acceptodds
Under review as a conference paper at ICLR 2027

Glyph Normalization with Discrete Generative Transport for Low Resource Tangut Cursive Script

Abstract

Tangut cursive script contains extensive stroke connections, omissions, and structural deformations, leading to substantial visual differences from standard printed glyphs and making direct recognition particularly difficult under limited training data. To reduce the influence of complex cursive variations on subsequent recognition and digital processing, we formulate Tangut cursive processing as a generative glyph normalization task, where cursive glyphs are progressively transformed into standard printed representations to provide more stable visual inputs for downstream recognition. Unlike conventional generative models that mainly emphasize visual quality or style transfer, our task places greater emphasis on preserving character structure and identity during transformation. We therefore propose a Source Conditioned Generative Transport Model (SC-GTM) based on discrete latent representations. Cursive and printed glyphs are encoded into a shared visual token space, in which the model progressively updates local probability states through multi-to-multi spatial transport and local token rewriting. The transport process reorganizes observable local information, while the rewriting process accounts for morphological changes and structures that may be partially omitted in cursive writing, without requiring predefined one-to-one correspondences between strokes, components, or tokens. To constrain the underdetermined intermediate transformation, we further introduce path relative-entropy regularization inspired by the Schr\"odinger bridge formulation together with paired endpoint supervision. Experiments on both unseen characters and unseen cursive instances show that SC-GTM achieves competitive visual reconstruction quality and favorable structural consistency, particularly in foreground overlap and skeleton-based measures. Although recognition accuracy is not directly used as an evaluation metric in this work, the results suggest that explicitly modeling spatial reorganization and local rewriting can effectively reduce structural discrepancies between cursive and printed glyphs, providing a promising direction for low-resource Tangut cursive normalization and subsequent automatic recognition.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Tangut Cursive, Low Resource Recognition, Glyph Normalization, Discrete Representation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_1MmvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper presents SC-GTM, a model for normalizing low-resource Tangut cursive glyphs into canonical printed forms. It operates in a shared discrete visual-token space, alternating capacity-constrained many-to-many spatial transport with bounded local token rewriting. Paired endpoint supervision guides restoration, while path relative entropy regularizes the intermediate dynamics against a fixed reference process. Evaluation uses a curated dataset of 1,048 character classes under unseen-character and seen-character/unseen-instance protocols, comparing against nine baselines using structural and perceptual metrics. Ablations examine transport and path regularization. The reported results favor SC-GTM on skeleton Chamfer distance, while foreground IoU and appearance metrics show mixed rankings across methods.

Critical strengths
  1. The transport–rewrite formulation explicitly separates spatial reorganization from token-category changes without correspondence annotations, providing a well-motivated inductive bias for restoring structurally distorted cursive glyphs.
  2. The unseen-character evaluation tests generalization beyond memorized classes. Improved skeleton Chamfer distance and supporting ablations provide initial evidence that transport and path regularization benefit structural restoration.
Critical weaknesses

1.The evaluation relies on image and structural similarity metrics, without expert assessment or direct verification of character identity. Small stroke or component changes may alter identity, so the reported metric improvements do not fully establish the intended normalization quality. 2.The ablations only remove transport or path regularization, leaving the bounded rewriting mechanism and transport capacity constraints largely untested. Additional controlled ablations would clarify whether these constraints are necessary for the observed improvements.

Questions for authors

Reference Critical Weaknesses.

Reviewer_ndJZOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SC-GTM for Tangut cursive glyph normalization, combining discrete spatial transport, local token rewriting, and path regularization. Experiments report improvements on several structural similarity metrics.

Critical strengths

The paper addresses a relatively underexplored task and reports some promising improvements on structural metrics.

Critical weaknesses
  1. The paper is substantially underdeveloped in its methodology and presentation. The main content is unusually brief, with no overall framework diagram and insufficient architectural and implementation details. The related work is superficial, and the novelty over existing generative transport and glyph transformation methods is not clearly established. These omissions also raise serious concerns about reproducibility.

  2. The experimental validation is inadequate. The evaluation is limited to a small dataset, and the reported improvements are modest or inconsistent across metrics. The ablation study does not sufficiently validate the individual contributions of the proposed components. More importantly, the experiments lack direct evaluation of character identity preservation, leaving the practical effectiveness of glyph normalization unsubstantiated.

Questions for authors
  1. What is the essential methodological novelty compared with existing generative transport and glyph transformation approaches?

  2. Can the authors provide complete implementation details and more comprehensive experimental evidence, particularly regarding component contributions and character identity preservation?

Reviewer_apMsOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper addresses paired normalization of Tangut cursive glyphs into canonical printed forms. SC-GTM operates in a shared discrete visual-token space, alternating capacity-constrained spatial transport with bounded token rewriting and a path-relative-entropy regularizer. Experiments consider unseen character classes and held-out instances of seen classes. The reported advantages are mainly in structural similarity measures such as skeleton distance and foreground IoU, rather than uniform improvements across perceptual metrics.

Critical strengths
  1. A worthwhile low-resource application. Recovering a canonical glyph, rather than only predicting a class label, could help inspect how a system resolves structural variation in historical script. The distinction between unseen-character and unseen-instance evaluation is useful.

  2. An interpretable decomposition. Separating spatial transport from token rewriting gives a concrete modeling hypothesis. The paper also appropriately avoids interpreting its internal transport as literal stroke genealogy, and reports metrics on which competing methods perform better.

Critical weaknesses
  1. The experimental evidence is not sufficiently reliable to establish the claimed advance. The custom dataset is described through broad references to facsimiles and public online sources, without enough source-level provenance, character-identity verification, or split manifests to audit the paired task and its generalization conditions. The Palette and CycleGAN examples on pp.9–10 also show conspicuous color/background or grid-like artifacts and extremely weak metrics, raising unresolved questions about training and preprocessing comparability. These observations do not establish data leakage or misconduct, but they do make effective baseline checks essential. Finally, the paper describes standard deviations over independent runs, whereas the supplied evaluation script computes variation across test samples for one model; a separate run-aggregation procedure may exist, but it is not established by the supplied materials.

  2. The methodological increment is limited and its central interpretation is not fully validated. The method combines vector quantization, capacity-constrained transport, local rewriting, and path regularization, but the evidence does not show that this combination yields a sufficiently distinct learning contribution beyond a tailored engineering design. The transport/path ablations are useful, yet simpler capacity-matched rewriting or iterative restoration controls would be more discriminating. Moreover, the stated strictly 3×3-local rewriting controller includes GroupNorm in the supplied implementation, introducing dependence on spatially aggregated statistics. This matters because restricted locality is part of the argument for the decomposition. Overall, neither the methodological advance nor its empirical validation currently meets the standard needed for acceptance at ICLR.

Questions for authors

Please refer to the weaknesses.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.