{"id":"840ca6d1-604c-4271-89b2-107487d2e221","arxiv_id":"1908.00634","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"Fuzzy perceptual-code LSTM and convLSTM systems are reported to recognize online Arabic and Latin handwriting with best rates up to 98.5%.","lead":"Two handwriting-recognition pipelines are described: one breaks online pen traces into fuzzy primitive strokes and classifies them with LSTM, the other feeds raw coordinates to convolutional LSTM. The authors report best recognition rates above 98% on Arabic and Latin databases, but give no code, data, or architecture details.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Fuzzy ground-truth targets are computed from the same stroke-length series used as input; the 98.5% headline may be an artifact of this target construction.","rationale":"The reader's weakest-assumption analysis focused on the unproven claim that nine perceptual codes can form any script and on the hand-tuned EPC overlap constant. That is a legitimate generalization concern, but I see a more direct load-bearing issue: the strongest single result, 98.5% in Table 4, is produced by a training procedure whose target values are computed from the same stroke-length features that are fed into the network. As written, Algorithm 2 does not specify how class references are built, whether they come only from training data, or how membership degrees are normalized. This leaves open a concrete artifact: the network may simply be learning to reproduce a distance-to-prototype soft label, making the reported improvement over framewise training an artefact of target construction rather than evidence for the perceptual-code architecture. This does not require accusing the authors of dishonesty; the manuscript simply provides too little detail to rule out the artifact. A focused control experiment with a nearest-reference baseline would settle it. I also agree with the reader that the central claim is not verifiable from the manuscript: no architecture hyperparameters, no train/test splits, no error bars, and no code or data are provided. The paper's own tables contain rates from 90% to 97.5% alongside the 98%+ numbers, so the abstract's blanket statement is internally inconsistent. Because this concern directly undermines the only result that supports the headline and the missing protocol prevents reproduction, the reader's REJECT verdict is appropriate and should remain unchanged.","tokens_in":15135,"tokens_out":6746,"duration_ms":66758,"concrete_test":"Re-run the MAYASTROUN digit experiments under three conditions with the same train/test split: (A) framewise one-hot LSTM; (B) fuzzy ground-truth LSTM exactly as described in Algorithm 2; (C) a non-LSTM nearest-reference classifier that assigns each test sample to the class with smallest Euclidean distance between {li} and {lr}, with class references estimated only from the training fold. If condition (C) reaches roughly 98.5%, or if removing fuzzy targets drops accuracy well below the reported 98.5%, the result is attributable to the fuzzy-target construction rather than to the perceptual-code LSTM. The same protocol should be applied to all five databases, and the per-database numbers should be reported with error bars.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.3.2 (Algorithm 2) defines fuzzy ground-truth memberships for OnHSR-LSTM by computing the Euclidean distance between a reference class length series {lr} and the sample's segmented stroke-length series {li}. These same {li} are derived from the Beta-elliptic segmentation that constitutes the network's input representation. The training target is therefore a deterministic function of the input features plus class prototypes, so the LSTM can achieve high agreement with that target by learning a distance-to-prototype mapping, without performing the semantic perceptual-code recognition the paper claims. The headline '>98%' rests on Table 4's 98.5% MAYASTROUN row, which is exactly the fuzzy-ground-truth condition; the framewise baseline on the same database is only 94.5%. No train/test split, number of runs, or prototype construction details are given. Furthermore, Tables 4 through 9 report many rows below 98% (90.0% to 97.5%), so the abstract's unqualified 'exceeds 98%' is not supported even by the paper's own reported numbers. The load-bearing assumption is not primarily the perceptual-code coverage; it is the validity of the fuzzy-target protocol that produces the only result above 98% for system 1.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two online handwriting recognition systems: OnHSR-LSTM, which segments a script into Beta-elliptic strokes, converts each stroke into fuzzy Elementary and Basic Perceptual Codes, and classifies the resulting code sequence with an LSTM and CTC; and OnHR-convLSTM, which applies a convolutional LSTM directly to raw (x,y,z) traces. Experiments are reported on MAYASTROUN, LMCA/ADAB, UNIPEN, IRONOFF, and an in-house mobile database, with rates up to 98.5%. The abstract claims that the recognition rate exceeds 98%, but this is contradicted by several rows in the paper's own tables. The best result for system 1 is obtained under a 'Fuzzy Ground Truth' training protocol whose target memberships are computed from the same stroke-length features that form the network input, which makes the reported gain over framewise training difficult to interpret.","tokens_in":15363,"tokens_out":6028,"duration_ms":58921,"significance":"If the reported results were valid, the paper would offer a distinctive integration of perceptual-code theory with deep sequence classifiers for multilingual online handwriting, with potential applications to Arabic and Latin scripts and to tremor/robustness analysis. The paper's strengths include the use of an established Beta-elliptic segmentation model, the use of CTC for sequence alignment, and the application of data augmentation and dropout. However, the central quantitative claim is not currently supported: the abstract's >98% statement does not match the tabulated rates, the evaluation protocol is severely underspecified, and the best system-1 result relies on a training-target construction that is a deterministic function of the input representation. The contribution as presented is therefore not established, despite the plausibility of the general approach.","major_comments":[{"comment":"The abstract's claim that 'the recognition rate exceeds 98%' is not supported by the paper's own numbers. Table 4 reports framewise rates of 91.5%, 94.5%, and 90.0% and Fuzzy Ground Truth rates of 97.5%, 98.5%, and 96.0%; Table 5 reports 90.5% and 97.5%; Table 9 reports 93% and 98%. Only a few settings reach or exceed 98%, so the unqualified headline claim is contradicted by the experimental section.","section":"Abstract and Tables 4-9"},{"comment":"The Fuzzy Ground Truth training target is constructed as the Euclidean distance between a reference stroke-length series {lr} and the input-derived segmented stroke-length series {li}, both obtained from the Beta-elliptic segmentation that constitutes the network's input representation. This makes the target a deterministic function of the input, so the LSTM can learn to approximate a distance-to-prototype mapping without necessarily learning the semantic perceptual-code recognition the paper claims. The headline 98.5% MAYASTROUN result in Table 4 is exactly the Fuzzy Ground Truth condition, and no independent label source, split, or prototype construction details are given; this circularity is load-bearing and must be resolved.","section":"Section 4.3.2, Algorithm 2"},{"comment":"The experimental methodology is critically underspecified. Section 5.1 states only that 'we use 24861 samples for training and test' without giving the train/test split, writer independence, number of runs, or hyperparameter values; no error bars or confidence intervals are provided. Tables 4-9 compare the proposed systems with prior work without standardizing train/test protocols, database subsets, or evaluation measures, making the numerical comparisons unreliable.","section":"Section 5.1 and all experimental tables"},{"comment":"The load-bearing assumption that 'we use only 9 basic perceptual shapes which can form any script' is asserted without proof, coverage analysis, or an error analysis of the segmentation stage. The EPC assignment also depends on a hand-chosen overlap constant 'cst equal to /16' and on hand-defined direction-region boundaries, with no sensitivity analysis. As written, the reported rates cannot be expected to transfer to new writing styles or scripts.","section":"Sections 3.1.1.2 and 3.1.1.3"},{"comment":"The tables contain unresolved internal contradictions. Tables 6 and 9 both report IRONOFF results but give different OnHSR-LSTM rates (97.5% vs 93%) and appear to refer to different tasks (digits vs words). Section 5.2 says IRONOFF digit results are illustrated in Table 8, but Table 8 actually reports ADAB word recognition, while Section 5.3 refers to Table 8 for the mobile database when the mobile results are in Table 7. These inconsistencies make the reliability of the reported numbers impossible to assess.","section":"Tables 6, 8, and 9 and Section 5.2"},{"comment":"The noise robustness experiment reports that OnHR-convLSTM with raw (x,y,z) input drops to 50.45% accuracy, which is near chance for a digit task, while the same architecture under other conditions reaches 96-98%. No noise level, noise generation procedure, number of trials, or train/test protocol is given, so the robustness claim is not substantiated and the 50.45% figure suggests a possible training or evaluation problem that needs explanation.","section":"Section 7, Table 10"}],"minor_comments":[{"comment":"The overlap constant is written as 'cst equal to /16'; this should be cst = π/16, and the use of π should be made explicit.","section":"Section 3.1.1.3"},{"comment":"Table numbering is inconsistent: the text refers to 'Table 2' for the comparison with Tagougui et al., while the actual comparison is in Table 4; several other table references (e.g., Section 5.2 citing Table 8 for IRONOFF) are mismatched and must be corrected.","section":"Throughout"},{"comment":"The Beta function in Equation (1) contains likely typographical errors in the exponents and in the definition of tc; please verify the equation against the standard Beta-elliptic model references.","section":"Section 3.1.1.2, Equation (1)"},{"comment":"The statement 'We used 54 shapes instead of 56 Shapes' is unexplained; if two LMCA letter shapes were excluded from training or testing, the authors must state which shapes and why.","section":"Section 5.1"},{"comment":"The notation uses p for both a Beta parameter and the number of strokes (e.g., Equations (5)-(6)), which is ambiguous; please use distinct symbols.","section":"Equations (3)-(7)"},{"comment":"Figure 16 is referenced twice with different content (noise sample and perceptual illusions); the figures need to be renumbered coherently.","section":"Figures 15 and 16"}],"recommendation":"reject","confidential_remarks":"The manuscript appears to be an early draft with duplicated and misnumbered tables, an unsupported abstract claim, and a central experimental protocol whose training targets are derived from the input features themselves. The Fuzzy Ground Truth circularity is the main scientific concern, because the paper's best result depends on it and no independent validation is provided. I would not recommend inviting a revision unless the authors can supply an evaluation with independent ground-truth labels, standardized splits, error bars, and a resolution of the internal contradictions in the reported tables."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me give you the short version: the paper builds a plausible pipeline (Beta-elliptic segmentation, perceptual codes, LSTM/CTC) and adds a fuzzy ground-truth training scheme, but the reported results are not credible as presented. The headline '>98%' is contradicted by the paper's own tables, and the fuzzy target construction is close enough to a shortcut that I'd want to see a rigorous analysis before believing the accuracy.\n\nWhat's actually new: the combination of the group's earlier perceptual-code representation with an LSTM and CTC is a reasonable extension, and the fuzzy ground-truth variant (Algorithm 2) is a specific, testable idea. Testing on five databases—MAYASTROUN, LMCA/ADAB, UNIPEN, IRONOFF, plus a mobile set—is a real attempt at breadth.\n\nThe soft spots are substantial. First, the empirical section lacks almost everything needed for verification: no train/test splits, no error bars, no hyperparameter details, and no code or data. Table 4 alone shows the framewise rate at 94.5% and the fuzzy rate at 98.5% on MAYASTROUN digits; but other tables report 90.0-97.5% for various tasks, so the abstract's blanket 'exceeds 98%' is not supported even internally. The 54/56 shape exclusion and manual segmentation from ADAB are unexplained and introduce selection bias. Second, and more load-bearing, the fuzzy ground-truth targets in Algorithm 2 are computed as Euclidean distances between the sample's stroke-length series {li} and class references {lr}. Those {li} are the same features fed into the network. That means the training target is a deterministic function of the input representation, so the LSTM could learn to approximate a distance-to-prototype mapping instead of doing the perceptual recognition the paper claims. This is not automatically fatal—if the prototypes are class-averaged and the test labels are true classes, the method could still generalize—but the manuscript gives no analysis or experimental separation to rule out the shortcut. Third, the citation pattern leans heavily on the authors' own prior work for the perceptual-code machinery; that is understandable, but it means the representation itself isn't independently established here.\n\nWho should read this: people working on online handwriting recognition who want to see one group's attempt to combine perceptual features with deep sequence models. The idea is worth thinking about, but the paper as written is not a reliable source for the numbers.\n\nMy recommendation: I would not accept it in this form, and I'd ask for major revisions or reject. But I would send it to peer review rather than desk-reject, because the fuzzy ground-truth question is real and a referee could extract what's needed. If you engage with it, focus on the fuzzy training protocol and demand a proper evaluation with fixed splits and error bars.","headline":"Plausible pipeline, but the accuracy claims don't survive contact with the paper's own tables, and the fuzzy ground-truth protocol needs serious scrutiny before anyone bets on it.","tokens_in":15944,"tokens_out":3905,"would_cite":false,"duration_ms":40134,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Decomposing online handwriting into nine fuzzy perceptual stroke codes lets LSTM recognizers exceed 98 percent accuracy on Arabic and Latin digits and words.","keywords":["online handwriting recognition","Beta-elliptic model","perceptual codes","fuzzy logic","LSTM","convolutional LSTM","fuzzy ground truth training","Arabic and Latin scripts"],"falsifier":"A direct reproduction of the MAYASTROUN digit experiment where the perceptual-code front end is replaced by raw (x,y,z) coordinates, keeping the same LSTM and fuzzy training, settles whether the reported 98.5 percent is caused by the perceptual representation; if raw-input accuracy is not clearly lower, the central claim is falsified.","tokens_in":14898,"feed_emoji":"✍️","tokens_out":12009,"duration_ms":99472,"temperature":0.7,"pith_summary":"The paper proposes two online handwriting recognition systems for Arabic and Latin scripts and claims both exceed 98 percent accuracy on several databases. The first, OnHSR-LSTM, segments handwriting into strokes with a Beta-elliptic model, labels each stroke with one of nine fuzzy perceptual codes, and trains a long short-term memory (LSTM) network with fuzzy ground truth to recognize the resulting code sequences. The second, OnHR-convLSTM, learns directly from raw (x,y,z) pen signals using a convolutional LSTM. The strongest reported result is 98.5 percent recognition of Arabic digits on the MAYASTROUN database. If the claim holds, a small fixed set of perceptual primitives plus sequence learning can replace much handwriting-specific feature engineering and transfer across scripts.","feed_headline":"Fuzzy perceptual codes lift handwriting recognition past 98 percent","feed_subtitle":"A nine-shape stroke alphabet and fuzzy labels push LSTM and convLSTM models past 98 percent on Arabic and Latin digits and words.","key_machinery":"The load-bearing object is the nine-code perceptual alphabet—Valley, left-oblique shaft, shaft, right-oblique shaft, four half-occlusions, and Occlusion—which the paper asserts can form any script. The Beta-elliptic model turns the pen trace into elliptic strokes bounded by three beta-points, and each stroke receives fuzzy membership in one of four elementary perceptual code regions via a hand-set overlap constant. An LSTM maps the code sequence to a label, CTC removes timing alignment, and fuzzy ground truth replaces one-hot targets with membership values computed from Euclidean distances between stroke lengths.","core_discovery":"The paper's central claim is that online handwriting is a concatenation of perceptual codes: with the Beta-elliptic model, any trace can be segmented into elliptic strokes, each stroke receives fuzzy membership over four elementary perceptual code regions, and sequences of these codes identify characters, digits, and words. On that representation, an LSTM with CTC labeling and fuzzy ground truth training reaches 98.5 percent on MAYASTROUN Arabic digits, 97.5 percent on the UNIPEN Latin set, and up to 98.4 percent on a mobile-acquired Arabic database. The companion system, OnHR-convLSTM, reads raw (x,y,z) signals through convolutional LSTM layers and reaches 98 percent on IRONOFF words. The authors therefore claim the perceptual front end is not required for high accuracy, but that it is the component that survives injected noise, dropping only to 96 percent where the raw-signal convLSTM falls to 50.45 percent.","pith_inferences":["A natural test the authors do not run is to apply the same nine-code pipeline to another cursive script such as Farsi or Urdu; if the alphabet is truly universal, accuracy should remain in the high 90s with no architectural change.","Because fuzzy membership assigns each stroke to several codes with degrees, the method could flag low-confidence strokes for human review or active learning; the paper reports accuracies but does not analyze per-stroke ambiguity.","The 50.45 percent collapse of the raw convLSTM under noise suggests the perceptual front end functions as an implicit denoiser; comparing it against standard tremor-removal filters on Parkinsonian handwriting would quantify how much of the gain is segmentation versus preprocessing."],"forward_implications":["On Arabic digits from MAYASTROUN, OnHSR-LSTM with fuzzy ground truth reaches 98.5 percent, exceeding the circular-trajectory baseline and the smart-device recognizer reported for the same task.","Fuzzy ground truth training improves accuracy over framewise training in every reported pairing: 91.5 to 97.5 percent on LMCA/ADAB letters, 90.5 to 97.5 percent on UNIPEN set 1(c), and 90 to 96 percent on ADAB.","OnHR-convLSTM reaches 98 percent on IRONOFF words and 97 percent on ADAB words, making the raw-signal convolutional LSTM the stronger word-level pipeline in these experiments.","Under injected noise, OnHSR-LSTM keeps 96 percent accuracy while OnHR-convLSTM on raw (x,y,z) drops to 50.45 percent, supporting the claim that Beta-elliptic preprocessing removes tremor and sampling noise before classification.","The same nine-code perceptual alphabet is used for both Arabic and Latin scripts, so the system is claimed to be multilingual without changing the feature representation."],"supporting_citations":[{"why":"serves as the smart-device online-handwriting baseline whose Arabic digit accuracy the paper claims to surpass with 98.5 percent.","marker":"[12]"},{"why":"provides the Beta-elliptic handwriting generation model that segments the online trace into strokes.","marker":"[26]"},{"why":"supplies the perceptual segmentation theory used to cut scripts into elementary strokes and codes.","marker":"[19]"},{"why":"gives the fuzzy-logic representation for assigning elementary perceptual codes to each stroke.","marker":"[25]"},{"why":"provides the MAYASTROUN multilingual database on which the best Arabic digit result is measured.","marker":"[34]"},{"why":"supplies the CTC loss used to label and align the unsegmented perceptual-code sequences.","marker":"[32]"},{"why":"provides the framewise training method that the fuzzy ground truth training is measured against.","marker":"[33]"},{"why":"supplies the IRONOFF online word database on which the convLSTM system reports 98 percent.","marker":"[38]"},{"why":"provides the multi-language online handwriting system and UNIPEN comparison used for the Latin character results.","marker":"[3]"}],"fun_headline_variants":["Fuzzy stroke codes push handwriting AI past 98%","Fuzzy perceptual representation makes handwriting AI noise-proof","LSTM plus fuzzy codes hits 98% on Arabic and Latin handwriting","Perceptual stroke codes boost handwriting recognition past 98%","Fuzzy perceptual front end shields handwriting AI from noise"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that every Arabic or Latin online trace can be reliably cut into strokes that each belong to one of nine perceptual shapes, a coverage claim the paper states without proof or segmentation error analysis.","fun_headline_variants_meta":{"raw":{"variants":["Fuzzy stroke codes push handwriting AI past 98%","Fuzzy perceptual representation makes handwriting AI noise-proof","LSTM plus fuzzy codes hits 98% on Arabic and Latin handwriting","Perceptual stroke codes boost handwriting recognition past 98%","Fuzzy perceptual front end shields handwriting AI from noise"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000782,"raw_usage":{"total_tokens":3413,"prompt_tokens":864,"completion_tokens":2549,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":480,"completion_tokens_details":{"reasoning_tokens":2468}},"tokens_in":480,"tokens_out":2549,"duration_ms":17090,"temperature":1.0,"reasoning_tokens":2468,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:42:05.101997+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct reproduction of the MAYASTROUN digit experiment where the perceptual-code front end is replaced by raw (x,y,z) coordinates, keeping the same LSTM and fuzzy training, settles whether the reported 98.5 percent is caused by the perceptual representation; if raw-input accuracy is not clearly lower, the central claim is falsified.","supporting_citations":[{"cited_title":"Recognition of online handwriting with variability on smart devices,","cited_arxiv_id":null,"evidence_quote":"serves as the smart-device online-handwriting baseline whose Arabic digit accuracy the paper claims to surpass with 98.5 percent."},{"cited_title":"A developmental theory of the relation between geometry and kinematics in handwriting,","cited_arxiv_id":null,"evidence_quote":"provides the Beta-elliptic handwriting generation model that segments the online trace into strokes."},{"cited_title":"The pertohs theory for on-line handwriting segmentation,","cited_arxiv_id":null,"evidence_quote":"supplies the perceptual segmentation theory used to cut scripts into elementary strokes and codes."},{"cited_title":"A new approach for the extraction of handwriting perceptual codes using fuzzy logic,","cited_arxiv_id":null,"evidence_quote":"gives the fuzzy-logic representation for assigning elementary perceptual codes to each stroke."},{"cited_title":"Mayastroun: A multilanguage handwriting database,","cited_arxiv_id":null,"evidence_quote":"provides the MAYASTROUN multilingual database on which the best Arabic digit result is measured."},{"cited_title":"Framewise and ctc training of neural networks for handwriting recognition,","cited_arxiv_id":null,"evidence_quote":"provides the framewise training method that the fuzzy ground truth training is measured against."},{"cited_title":"The ireste on/off (ironoff) dual handwriting database,","cited_arxiv_id":null,"evidence_quote":"supplies the IRONOFF online word database on which the convLSTM system reports 98 percent."},{"cited_title":"Multi-language online handwriting recognition,","cited_arxiv_id":null,"evidence_quote":"provides the multi-language online handwriting system and UNIPEN comparison used for the Latin character results."}],"review_version":1}