acceptodds
Under review as a conference paper at ICLR 2027

From Signs to Sequences: Compositional Synthesis for Cuneiform Recognition Across Periods and Collections

Abstract

Training data that pair cuneiform line images with sign sequences are scarce, even though isolated sign images and machine-readable texts are available separately. This scarcity keeps large collections of digitized tablets difficult to search and analyze automatically. We introduce a compositional supervision method that combines these independently available resources. Given a normalized sign sequence from a text corpus, we sample corresponding real sign crops from a tablet-disjoint visual dictionary (augmented for rare signs) and compose them into a synthetic line image. These image–sequence pairs pre-train three recognition architectures—a CRNN, an SVTRv2-inspired CTC model, and PaddleOCR-VL—before adaptation on a limited number of annotated tablets. In the principal evaluation, sign locations are supplied by ground-truth annotations and the input lines are reconstructed from their crops. With a fixed sign inventory, adaptation on 100 Ur III tablets yields sign error rates (SER) of 18–25% on in-period data. Removing compositional pre-training increases error by 24–52 percentage points across architectures and evaluation domains. Controlled experiments within the SVTRv2-inspired model further show that multi-sign sequence training provides substantially greater benefits than isolated-sign training, whereas preserving attested sign order and co-occurrence produces smaller, domain-dependent gains. Adaptation primarily improves recognition within the target period, while cross-period and cross-collection transfer remains difficult. An exploratory modular pipeline with predicted sign locations and reading order reaches 38.05% whole-side SER on reused in-period tablets, versus 68.92% for its strongest original geometric baseline. These results establish a practical way to repurpose existing Assyriological image and text collections as sequence-level supervision, while clarifying the remaining challenges for scalable cuneiform image-to-text systems.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
cuneiform recognition, synthetic data, compositional supervision, sequence recognition, low-resource OCR, domain generalization

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_C8w4OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a compositional supervision method for cuneiform recognition by combining real sign crops with text-derived sequences to generate synthetic training data. Experiments across three recognition models show substantial reductions in Sign Error Rate (SER), while additional analyses investigate sequence training benefits and cross-domain generalization.

Critical strengths
  1. Extensive empirical evaluation supports the effectiveness of compositional pre-training.
  2. The paper provides useful analyses of training formats and cross-domain limitations.
Critical weaknesses

1.The proposed compositional synthesis method mainly combines existing sign crops and text sequences, following established synthetic data generation approaches in OCR. Its methodological contributions beyond prior work remain limited. 2.The multi-sign training improvements may partly result from changes in input length and CTC alignment rather than better sequence understanding. Moreover, the per-sign DeiT3 classifier outperforms SVTRv2* on both cross-domain benchmarks. 3.The main experiments use ground-truth sign boundaries and reconstructed line images, which simplify the recognition task and limit the applicability of the findings to real-world cuneiform recognition.

Questions for authors

1.How well does compositional pre-training transfer to naturally cropped cuneiform lines without oracle sign localization and reconstruction? 2.In the whole-side recognition experiment, how much of the performance improvement can be attributed specifically to compositional pre-training rather than the stronger detection, classification, language modeling, and decoding components? 3.How stable are the relatively small gains from preserving attested sign order and co-occurrence across independent Stage-A pre-training seeds?

2 comments
AuthorsOpenReview ↗

Contribution, real-world applicability, attribution and seed stability

Thank you for the review; each of your three questions is answered by a new experiment below.

W1. Contribution. What is new for cuneiform is the source of supervision: two resources that exist independently and at scale, isolated sign images and transliterated corpora (here 199,243 sign crops and 387,315 text lines), are turned into sequence-level training data without any paired line images. This extends prior cuneiform work on sign detection, classification and synthesis (Dencker et al., 2020; Mikulinsky et al., 2025; Bullenkamp et al., 2026). Composing crops into lines has precedents in OCR (e.g., Senthil et al., 2020; StackMix), and we do not claim a new synthesis algorithm. Our methodological contribution is a tested route from independently collected Assyriological resources to image-to-sequence training, with controlled evidence of which parts of this supervision matter and which transfer. Under tablet- and text-disjoint controls, the supervision lowers SER by 24-52 points across three architectures. Most of the gain comes from the multi-sign line format (23-46 points on strict subsets); attested sign order adds a smaller in-period gain. The gain also holds beyond reconstructed lines with oracle sign boundaries: on natural line images (89.2% to 52.8%), on automatically found lines, and on tablets whose images are outside all our data.

W3 / Q1. Without oracle localization. The benefit carries over. On natural line crops cut from the tablet image (no stitching, no sign boundaries given to the model), Stage-A lowers in-period SER from 89.2% to 52.8%, and by 19.0 and 7.0 points cross-period and on HT. It also holds with no line annotation at test time (outputs tagged OOV are excluded, as the references exclude out-of-inventory regions). With line regions from the system's predicted rows, whole-side SER falls from 90.6% to 55.7% in-period. With an independent line segmenter trained on tablets of other periods and used without correction, it falls by 29.5 points [27.4, 31.7] in-period and by 26.7 points [22.5, 31.0] on source-disjoint tablets (Section 6.1, Appendix I, Table 23).

Q2. How much of the whole-side gain is due to pre-training. In the full system, pre-training acts only through the sequence verifier, which lowers SER by 1.83 points in-period, 2.63 cross-period and 2.18 on the source-disjoint tablets, with every other component fixed. Much of this share comes from Stage-A: with both verifier weights chosen by the same DEV rule, the Stage-A verifier beats the no-Stage-A one by 1.48, 1.91 and 1.29 points (each CI above zero), and with both fixed at the paper's weight, by 1.28 and 1.36 points in-period and cross-period. The rest of the drop from 68.92% to 38.05% comes from the stronger detection, classification and LM decoding, as Section 6.3 states; the original baselines already use Stage-A recognizers, so that comparison does not isolate pre-training. Where the sequence model itself reads the image, Stage-A is worth far more: 35.8 points as a decoder on predicted crops (32.7 with the language model), and 27-35 points on automatically found natural lines (Appendix M, Table 28; Appendix I, Table 23).

Q3. Stability across Stage-A seeds. Stable in-period, seed-dependent off-domain. With two extra seeds, the in-period gain is positive for every seed (mean 4.64 points on all lines, across-seed 95% interval [3.09, 6.19]; 2.19 on the strict subset). Off-domain, seed 456 reverses it (strict means 1.28 cross-period and 1.01 on HT), and Section 6.2 now says so. The full table is in our reply to 4j2R (Q2) and in Appendix L. The paper's large effects, Stage-A itself and the line format (W1), do not rest on this contrast. Across the three Stage-A seeds, the shuffled model's strict-subset SER moves by at most 2.1, 5.7 and 4.5 points, far less than the 45.1, 46.2 and 23.1-point line-format gain (Appendix L).

W2. Input length and CTC alignment. Output length explains only part of the line-format gap. The single-sign model does under-emit on multi-sign lines, yet on lines where both models output exactly the right number of signs, the line-format model is still far more accurate per position: 86.7% vs. 63.8% in-period, 64.7% vs. 23.8% cross-period and 31.9% vs. 11.9% on HT. The SER gap is already 33.4 points on 3-sign lines, the shortest we evaluate (Section 6.2, Appendix Q). We do not claim linguistic understanding; the attested-vs.-shuffled test (Q3) is the direct test of text structure. On DeiT3, the two models are complementary: rescoring DeiT3's top-5 candidates with the Stage-A SVTRv2* beats DeiT3 alone on all three sets (21.0% to 14.6% in-period, 27.8% to 25.1% cross-period, 62.4% to 61.7% on HT; on the strict subsets, 22.2% to 17.0% in-period and 28.3% to 26.1% cross-period), and with the no-Stage-A model the off-domain gains disappear (Section 6.1, Table 2, Appendix N).

Thank you for your response. I will maintain my rating.

Reviewer_PYBuOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper combines isolated cuneiform sign images and text sequences to create labeled training lines. It evaluates pretraining across three recognizers, historical periods, and collections, and explores automatic tablet reading.

Critical strengths
  1. Reusing existing image and text collections addresses the shortage of annotated cuneiform lines without requiring paired real-line data for pretraining.

  2. Pretraining reduces sign error by 24–52 percentage points across three recognizers. Gains persist under strict text-overlap controls, supporting its effectiveness with limited annotation.

Critical weaknesses

The practical benefit beyond the controlled setting remains insufficiently validated. With ground-truth sign locations, per-sign DeiT3 outperforms SVTRv2* by 10.6 and 8.8 percentage points on the cross-period and cross-collection tests (Table 2). In the automatic pipeline, sequence verification reduces SER from 39.88% to 38.05%, but this statistically supported gain is measured on previously inspected tablets (Table 12; Appendix D.7). Thus, the strong pretraining results have not yet translated into convincing evidence of improved automatic recognition on an untouched test set.

Questions for authors

Does SVTRv2* offer a measurable inference-speed or memory advantage over the more accurate per-sign baseline? Such an advantage would strengthen its practical justification. What is whole-side SER when the compositionally pre-trained SVTRv2* is the decoder rather than a +1.8-pp verifier, and when the verifier itself is trained without Stage-A? This determines whether your method contributes on real images.

1 comment
AuthorsOpenReview ↗

Compositional pre-training improves automatic recognition on real images and on source-disjoint tablets

Thank you. We ran the tests you named, and they show that compositional pre-training (Stage-A) contributes on real images. It measurably improves automatic recognition on tablets whose images appear nowhere in our data, and whenever the sequence model reads the image itself, removing Stage-A costs 24 points of whole-side SER or more in every automatic setting we ran, including decoding with the language model and a no-Stage-A model trained for 300 epochs.

Untouched test set. We built one: 34 Ur III sides (18 tablets and envelopes, 13 texts) from HeiCuBeDa whose images appear in none of our training, development or test data, nor in the sign dictionary. References come from the CDLI transliterations and were hashed before any model run; the language model was rebuilt without these texts. On this set the paper's frozen system reaches 46.49% whole-side SER, against 48.66% without its Stage-A verifier (2.18 points [0.64, 3.66]). The same verifier gains 1.83 points in-period and 2.63 cross-period, each CI above zero, with every other component fixed. Your two questions, on the source-disjoint set:

Source-disjoint sides, whole-side SER %No Stage-AStage-AReduction [95% CI]
Verifier, weight chosen on DEV for each (Q2b)48.66 (weight 0)47.381.29 [0.13, 2.32]
SVTRv2* as the decoder (Q2a)85.7656.6829.1 [25.0, 33.3]

Q2b. A verifier trained without Stage-A. Under the same DEV rule, the no-Stage-A verifier is given weight 0 in four of the five distinct DEV selections (the cross-period and source-disjoint runs reuse them), so it adds 0.3 points in-period and nothing on the other two sets. The Stage-A verifier is better on all three test sets: by 1.48 [0.95, 2.05] points in-period, 1.91 [1.41, 2.48] cross-period and 1.29 [0.13, 2.32] source-disjoint. With both weights fixed at the paper's 0.25, it is still better by 1.28 [0.66, 1.92] points in-period and 1.36 [0.54, 2.20] cross-period; on the 34 source-disjoint sides the difference, 0.99 [−0.58, 2.59], is not significant. More real data does not replace Stage-A: adapted on all 214 tablets used for DeiT3, a no-Stage-A model reaches 53-57% DEV LCS Recall, against 85-86% with Stage-A.

Q2a. SVTRv2 as the decoder.* On the system's predicted crops and order, the Stage-A model reaches 46.39% whole-side SER in-period, 68.29% cross-period and 56.68% source-disjoint; without Stage-A, 82.22%, 93.63% and 85.76%. With the system's six-gram language model in the decoder (its weights chosen on DEV for each model), the Stage-A model reaches 43.88% in-period, against 76.54% without Stage-A and 74.38% for a no-Stage-A model trained for 300 epochs (30.50 points [28.42, 32.46]). As a stand-alone decoder it still trails the full system (38.05% in-period), so the system uses it as a verifier on top of DeiT3, where it adds the gains above.

Real line images. Stage-A also carries over when the sequence model reads real line images without sign boundaries. On natural line crops (no stitching) it lowers in-period SER from 89.2% to 52.8%. With line regions taken from the system's predicted rows, or from an independent line segmenter trained on tablets of other periods, it lowers whole-side SER on the source-disjoint tablets by 34.4 and 26.7 points (outputs tagged OOV excluded; Section 6.1, Appendix I).

Per-sign DeiT3. With true sign boxes the two models are complementary. Rescoring DeiT3's top-5 candidates with the Stage-A SVTRv2* beats DeiT3 alone on every test set: 21.0% to 14.6% in-period (6.46 points [5.43, 7.58]), 27.8% to 25.1% cross-period and 62.4% to 61.7% on HT. On the strict subsets it lowers DeiT3's SER by 5.18 points [3.63, 6.81] in-period and 2.28 [1.53, 3.07] cross-period. Compared directly, the Stage-A combination beats the no-Stage-A one by 4.24 points [3.26, 5.27] in-period and 2.95 [2.28, 3.63] cross-period, and with the no-Stage-A model the off-domain gains over DeiT3 disappear (28.1% and 62.4%; Table 2, Appendix N). The whole-side system uses the same division of labor with predicted boxes: DeiT3 proposes, and the Stage-A sequence model verifies.

Q1. Speed and memory. Yes. On CPU (8 threads), SVTRv2* reads a line about 12 times faster than DeiT3 classifies its crops (9.6 vs. 118.4 ms median per line), with about 28 times fewer FLOPs (2.8 vs. 79.9 GFLOPs for a 5-sign line) and 11 times fewer parameters (3.5M vs. 38.7M). With each model alone on the GPU, its peak memory when reading one line is also lower (189 vs. 228 MiB). It also reads a whole line without needing a box for every sign (Appendix O).

We would be grateful if you would reconsider your score in light of these results.

Reviewer_4j2ROpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper composes isolated cuneiform sign images according to text-derived sequences, creating synthetic line supervision before adaptation on a small set of target tablets. Experiments cover three recognizers and cross-period/collection settings. Multi-sign training provides substantially larger gains than preserving attested text structure, while target-period adaptation offers limited out-of-domain improvement. The main evaluation uses lines reconstructed from ground-truth sign crops; whole-tablet recognition is evaluated separately.

Critical strengths

The main empirical result is convincing. Pre-training gains persist across three architectures, strict text-overlap controls, and an extended-training baseline, making simple memorization an unlikely sole explanation.

The study provides useful, sometimes unfavorable findings. Multi-sign training matters more than attested order, and per-sign classification performs better out of domain when boundaries are known. These are actionable insights.

The limitations are unusually transparent. Oracle localization, vocabulary exclusions, reused evaluation data, and conditional confidence intervals are explicitly distinguished from full transcription.

Critical weaknesses

Realistic validation remains the main weakness.

Questions for authors

Does compositional pre-training help on natural line images under a matched protocol? Both training and primary evaluation use stitched inputs, while Appendix I changes checkpoints and evaluation sets. A matched with/without-pre-training comparison on natural lines would clarify practical transfer.

Are the smaller text-structure gains robust to pre-training randomness? Each primary attribution arm uses one Stage-A checkpoint, whose variability is excluded from the intervals. Repeating the shuffled and attested arms would be particularly informative.

Could the whole-side pipeline be tested on a previously untouched set? Current tablets were reused during development, and DEV overlaps detector training. The existing ablation usefully isolates a 1.83-point verifier gain; an independent test would strengthen that result.

Could label mappings be harmonized and executable reproduction materials provided? CTC and VLM use different evaluation line sets, and the described supplement mainly supports rescoring saved outputs. A common-subset comparison and runnable synthesis, training, and inference code would improve verifiability.

1 comment
AuthorsOpenReview ↗

Natural lines, Stage-A seeds, a source-disjoint test set, and code

Thank you. Each question is answered by a new experiment. Together they test realistic inputs at three levels: real line images, line regions found without any annotation, and whole sides from tablets whose images never entered our data.

Q1. Natural lines under a matched protocol. Yes, by a wide margin. Each line is one rectangle cut from the tablet image, with no stitching. The two models start from the primary text-disjoint Stage-A checkpoint or from random weights and are otherwise trained with the same recipe: same 100 tablets, learning rate chosen on DEV from the same grid, three seeds. Signs outside the inventory form one explicit OOV class, and every output is scored.

SER % on natural linesNo Stage-AStage-AReduction [95% CI]
In-period (906 lines)89.252.836.5 [34.2, 38.7]
In-period, strict subset88.862.426.4 [23.3, 29.4]
Cross-period99.680.619.0 [16.6, 21.7]
HT99.292.27.0 [5.9, 8.1]

On the 670 primary in-period lines alone the reduction is 36.2 points [33.8, 38.6]. In-inventory scoring gives 38.3 points, and a no-Stage-A model trained for 300 epochs is still 32.4 points worse. The gain also holds with no line annotation at test time: with line regions from the system's predicted rows, or from an independent line segmenter trained on tablets of other periods and used without correction, Stage-A lowers whole-side SER by 34.9 and 29.5 points in-period and by 34.4 and 26.7 points on the source-disjoint tablets of Q3 (outputs tagged OOV excluded; 36-43 points if they count as errors; Section 6.1, Appendix I).

Q2. Robustness to Stage-A randomness. Yes in-period; off-domain the gain depends on the seed. We retrained the attested and shuffled Stage-A models with two more seeds and ran the full adaptation (three Stage-B seeds) for each. Gain from attested order (shuffled minus attested SER, points):

SetSeed 42 (Figure 2)Seed 123Seed 456Mean
In-period, all lines4.415.354.164.64
In-period, strict1.163.871.552.19
Cross-period, all lines4.202.77−1.341.88
Cross-period, strict3.542.21−1.911.28
HT, all lines1.722.37−0.961.04
HT, strict1.802.26−1.021.01

In-period the gain is positive for every seed, and on all lines every per-seed CI excludes zero (across-seed 95% interval [3.09, 6.19]). Off-domain, seed 456 reverses it, and Section 6.2 now says the off-domain text-structure benefit is seed-dependent. The paper's large effects, Stage-A itself (Table 1) and the line format (Figure 2), do not rest on this contrast. Across the three Stage-A seeds, the shuffled model's strict-subset SER moves by at most 2.1, 5.7 and 4.5 points, far less than the 45.1, 46.2 and 23.1-point line-format gain (Appendix L).

Q3. Whole-side test on an untouched set. Done, on 34 HeiCuBeDa Ur III sides (18 tablets and envelopes, 13 texts) whose images appear in none of our training, development or test data, nor in the sign dictionary. References come from the CDLI transliterations and were hashed before any model run; the language model was rebuilt without these texts.

Whole-side SER %In-periodCross-periodSource-disjoint
Base system, no verifier39.8860.7248.66
Paper's frozen system (Stage-A verifier)38.0558.0946.49
Reduction [95% CI]1.83 [1.15, 2.53]2.63 [1.91, 3.44]2.18 [0.64, 3.66]
Stage-A vs. no-Stage-A verifier, same DEV rule [95% CI]1.48 [0.95, 2.05]1.91 [1.41, 2.48]1.29 [0.13, 2.32]
SVTRv2* as decoder: no Stage-A / Stage-A82.22 / 46.3993.63 / 68.2985.76 / 56.68

The verifier gain holds on all three sets, and verifier weights are always chosen on DEV. Transliterations list more signs than box annotations (on in-period data they raise the same system's SER by 5.8 points), so compare within a column. Of the 387,315 Stage-A lines, 100 share text with these tablets; they are rendered from other tablets' crops (Section 6.3, Appendix M).

Q4. Harmonized mappings and runnable code. On lines whose references are token-identical under both label mappings, removing Stage-A costs (points; Appendix P):

Set (lines)SVTRv2*CRNNPaddleOCR-VL
In-period (572)49.647.226.6
Cross-period (1,155)51.748.235.0
HT (1,056)26.924.324.9

All nine are within 1.3 points of Table 1. The supplementary material now runs synthesis, Stage-A/B training, inference, scoring and every new experiment (the independent segmenter's line predictions are external), with configurations, seeds, manifests and label maps. A CPU smoke test runs every stage end to end in about 72 seconds, and the saved per-line records recompute all main-text tables and most appendix tables without retraining.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of new evidence: compositional pre-training on real images and source-disjoint tablets

Title: Summary of new evidence: compositional pre-training on real images and source-disjoint tablets

We thank the reviewers. All three reviews ask whether compositional pre-training (Stage-A) still helps once the inputs are realistic. It does in every setting below. In each with/without comparison, both models get the same data, training recipe and DEV selection rule; they differ only in whether the recognizer starts from Stage-A. The 300-epoch no-Stage-A model in item 4 is an extra, longer-trained control.

  1. Real line images (4j2R Q1, C8w4 Q1). On natural line crops, cut from the tablet image with no stitching, Stage-A lowers in-period SER from 89.2% to 52.8%, and by 19.0 and 7.0 points cross-period and on HT (Section 6.1, Appendix I).
  2. No line annotation at test time (4j2R Q1, C8w4 W3). With line regions from the system's predicted rows, or from an independent line segmenter trained only on tablets of other periods, Stage-A lowers whole-side SER by 27-35 points, both in-period and on source-disjoint tablets (Appendix I, Table 23).
  3. A source-disjoint whole-side test (4j2R Q3, PYBu). On 34 sides from tablets whose images appear nowhere in our data, the paper's frozen Stage-A verifier lowers whole-side SER by 2.18 points [0.64, 3.66]. When both verifiers' weights are chosen by the same DEV rule, the one without Stage-A gets weight 0 in four of the five distinct DEV selections, and the Stage-A verifier is better on all three test sets, by 1.29 points [0.13, 2.32] on these sides (Section 6.3, Appendix M).
  4. The sequence model as decoder (PYBu, C8w4 Q2). Reading the system's predicted crops, it reaches 46.4% whole-side SER in-period with Stage-A and 82.2% without it; with the system's language model, 43.9% and 76.5%, and 74.4% for a no-Stage-A model trained for 300 epochs.
  5. Per-sign classifier (PYBu, C8w4 W2). Rescoring the per-sign DeiT3's candidates with the Stage-A sequence model beats DeiT3 alone on all three test sets (in-period 21.0% to 14.6% SER; on strict in-period lines 22.2% to 17.0% SER); with the no-Stage-A model the off-domain gain disappears (Table 2, Appendix N).
  6. Stage-A seeds (4j2R Q2, C8w4 Q3). Two more pre-training seeds confirm the in-period text-structure gain for every seed; off-domain it depends on the seed, and Section 6.2 now says so.

We also add a common CTC/VLM test subset, an inference-cost comparison, a length/alignment analysis and runnable code. Our summary-of-changes comment lists where each item is in the revised paper.

AuthorsOpenReview ↗

Revised PDF and supplementary material uploaded

We have uploaded a revised PDF and an updated supplementary package. Changes to the submitted version are marked in blue. The title and main line are unchanged; the abstract, contributions, results and conclusion now include the new evidence, and the main text stays within 10 pages.

  1. Natural line images (Sections 3.1 and 6.1; Appendix I, Tables 22 and 23): a matched with/without Stage-A comparison on natural lines, and whole-side recognition with line regions from the system's predicted rows or from an independent line segmenter. (4j2R Q1, C8w4 W3/Q1)
  2. A source-disjoint whole-side test and the sequence model's share in the system (Section 6.3; Appendix M, Table 28): 34 sides from tablets whose images are outside all of our data, the verifier with and without Stage-A under the same DEV rule, and SVTRv2* as decoder. (4j2R Q3, PYBu Q2, C8w4 Q2)
  3. Independent Stage-A seeds for the attested-order contrast (Section 6.2; Appendix L, Table 27). (4j2R Q2, C8w4 Q3)
  4. Per-sign DeiT3 combined with SVTRv2* (Section 6.1, Table 2; Appendix N). (PYBu, C8w4 W2)
  5. Inference cost (Section 6.1; Appendix O). (PYBu Q1)
  6. A common CTC/VLM line subset with identical references (Table 1 caption; Appendix P). (4j2R Q4)
  7. Length and alignment in the training-format contrast (Section 6.2; Appendix Q). (C8w4 W2)
  8. Related work: Bullenkamp et al. (2026) (Section 2).
  9. Data corrections (Sections 4.1, 5.1 and 5.3, Table 4, Appendices B, D and F): two tablets missed by our tablet-to-text mapping (cross-period HS 1032, DEV HS 2445) and one cross-period tablet (HS 2080) whose envelope's crops are in the sign dictionary. Removing HS 1032 and HS 2080 from the cross-period set changes the contrasts in Table 1, on all lines and on the strict subsets, by at most 0.5 points.
  10. The supplementary material now contains runnable code for synthesis, training, inference, scoring and every new experiment (the independent segmenter's line predictions are external), plus a CPU smoke test. (4j2R Q4)

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.