Semantic parsing bridges the gap between human language and machine-executable code, translating natural language queries into structured representations like SQL, lambda calculus, or formal logical forms. However, training these parsers traditionally requires massive parallel corpora of sentences paired with their exact logical equivalents. This requirement creates a severe annotation bottleneck: while raw text is abundant, expert-crafted logical forms are exceptionally expensive, time-consuming, and difficult to scale across new domains or database schemas.
To bypass this dependency, researchers have developed paradigms that learn semantic mappings without direct sentence-to-logic supervision. These approaches shift the learning signal from explicit human annotations to implicit structural, environmental, or generative cues.
Learning from the World's Response
If we do not provide a model with the correct logical form, how can it learn? The key intuition is to leverage the execution environment. Instead of telling a model how to write the query, we only tell it whether the query's result is correct.
As demonstrated by James Clarke et al. in Driving Semantic Parsing from the World's Response, a semantic parser can treat the database or environment as a black box that returns feedback—such as a denotation (e.g., "Paris") or a binary success signal. In a comprehensive overview of the field, Percy Liang explains in Learning Executable Semantic Parsers for Natural Language Understanding that this feedback allows the system to guide its search through the space of possible logical forms. If a generated program yields the correct answer, the model reinforces the actions that led to it.
However, weak supervision introduces a major challenge: spurious programs. A parser might generate an incorrect logical form that accidentally evaluates to the correct answer (for instance, calculating the average of a column instead of the maximum, where both happen to equal 5 for a specific data subset). To combat this, Omer Goldman introduced the concept of abstract templates in Weakly Supervised Semantic Parsing with Abstract Examples. By grouping sentences and programs into generalized structural patterns, the model learns to generalize across semantic types rather than memorizing brittle, accidental paths to the correct answer.
Grounding Language in Existing Schemas
Another powerful source of supervision is the target environment's inherent structure—its ontology or database schema. A database is not a random collection of strings; it defines a rigid, meaningful network of entities and relations.
In Unsupervised Semantic Parsing, Hoifung Poon et al. showed that models can automatically discover semantic mappings by clustering syntactic variations of text that correspond to the same underlying database relations. Hoifung Poon extended this in Grounded Unsupervised Semantic Parsing by directly "grounding" language into database schemas, allowing the system to learn valid logical forms purely by matching the statistical distribution of natural language predicates with the relational structure of the database.
When natural language does not map cleanly to a database's rigid schema—a problem known as ontological mismatch—decoupling linguistic structure from the target ontology becomes necessary. Eunsol Choi et al. addressed this in Scaling Semantic Parsers with On-the-Fly Ontology Matching by using an intermediate layer that separates the sentence's syntax from the database schema, learning an alignment model on the fly. Similarly, Min-Chul Yang et al. utilized joint relational embeddings in Joint Relational Embeddings for Knowledge-based Question Answering to project text phrases and knowledge graph predicates into a shared latent space, bypassing the need for hand-crafted lexical rules.
Generative Reconstruction and Latent Variables
If we view semantic parsing as a translation task from language to logic, we can also frame it as a generative autoencoding problem. The intuition here is simple: if a model parses a sentence into a logical form, it should be able to reconstruct the original sentence from that logical form.
This idea is formalized by Pengcheng Yin et al. in StructVAE: Tree-structured Latent Variable Models for Semi-supervised Semantic Parsing. By treating the target logical form as a tree-structured latent variable within a Variational Autoencoder (VAE), the model can leverage vast amounts of unannotated text. The parser acts as an encoder, and a generator acts as a decoder. Because the latent variable is forced to adhere to valid logical syntax, the model learns to extract structurally sound meaning representations without explicit parallel labels.
Zero-Shot Parsing and Agentic Search
With the advent of Large Language Models (LLMs), the strategy for mitigating the annotation bottleneck has shifted from training specialized parsers to designing zero-shot prompting and search frameworks. LLMs possess broad linguistic capabilities but struggle with complex, multi-step logical reasoning over specific, large-scale knowledge bases.
To resolve this, Dheeraj Mekala et al. introduced ZEROTOP: Zero-Shot Task-Oriented Semantic Parsing using Large Language Models, which reframes hierarchical semantic parsing into a sequence of simpler, zero-shot question-answering tasks. Rather than generating a massive, complex query in one shot, the system decomposes the problem into manageable, localized decisions.
Taking this a step further, Haoran Luo et al. developed KBQA-o1: Agentic Knowledge Base Question Answering with Monte Carlo Tree Search. Instead of treating semantic parsing as a direct translation problem, their framework uses Monte Carlo Tree Search (MCTS) to navigate the knowledge base iteratively, evaluating candidate paths dynamically. This agentic approach allows smaller models to match or exceed the accuracy of heavily supervised systems, proving that structured search over an environment can effectively substitute for massive annotated datasets.
Key Foundational Work
Unsupervised Semantic Role Labellin (2004) — contributed 1 concept elements to this direction.