Semantic Parsing

The core output is a logical form—a structured, unambiguous representation of meaning. This is typically expressed in languages like Lambda Calculus, SQL, Prolog, or custom domain-specific languages (DSLs). The parser must resolve scope ambiguities and map syntactic structures to semantic predicates.

• Example: "Who is the CEO of the company that bought Twitter?" becomes a nested database query.
• Key Challenge: Handling compositional semantics where the meaning of a phrase depends on the meaning of its parts.

Historically, parsing relied on hand-crafted Combinatory Categorial Grammar (CCG) rules paired with lexicons. Modern systems use sequence-to-sequence (seq2seq) neural models that directly translate text to code.

• Grammar-Based: High precision, brittle to linguistic variation.
• Neural: Robust to paraphrasing, trained on text-code pairs.
• Intermediate Decoding: Constrained decoding ensures syntactically valid logical forms.

Parsing must be grounded against a specific knowledge base or environment. The same phrase "list users" maps to different SQL schemas or API calls depending on the context.

• Schema Linking: Aligning words like "customer" to a table named clients.
• Executable Verification: The gold standard is whether the parsed logic executes correctly and returns the right answer, not just surface-level string match.

A critical benchmark is the ability to understand novel combinations of known primitives. A model trained on "find flights to Boston" and "show me the cheapest option" must generalize to "find the cheapest flight to Boston."

• Systematicity: Recombining known linguistic atoms.
• Benchmarks: Datasets like CFQ and COGS specifically test this capability, revealing brittleness in standard seq2seq models.

In multi-turn dialogues, parsing is contextual. The command "What about Paris?" requires resolving the anaphora based on the previous query (e.g., "flights to London").

• Dialogue State Tracking: The parser updates a belief state rather than generating a standalone logical form.
• Clarification Requests: The parser must identify ambiguity and ask for clarification (e.g., "Paris, France or Paris, Texas?") before executing.

Standard string-matching metrics like BLEU are insufficient. The primary metric is Exact Set Match Accuracy—whether the predicted logical form is semantically equivalent to the gold standard.

• Denotation Accuracy: Executing the predicted and gold logical forms against a database and comparing the returned results.
• Abstract Syntax Tree (AST) Match: Comparing the structure of the code, not just the tokens.

The process of identifying and classifying key elements like people, organizations, dates, and locations from unstructured text. These extracted entities become the arguments bound to predicates in the final logical form.

• Anchors a parse to real-world concepts
• Uses sequence labeling models like BERT-CRF or spaCy
• Critical for populating slots in task-oriented parsing

The process of extracting specific attributes from a query to populate a predefined semantic frame or template. Slot filling is the downstream consumer of entity extraction and the direct precursor to API execution.

• Example: 'Book a flight to London on Tuesday' fills destination: London and date: Tuesday
• Uses BIO tagging or sequence-to-sequence models
• Essential for task-oriented dialogue systems

The process of breaking a complex, multi-faceted query into a set of simpler atomic sub-queries that can be independently parsed and executed. The results are then synthesized into a composite answer.

• Handles comparative and aggregate questions
• Example: 'Which company had higher revenue, Apple or Google?' decomposes into two revenue lookups
• Often implemented with chain-of-thought prompting or recursive parsing

The process of connecting a textual mention to its unique, unambiguous entry in a knowledge base such as Wikidata or a proprietary graph. Entity linking resolves the referential ambiguity that would otherwise produce incorrect logical forms.

• Disambiguates 'Paris' (city vs. person vs. mythological figure)
• Uses dense retrieval against entity embeddings
• Provides the canonical IDs used in SPARQL or GraphQL queries

The computational task of identifying which meaning of a polysemous word is intended in context. WSD is critical for selecting the correct predicate or relation in a knowledge base schema.

• Determines if 'bank' means financial institution or river edge
• Uses contextual embeddings from models like BERT
• Directly impacts the accuracy of generated logical forms