Semantic Parsing with Dual Learning

Chen Liu; Jieyu Li; Kai Yu; Ruisheng Cao; Su Zhu

arxiv: 1907.05343 · v2 · pith:WOXZBGCXnew · submitted 2019-07-10 · 💻 cs.CL · cs.AI

Semantic Parsing with Dual Learning

Ruisheng Cao , Su Zhu , Chen Liu , Jieyu Li , Kai Yu This is my paper

Pith reviewed 2026-05-25 00:09 UTC · model grok-4.3

classification 💻 cs.CL cs.AI

keywords semantic parsingdual learninglogical formsunlabeled datareward signalATISOvernight

0 comments

The pith

A dual-learning game between a semantic parser and its reverse model improves results by using unlabeled data and logical-form structure knowledge.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a dual-learning framework where a model mapping natural language to logical forms plays against a model mapping logical forms back to natural language. The game lets both models regularize each other and draw feedback signals from known properties of logical-form structure, allowing effective use of unlabeled queries alongside labeled ones. A new reward signal checks outputs at surface and semantic levels to favor complete and reasonable logical forms. If the approach holds, semantic parsing becomes less dependent on expensive labeled data. Experiments report new state-of-the-art accuracy on the ATIS dataset and competitive results on Overnight.

Core claim

The authors introduce a dual-learning algorithm that pairs a primal semantic parser with a dual logical-form-to-query model; the resulting mutual regularization and prior-knowledge feedback, combined with a surface-and-semantic-level reward, enable fuller use of labeled and unlabeled data and deliver new state-of-the-art performance on ATIS together with competitive performance on Overnight.

What carries the argument

The dual-learning game between the semantic parser and the logical-form-to-query model, which supplies mutual regularization and a novel multi-level reward signal derived from logical-form structure.

If this is right

Semantic parsers can reach higher accuracy with the same number of labeled examples by adding unlabeled natural-language queries.
The dual model supplies a training signal that favors logically complete and structurally valid outputs.
New state-of-the-art results appear on ATIS without requiring extra labeled data.
The same dual-game pattern can be applied to other structured-prediction tasks that suffer from scarce annotations.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Collecting large logs of unlabeled user queries could further improve performance in deployed systems.
The method may lower annotation costs when moving semantic parsing to new domains whose logical-form grammar is already known.
If the reward signal proves robust, it could transfer to other sequence-to-structure generation problems where partial outputs are common errors.
Testing the dual game on logical forms with greater nesting depth would reveal how far the structure-based feedback scales.

Load-bearing premise

The game between the two models reliably supplies useful regularization and feedback rather than simply amplifying each model's own mistakes.

What would settle it

Training both models on the same labeled data without the dual game or the new reward produces equal or higher accuracy than the full dual setup.

read the original abstract

Semantic parsing converts natural language queries into structured logical forms. The paucity of annotated training samples is a fundamental challenge in this field. In this work, we develop a semantic parsing framework with the dual learning algorithm, which enables a semantic parser to make full use of data (labeled and even unlabeled) through a dual-learning game. This game between a primal model (semantic parsing) and a dual model (logical form to query) forces them to regularize each other, and can achieve feedback signals from some prior-knowledge. By utilizing the prior-knowledge of logical form structures, we propose a novel reward signal at the surface and semantic levels which tends to generate complete and reasonable logical forms. Experimental results show that our approach achieves new state-of-the-art performance on ATIS dataset and gets competitive performance on Overnight dataset.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Dual learning with structure-based rewards is a straightforward way to pull signal from unlabeled data in semantic parsing, but the SOTA claim on ATIS rests on experiments the abstract does not show.

read the letter

The core idea is to run a dual game between a semantic parser and a logical-form-to-text model so they regularize each other, then add surface- and semantic-level rewards that favor complete, well-formed logical forms based on known structure priors. This is presented as a way to make better use of both labeled and unlabeled data without new annotations. That framing is new enough in this domain and the reward design follows directly from the logical-form properties the authors already know how to check. The reported outcome is new SOTA on ATIS and competitive numbers on Overnight, which would matter if the controls hold. The paper does a clean job of spelling out how the dual objective supplies feedback that standard supervised training lacks. The main soft spot is that the abstract supplies almost no experimental detail—no baseline list, no ablation on the reward terms, no error analysis—so it is impossible to tell whether the gains come from the dual game itself or from other training choices. The weakest assumption is that the structure priors produce reliable signals rather than just encouraging the model to stay inside the space it already knows how to generate. If the full paper contains those controls and they survive, the central claim stands; nothing in the description suggests an internal contradiction or circularity. This work is aimed at researchers who care about data efficiency in semantic parsing or other structured prediction tasks. A reader already following dual learning or low-resource NLP would find the setup useful to try even if the exact numbers need checking. It is worth sending to peer review because the problem is real, the method is a clear extension, and the experiments can be evaluated once the details are on the table.

Referee Report

2 major / 1 minor

Summary. The paper proposes a dual-learning framework for semantic parsing that pairs a primal natural-language-to-logical-form model with a dual logical-form-to-query model. The dual game is intended to regularize both models and supply feedback signals derived from prior knowledge of logical-form structures; a novel surface- and semantic-level reward is introduced to encourage complete and reasonable outputs. The central empirical claim is that the method attains new state-of-the-art results on ATIS and competitive performance on Overnight while making effective use of both labeled and unlabeled data.

Significance. If the reported gains are reproducible and attributable to the dual-learning mechanism rather than to unstated modeling choices, the work would offer a practical route to improving semantic parsers under limited supervision, a persistent bottleneck in the field. The explicit incorporation of logical-form structural priors into the reward is a potentially transferable idea.

major comments (2)

[Abstract] Abstract: the claim of 'new state-of-the-art performance on ATIS' is presented without any experimental details, baselines, ablation studies, error analysis, or quantitative tables, rendering the central performance claim unverifiable from the supplied text.
[Abstract] Abstract: the description of the dual-learning game and the 'novel reward signal at the surface and semantic levels' supplies no equations, pseudocode, or implementation specifics for how the regularization, feedback signals, or structure-aware reward are computed, which is load-bearing for the claim that the game produces effective regularization from logical-form priors.

minor comments (1)

[Abstract] The abstract would be clearer if it briefly indicated the size of the labeled/unlabeled splits used and the magnitude of improvement over the previous best system.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major comment point by point below.

read point-by-point responses

Referee: [Abstract] Abstract: the claim of 'new state-of-the-art performance on ATIS' is presented without any experimental details, baselines, ablation studies, error analysis, or quantitative tables, rendering the central performance claim unverifiable from the supplied text.

Authors: Abstracts are concise summaries by design and do not replicate full experimental details, baselines, ablation studies, error analyses, or tables; those elements appear in the Experiments section of the full manuscript. The central performance claim is supported by the complete evaluation reported in the body of the paper. We do not plan to alter the abstract, as doing so would exceed conventional length limits without improving verifiability of the full work. revision: no
Referee: [Abstract] Abstract: the description of the dual-learning game and the 'novel reward signal at the surface and semantic levels' supplies no equations, pseudocode, or implementation specifics for how the regularization, feedback signals, or structure-aware reward are computed, which is load-bearing for the claim that the game produces effective regularization from logical-form priors.

Authors: The abstract supplies a high-level overview of the approach. The equations defining the dual-learning objective, the surface- and semantic-level rewards, the regularization terms, the feedback signals, and the associated pseudocode are all provided in the Method section of the manuscript. This organization is standard and keeps the abstract accessible while placing the technical specifics where they belong. No revision to the abstract is required. revision: no

Circularity Check

0 steps flagged

No significant circularity; dual-learning regularization presented as independent signal

full rationale

The paper's central mechanism is a dual-learning game between a semantic parser and a logical-form-to-query model that supplies regularization and structure-aware rewards derived from logical-form priors. No equation or claim reduces a reported performance gain to a fitted parameter or self-citation by construction; the dual-game feedback is described as an external source of signal rather than a renaming or tautological re-use of the model's own outputs. Minor self-citation of prior dual-learning work exists but is not load-bearing for the ATIS SOTA claim. The derivation therefore remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that mutual regularization via the dual game supplies useful learning signals beyond standard supervised training; no free parameters or invented entities are mentioned in the abstract.

axioms (1)

domain assumption Dual learning between primal and dual models can regularize each other and extract feedback from prior knowledge of logical form structures
Invoked to justify use of unlabeled data and the novel reward signal.

pith-pipeline@v0.9.0 · 5660 in / 1059 out tokens · 41920 ms · 2026-05-25T00:09:20.985892+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/AbsoluteFloorClosure.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

dual-learning game between primal semantic parser and dual logical-form-to-query model... validity reward... grammar_error_indicator
IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

validity reward at surface and semantic levels

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.