acceptodds
Under review as a conference paper at ICLR 2027

A Unified Manifold Framework for Modeling Chinese Script Evolution with Application to Oracle Bone Decipherment

Abstract

Of the approximately 4,500 Oracle Bone Inscription (OBI) characters discovered from the Shang dynasty, only about 1,600 have been deciphered. Existing computational methods mainly compute the similarity between OBI and single historical period characters independently. However, during the evolution of Chinese characters, significant structural or semantic changes often occur in uncertain dynasties. Relying solely on a single dynasty for reference leads to numerous deciphering errors. Therefore, we propose the Manifold-based Script Evolution Framework (MSEF), the first unified framework that models the evolution series (OBI, Bronze, Seal, Clerical, Regular) of Chinese characters as the continual evolution of a manifold space. MSEF represents each character as an era-specific manifold point and learns continuous inter-era transition rules via Neural Ordinary Differential Equations. Both manifold space and transition dynamics can be trained end-to-end through character evolution pairs across any two eras. To support this training scheme, we construct the first unified cross-era dataset with fine-grained temporal and regional granularity. Based on MSEF, we propose the Cascaded Bidirectional Evolutionary Decipherment (CBED) algorithm, where forward predictions recall potential candidates and backward consistency checking filters false matches, effectively avoiding errors caused by single-dynasty comparisons. To ground our approach, we provide a pre-exploratory diagnosis of existing generative and retrieval paradigms, showing why continuous manifold dynamics are well suited for fragmented cross-era script data. We further conduct mechanistic post-analysis, showing that the learned representations and transition dynamics capture patterns consistent with known paleographic transformations. Experiments on three commonly used benchmarks and our dataset demonstrate that MSEF achieves state-of-the-art performance among current oracle deciphering algorithms. This work opens the possibility of using continuous manifold dynamics in cross-era paleographic decipherment. Dataset and code are released through our project page https://iclr-oracle-submission.github.io.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Oracle Character, Ancient Language, Language Decipherment

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_jkKTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Manifold-based Script Evolution Framework (MSEF), a unified framework that models the evolution series (OBI, Bronze, Seal, Clerical, Regular) of Chinese characters as the continual evolution of a manifold space. MSEF represents each character as an era-specific manifold point and learns continuous inter-era transition rules via Neural Ordinary Differential Equations. To support this training scheme, a unified cross-era dataset with fine-grained temporal and regional granularity is constructed. Experiments on three benchmarks and the proposed dataset demonstrate that MSEF achieves state-of-the-art performance (72.5% Recall@1 on FGCCES and 71.5% Top-1 accuracy on OBS-OCR) among the evaluated oracle deciphering algorithms.

Critical strengths
  1. The Manifold-based Script Evolution Framework (MSEF), which could model the evolution series (OBI, Bronze, Seal, Clerical, Regular) of Chinese characters as the continual evolution of a manifold space, is proposed. Such modeling is quite new compared to the existing works.
  2. The authors build a unified cross-era oracle dataset, FGCCES, for training MSEF, which provides fine-grained temporal (e.g., Early/Mid/Late Western Zhou) and regional granularity, explicitly addressing the data fragmentation issue in cross-era script matching
  3. Experiments on three commonly used benchmarks and the proposed dataset demonstrate that MSEF achieves state-of-the-art performance among the evaluated oracle deciphering algorithms.
Critical weaknesses
  1. Character evolution is fundamentally discrete and prone to non-diffeomorphic transformations (e.g., radical splitting, radical merging, and the abrupt stroke transformation). Because Neural ODEs define smooth, reversible, bijection-preserving trajectories, how does the framework theoretically handle cases where a single archaic glyph splits into multiple modern characters, or multiple ancient glyphs merge? The authors should formalize why a continuous vector field is an appropriate inductive bias for inherently discrete and punctuated cultural evolution, or clearly reframe the ODE as a continuous interpolation proxy rather than a physical law of script evolution.
  2. The "continuous" time intervals t∈[0,1]t \in [0, 1] used for the Neural ODE (e.g., OBI as [0,0.30)[0, 0.30)) act as a parameterized model coordinate rather than a historically grounded duration, meaning velocity magnitudes cannot be reliably interpreted as historical rates of script change.
  3. Paragraphs should be separated by necessary line breaks to enhance the readability of the article, such as the lengthy paragraph in the introduction, different assumptions in Page 4: Sec 2.2, highlighted module introductions (Sec. 2.3), highlighted subsubsections (Baselines, Metrics, Implementation Details in Sec. 3.1), conclusions in Sec. 3.2, and other place in this manuscript.
  4. There is a lack of comparison with the general Oracle retrieval work and methods to illustrate the validity and necessity of modeling this in the manifold space as presented in this paper.
  5. The 352-dimensional input feature incorporates 64-dimensional semantic dictionary definitions. In a true decipherment setting, such information is strictly unavailable for undeciphered characters.
  6. Supervising the extinction probability with binary labels conflates the absence of archaeological excavation with true historical extinction. The authors should reformulate the survival module using positive-unlabeled (PU) learning or explicitly calibrate the confidence threshold to account for unobserved intermediate records.
  7. There is a lack of cross-dataset validation experiments, such as using the OBC306 dataset (Huang, Shuangping, et al. "Obc306: A large-scale oracle bone character recognition dataset." 2019 International Conference on Document Analysis and Recognition (ICDAR). IEEE, 2019.) with texture noise as the query, while the retrieval database is composed of clean handwritten texts like HWOBC and EVOBC.
  8. Lack the visualization of the proposed FGCCES dataset.
  9. The organization of the paper could be further improved. Many key findings and visualizations should be moved into the main text rather than being relegated to the appendix/supplementary material.

The main concern with this paper is that MSEF is built upon supervised settings, i.e., the evolutionary relationships between image pairs are assumed to be known in advance, whereas the applicability of the proposed method to deciphering undeciphered oracle bone script characters remains insufficiently discussed. It is questionable whether the decipherment of oracle bone inscriptions should be formulated as a retrieval task. This modeling choice appears ill-suited, as decipherment involves inferring unknown meanings rather than matching against known candidates.

Questions for authors
  1. How does the selected value t (Tab. 1 and Sec. K) affect the result? Lack of hyperparameter robustness experiments for time t.
  2. The website link of BNU Character Database needs to be updated, where current link at page55: line 2925 does note point to the correct website.
Reviewer_nK5EOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes MSEF for oracle bone inscription decipherment. It treats the Chinese characters as continous evolution flow in latent space. A time-conditioned encoder was used to map the scripts from each era, and the transition between eras is learned together with a survival netowrk to predict if a character become extinct. Based on it, authors propose CBED, which performs cascade retrieval from OBI to modern script, then uses the inverse flow to verify cnandidates backward. A cross-era dataset FGCCES was built to facilitate the proposed meethods. Experiments on HUSTOBS, EVOBC, PictOBI, and FGCCES show the effectiveness.

Critical strengths
  1. The forward recall + backward verification mechanism proposed is intuitive and interesting, which models OBI evolution as a continuous dynamic across eras.
  2. The experiment results show that using intermediate scripts improve the performance a lot, adding each era bring consistent gain.
  3. Training with fragmented evolution pairs is practical for this low resource task.
Critical weaknesses
  1. Using intermediate eras for OBI is not new, EVOBC, CrossFont and Sundial-GAN already use multi-era scripts. And it remains unclear whether the gain comes from the Neural ODE/manifold modeling, as in the Table 9, the baseline CrossFont with simple cross-era consistency filtering (no ODE) already achieve the same 72.5 R@1 as the proposed MSEF, but in Table 4, it is reported only the original 54.6, the benifit of continuous dynamics itself is not shown.
  2. It seems that input features include dictionary definition and artifact context but L.3/H.3 mention that it is not established that the answer masking was actually executed, is there a possible information leakage?
  3. Many baselines are re-implemented by Paper2Code, the PictOBI overall of MSEF not consistent with its own normal/complex scores, these baseline comparisons do not look convincing enough.
  4. The paper is very hard to follow, the appendix has way too many sections (from Sec. A to Sec. V), with a lot of pre-exploratory analysis and repeated tables. Also the appendix contain many self-disclaimers, such as "source assertions, not recomputed", multi-round results are “unvalidated”, and the 140M params can not match a 12-layer 256-d transformer. This make it hard to trust the main results, I suggest authors to largely simplify the appendix and keep only the parts that really support the main claims.
  5. The expert evaluation is quite limited, only 2 experts and 100 cases, no inter-annotator agreement is reported, and the 12% “later validated” cases have no source or record, it is hard to know what this category really mean.
Questions for authors
  1. For the test queries, are the semantic (dictionary definition) and context features really zero-filled? What is the performance if all query-side semantic and context features are removed?
  2. Which baselines are re-implemented by Paper2Code, and did the authors check the reproduced results against the numbers in original papers?
  3. How a retrieval model like MSEF is evaluated on OBS-OCR (Table 2) and the 4-choice PictOBI task (Table 3)? Also how the PictOBI overall score of MSEF is computed, since it seems not consistent with the normal/complex scores.
  4. Under the same split, candidate vocabulary and reference graph, how does MSEF compare with the joint-consistent CrossFont in Table 9? Can this be added in Table 4?
Reviewer_213GOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose the Manifold-based Script Evolution Framework (MSEF) that models the evolution series— OBI → Bronze → Seal → Clerical → Regular —as the continual evolution of a manifold space. Each character is represented as an era-specific manifold point, and continuous inter-era transition rules are learned via Neural Ordinary Differential Equations; both the manifold space and transition dynamics can be trained end-to-end using character evolution pairs across any two eras. To support this, the authors construct the first unified cross-era dataset with fine-grained temporal and regional granularity. The paper further provides a pre-exploratory diagnosis of existing generative and retrieval paradigms, motivating why continuous manifold dynamics are well suited to fragmented cross-era script data, and a mechanistic post-analysis showing that the learned representations and transition dynamics capture patterns consistent with known paleographic transformations. Experiments on three commonly used benchmarks and the authors' own dataset show that MSEF achieves state-of-the-art performance among evaluated oracle-deciphering algorithms, opening the use of continuous manifold dynamics in cross-era paleographic decipherment.

Critical strengths
  1. Framing the entire script evolution series (OBI → Bronze → Seal → Clerical → Regular) as the continuous evolution of a manifold, learned with Neural ODEs, is a distinctive idea that moves beyond pairwise similarity matching. Modeling transitions as continuous dynamics naturally accommodates the irregular, non-discrete nature of historical script change.
  2. The cascaded bidirectional scheme (CBED)—forward candidate recall followed by backward consistency checking—is a principled way to exploit the full evolution chain rather than a single reference, directly targeting false matches from one-sided comparison.
  3. The pre-exploratory diagnosis of existing generative and retrieval paradigms justifies the manifold choice, and the mechanistic post-analysis showing that learned dynamics align with known paleographic transformations adds interpretability and confidence, rather than reporting accuracy in isolation.
Critical weaknesses
  1. The author’s description of the methodology is somewhat confusing. As I currently understand it, this is a generative approach—how, then, does it compare to an LLM?
  2. Language can undergo dramatic changes across different eras; what, then, is the basis for comparing distributions (such as tt-distributions) across these different periods?
  3. How can accuracy be ensured through cross-era projection retrieval and backward verification via step-by-step pruning, given that the encoder M may not be accurate?
  4. The overall methodology lacks sufficient innovation; apart from the cross-era projection and reverse verification, the rest of the content consists of material that already exists.
Questions for authors

See Weakness

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.