies
• Used in span-based and token-based SRL architectures