Paper deep dive
HANS: A Handwritten Answer Sheet Dataset for Noisy Hybrid Document Parsing
Xiazhen Wu, Wansong Qin, Yangbin Zheng, Liangda Fang, Zhan Li, Xiujie Huang, Liushen Zhou, Quanlong Guan
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 90%
Last extracted: 10/11/2026, 2:13:38 AM
Summary
This paper introduces HANS, a dataset of handwritten student answer sheets featuring complex interleaved text, math, and noise artifacts like strikethroughs. It also proposes NA-GOT, an end-to-end recognition framework that uses a Lightweight Noise Suppression Module (LNSM) and noise-aware attention to improve accuracy on this noisy data.
Entities (11)
Relation Signals (7)
NA-GOT ā contains ā Lightweight Noise Suppression Module
confidence 95% Ā· we design a Lightweight Noise Suppression Module (LNSM) that attenuates noise at the feature level.
NA-GOT ā contains ā Weighted Attention Correction
confidence 95% Ā· we introduce a noise-aware attention mechanism... Weighted Attention Correction (WAC) introduces a negative bias
NA-GOT ā extends ā GOT
confidence 95% Ā· NA-GOT, an end-to-end noise-aware recognition framework extending the GOT [21] architecture.
HANS ā usedfor ā NA-GOT
confidence 95% Ā· Building upon HANS, we propose NA-GOT... Experimental results demonstrate that HANS poses substantial challenges
Lightweight Noise Suppression Module ā uses ā Noise Mask Head
confidence 90% Ā· We takes the patch-level noise labels... and the output features V... as inputs to the Noise Mask Head (NMH)
YOLOv8 ā usedfor ā Noise Region Detection
confidence 85% Ā· YOLOv8 [6] is adopted for the automatic localization of student correction regions
DeepSeek-OCR ā usedfor ā content recognition
confidence 85% Ā· DeepSeek-OCR [22] is employed for preliminary content recognition of answer sheets
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Intelligent grading and automated scoring technologies constitute critical infrastructure for smart education. However, existing document parsing and handwriting recognition benchmarks are predominantly designed for well-structured printed documents or isolated mathematical expressions, lacking datasets that capture the complex characteristics inherent to student answer sheets, including multi-line derivation processes, heterogeneous mixtures of text and mathematical formulae, and noise artifacts such as strikethroughs. To address this gap, we introduce HANS, the first dataset explicitly constructed for real-world educational scenarios, encompassing mathematical expressions, natural language text, hand-drawn tables, and diverse noise patterns including corrections and deletions, accompanied by fine-grained annotations that establish a reliable foundation for robust recognition research. Building upon HANS, we propose NA-GOT, an end-to-end framework that achieves two-stage noise suppression through a lightweight noise suppression module operating at the feature level, complemented by a noiseaware attention mechanism incorporated into the decoding stage. Experimental results demonstrate that HANS poses substantial challenges to existing methods, while NA-GOT achieves significant improvements in both accuracy and stability for answer process recognition. The dataset will be made publicly available upon publication.
Tags
Links
- Source: https://arxiv.org/abs/2610.12363v1
- Canonical: https://arxiv.org/abs/2610.12363v1
Trouble viewing inline? Open PDF directly ā
Full Text
39,038 characters extracted from source content.
Expand or collapse full text
HANS: A Handwritten Answer Sheet Dataset for Noisy Hybrid Document Parsing Xiazhen Wu, Wansong Qin, Yangbin Zheng, Liangda Fang, Zhan Li, Xiujie Huang, Liushen Zhou, and Quanlong Guan Jinan University, Guangzhou, China wuxiazhen@stu2024.jnu.edu.cn songbai,zybbyz@stu2025.jnu.edu.cn fangld,lizhan,t_xiujie,crystal,gql@jnu.edu.cn Abstract. Intelligent grading and automated scoring technologies con- stitute critical infrastructure for smart education. However, existing doc- ument parsing and handwriting recognition benchmarks are predomi- nantly designed for well-structured printed documents or isolated math- ematical expressions, lacking datasets that capture the complex charac- teristics inherent to student answer sheets, including multi-line deriva- tion processes, heterogeneous mixtures of text and mathematical formu- lae, and noise artifacts such as strikethroughs. To address this gap, we introduce HANS, the first dataset explicitly constructed for real-world educational scenarios, encompassing mathematical expressions, natural language text, hand-drawn tables, and diverse noise patterns including corrections and deletions, accompanied by fine-grained annotations that establish a reliable foundation for robust recognition research. Build- ing upon HANS, we propose NA-GOT, an end-to-end framework that achieves two-stage noise suppression through a lightweight noise sup- pression module operating at the feature level, complemented by a noise- aware attention mechanism incorporated into the decoding stage. Exper- imental results demonstrate that HANS poses substantial challenges to existing methods, while NA-GOT achieves significant improvements in both accuracy and stability for answer process recognition. The dataset will be made publicly available upon publication. Keywords: Smart educationĀ· Handwritten document parsingĀ· HANS dataset 1 Introduction Intelligent grading and automated scoring technologies are increasingly emerg- ing as pivotal driving forces in the advancement of smart education. In con- trast to outcome-oriented evaluation paradigms that focus exclusively on final answers, modern pedagogical practice, particularly in mathematics education, places greater emphasis on understanding and diagnosing studentsā problem- solving thought processes. This demands that intelligent systems be capable of recognizing the derivation steps and natural language explanations written by arXiv:2610.12363v1 [cs.AI] 8 Oct 2026 2X. Wu et al. ę¾³ Fig. 1: Illustration of representative characteristics of the HANS dataset. HANS is a real-world student handwritten mathematics answer sheet dataset, encompassing the dual challenges of complex formula-text interleaving (blue bounding boxes) and correction noise artifacts (red bounding boxes). students on answer sheets, thereby enabling step-wise scoring, error cause anal- ysis, and personalized feedback. Consequently, the accurate recognition of stu- dentsā handwritten problem-solving processes constitutes a critical component of intelligent grading systems, as its accuracy and robustness directly determine the viability of downstream scoring and diagnostic tasks. Recent years have witnessed substantial progress in document parsing and handwriting recognition. Document parsing methods are predominantly designed for well-structured printed documents, with representative approaches includ- ing multi-stage pipeline [8] processing and end-to-end vision-language models [10,15,16,20,23], exemplified by benchmark datasets such as OmniDocBench [13] and FOX [11]. Concurrently, research on handwriting recognition using classi- cal benchmarks ā including IAM [12], CROHME [24], and ICDAR Math ā has driven an architectural evolution from CNN and RNN-CTC paradigms to- ward Transformer-based [17] encoders with Seq2Seq decoders and unified vision- language modeling frameworks. Nevertheless, existing benchmarks exhibit several critical limitations. 1) Pre- vailing datasets predominantly focus on printed academic documents, neglecting real-world handwritten document types such as student answer sheets. 2) There is a systematic absence of research on handwriting noise: most benchmarks as- sume high document quality and lack systematic annotation and targeted evalua- tion of common noise artifacts such as strikethroughs, erasures, and overwriting. 3) Existing handwriting datasets are confined to either pure text lines or isolated mathematical expressions; in particular, handwritten mathematical expression recognition has largely been restricted to single-line expressions, with limited investigation into continuous multi-line writing scenarios involving interleaved text and formulae. HANS: Noisy Hybrid Document Parsing3 To address these limitations, we introduce HANS, the first dataset explicitly constructed for real-world educational scenarios. HANS is sourced from scanned student answer sheets collected during eighth-grade mathematics quality as- sessment examinations conducted over three semesters in a municipal school district, specifically from the open-ended mathematical problem-solving sec- tions. To protect student privacy, all personally identifiable information was cropped prior to data release and use. Unlike prior datasets comprising pure text or isolated formulae, the student problem-solving processes in HANS en- compass mathematical expressions, natural language text, hand-drawn tables, and manual strikethroughs, faithfully reflecting the typical interplay between mixed handwriting and structured writing characteristic of authentic exami- nation contexts, as shown in Fig. 1 . More critically, we provide fine-grained annotations across distinct writing modalities within the problem-solving pro- cess, including content-level annotations for expressions, text, and tables, as well as spatial annotations for noise elements such as strikethroughs and deletions. These annotations furnish a supervisory foundation for explicitly distinguishing noise from valid handwriting, thereby supporting subsequent robust modeling and analysis. Building upon the noise annotations provided by HANS, we further pro- pose NA-GOT, an end-to-end noise-aware recognition framework extending the GOT [21] architecture. The proposed framework suppresses noise interference through a two-stage mechanism. In the first stage, we design a Lightweight Noise Suppression Module (LNSM) that attenuates noise at the feature level. In the second stage, we introduce a noise-aware attention mechanism that reduces at- tention weights assigned to noise-bearing regions via learned noise bias terms. Experimental results demonstrate that incorporating the noise-aware mechanism yields significant improvements in both recognition accuracy and output stability for answer sheet problem-solving process recognition. In summary, the contributions of this paper are as follows: 1. We construct HANS, a real-world educational answer sheet dataset en- compassing mixed handwritten problem-solving processes with heterogeneous content types including mathematical expressions, natural language text, and ta- bles. Fine-grained annotations are provided, with particular emphasis on spatial bounding box annotations for real-world noise artifacts such as strikethroughs and deletions, establishing a foundation for robust recognition research. 2. We propose NA-GOT, an end-to-end noise-aware recognition framework that achieves two-stage noise suppression through a Lightweight Noise Suppres- sion Module(LNSM) and a noise-aware attention mechanism. 3. Extensive experimental results demonstrate that the HANS dataset presents substantial challenges to current state-of-the-art methods, and validate that the proposed approach significantly outperforms existing methods. 4X. Wu et al. 2 Related Work 2.1 Document Parsing and Handwriting Recognition Datasets In the domain of document parsing, datasets such as DocLayNet [14] and Om- niDocBench [13] provide important benchmarks for document layout segmen- tation and multi-task learning, with particular significance for layout element detection and segmentation. Concurrently, datasets including FUNSD [7] and FOX have advanced the parsing of form and handwritten documents, encom- passing diverse components such as tables, text, and formulae while extend- ing support to more complex document structures. The data distributions of these benchmarks are predominantly composed of printed text, with relatively limited representation of handwritten content, and systematic investigation of noise-related challenges remains largely absent. Existing handwriting recogni- tion datasets, such as IAM, CROHME, and MathWriting [5] , primarily focus on offline handwriting recognition and mathematical formula transcription tasks. Although substantial progress has been achieved in handwritten text and math- ematical expression recognition, these datasets continue to suffer from limited diversity in writing modalities. 2.2 Document Parsing and Handwriting Recognition Models In the area of generative document understanding, models such as Nougat [2] and GOT-OCR2.0 realize end-to-end conversion from document images to structured text via visual Transformer architectures, supporting joint modeling of heteroge- neous content types. LayoutLM [25] achieves strong performance through joint modeling of layout and content by integrating visual and textual information within a unified framework. In the field of handwriting recognition, TrOCR [9] adopts an architecture comprising an image Transformer encoder coupled with a text Transformer decoder, enabling efficient end-to-end handwritten text tran- scription. In the domain of handwritten mathematical expression recognition, works such as TAMER introduce tree-structure-aware modules to jointly opti- mize sequence prediction and tree structure prediction tasks, thereby improv- ing model generalization over complex mathematical expression structures and demonstrating strong adaptability ā particularly for the recognition of intricate formulae and expressions. Despite the remarkable progress achieved in existing document parsing and handwriting recognition techniques, current methodologies remain inadequate in addressing the noise-corrupted and structurally complex handwriting character- istics inherent to educational answer sheets. Consequently, future research should build upon existing datasets and methods to develop more specialized datasets and more adaptive model frameworks, with the aim of enhancing robustness and accuracy in the educational answer sheet recognition task. HANS: Noisy Hybrid Document Parsing5 Table 1: Comparison of HANS against existing datasets.HS denotes the number of handwritten samples,and Level indicates the sample granularity(page or line). HME, HT, and NA indicate the presence of handwritten mathematical expres- sions,handwritten text, and noise annotations,respectively. DatasetHSLevelHMEHTNA OmniDocBench 116Pagesāā C-OCR [26] 200Pagesāā FOX0Pagesā IAM1539Linesāāā CROHME8836Linesāā HANS(ours) 5213 Pagesā 3 HANS Dataset 3.1 Dataset Construction and Statistics We construct HANS, a dataset designed for real-world educational answer sheet scenarios. The data are sourced from eighth-grade mathematics quality assess- ment examinations conducted over three semesters in a municipal middle school district in China, specifically from the open-ended problem-solving sections com- prising six mathematical questions per examination, drawn from scanned student answer sheets, as shown in Tab. 1. The dataset contains 5,213 answer question images, covering 745 students across 18 distinct problem types. To protect stu- dent privacy, all personally identifiable information was cropped prior to data release and use. The writing patterns in HANS exhibit considerable complexity. Within a single page, natural language explanations and mathematical expressions fre- quently appear in interleaved configurations, accompanied by extensive multi- line derivation steps, chains of equations, and cross-line intermediate results that collectively encode rich procedural information. The mathematical expressions themselves present high structural complexity, commonly featuring fractions, radicals, superscripts and subscripts, nested parentheses, and diverse geometric and logical symbols, further compounding the difficulty of symbol-level accurate transcription and structural consistency constraints. In addition, HANS con- tains a non-trivial proportion of hand-drawn tables, whose cells typically contain both natural language and mathematical content simultaneously, substantially increasing the coupling between structural layout recognition and content tran- scription. Furthermore, the answer sheets exhibit pervasive real-world noise and correction behaviors characteristic of authentic examination contexts, including erasures, strikethroughs, and overwriting, which introduce stroke discontinuities, stroke superposition, local occlusion, and structural degradation. These artifacts result in severe spatial entanglement between valid handwriting and noise traces, significantly elevating the uncertainty and difficulty faced by recognition models. 6X. Wu et al. ę¾³ ę¾³ ŅÆŌą¦Óą”Ņ A^_D^_^_CBJO\ ŅæŠ #^#_ Dac 2GHKR TUOYK (HU AC .KOMNZ =OJZN Č 4UOY_GTTUZGZOUT ._HXOJ GTTUZGZOUT Fig. 2: Characteristics of the HANS annotated dataset. Each sample in the dataset is annotated with both the complete problem-solving process transcription and the corresponding correction noise regions. Blue bounding boxes denote the annotation of interleaved mathematical expressions and natural language text, while red bounding boxes indicate the spatial annotations of correction noise artifacts. 3.2 Annotation Details Unlike benchmarks that provide only final answer transcriptions or annotations for isolated formulae, we provide annotations more closely aligned with the re- quirements of automated grading. The fundamental sample unit of the dataset is defined as the answer region image corresponding to each open-ended problem, with each sample containing two categories of annotations: (1) problem-solving process recognition annotations, and (2) non-semantic noise region annotations for erasures and strikethroughs. Full details are illustrated in Fig. 2. Problem-Solving Process Recognition Annotations Each sample is pro- vided with a recognition transcription corresponding to the complete problem- solving process visible in the image; content that has been deleted by the student is excluded from the final transcription. In terms of structural representation, the problem-solving process is segmented according to the semantic information of each handwritten line, strictly preserving the original line breaks as written. The logical ordering between lines is inferred from handwriting layout conventions, with line break symbols employed to denote logical sequencing. Noise Region Detection Annotations For images containing noise, we pro- vide pixel-level annotation files for noise regions, stored in JSON format. All noise instances within an image are uniformly assigned the label noise and are permitted to adopt one of two geometric annotation structures: block-level era- sures or strikethroughs are annotated using rectangle, defined by two diagonal corner coordinates, while irregularly shaped corrections or localized ink clusters are annotated using polygon, defined by a closed multi-point contour. In cases HANS: Noisy Hybrid Document Parsing7 where noise regions partially overlap with valid content, conservative boundary fitting is applied to minimize erroneous coverage of effective derivation areas. For strikethrough scenarios, annotations primarily delineate the extent of visi- ble occlusion caused by deletion strokes. For overwriting scenarios, annotations encompass both the discarded content and the superimposed strokes, while en- deavoring to preserve the final legible and correct writing. For erasure scenarios, annotations cover both erasure residuals and regions rendered illegible by the erasing action. 3.3 Annotation Pipeline To ensure both annotation quality and operational efficiency, we design a three- stage pipeline for the HANS annotation task, comprising automatic pre-annotation, annotator correction, and expert quality inspection. Automatic pre-annotation. We employ state-of-the-art detection and recog- nition models to generate pre-annotations for both noise region detection and problem-solving process transcription. Specifically, YOLOv8 [6] is adopted for the automatic localization of student correction regions, while DeepSeek-OCR [22] is employed for preliminary content recognition of answer sheets, encom- passing annotations for text, mathematical expressions, and tables. Annotator correction. Following automatic pre-annotation, human anno- tators systematically revise the initial annotation results. For student correction regions, annotators identify and supplement any missed noise areas, and refine the bounding box boundaries to more precisely conform to the extents of the correction artifacts. Each character is individually verified to ensure transcrip- tion accuracy. With respect to the overall problem-solving process, annotators perform character-level inspection and correction of the preliminary recognition drafts, ensuring the accurate representation of text, mathematical symbols, and their structural relationships. Omitted content is supplemented where necessary, and output formatting is standardized to satisfy the requirements of subsequent model training and evaluation. Quality inspection. Despite the comprehensive corrections performed by annotators, the structural complexity of mathematical formulae and tables may nonetheless introduce residual errors. Accordingly, domain experts conduct a dedicated quality inspection pass to verify annotation accuracy and ensure con- sistency across the dataset. 4 Method We propose NA-GOT, a noise-aware end-to-end answer sheet recognition frame- work, as illustrated in Fig. 3, designed to mitigate the interference of real-world noise artifacts, including erasures, strikethroughs, and overwriting on the ac- curate recognition of valid handwritten characters. Building upon GOT, the proposed framework suppresses noise through a two-stage mechanism. In the 8X. Wu et al. Fig. 3: Overall architecture of the NA-GOT framework. first stage, we design a Lightweight Noise Suppression Module (LNSM) that at- tenuates noise influence at the feature level. In the second stage, we introduce a Noise-aware attention mechanism that reduces attention weights assigned to noise-bearing regions via learned noise bias terms. Specifically, the input image is first processed by a visual encoder to extract patch-level features, while a mask matrix generated from noise annotations provides noise-aware supervision sig- nals during training. The noise-aware module performs weighted suppression of features according to the predicted noise probabilities, and subsequently modu- lates the decoderās attention distribution to direct the modelās focus toward valid handwriting regions. The resulting visual features are then projected via a linear layer and fed into the decoder, which generates the final Markdown-formatted output, achieving complete transcription of natural language text, mathematical formulae, and step-by-step derivations. 4.1 Lightweight Noise Suppression Module We first construct patch-level supervision signals aligned with the visual patch tokens from the noise annotations. Specifically, for each image, all annotations are first converted into a binary pixel mask matrix M according to the original image resolution. To maintain consistency with the patch token layout of the GOT visual encoder, M is uniformly resized to a resolution of 1024Ć1024, fol- lowed by 16Ć16 average pooling to compute the noise occupancy ratio over the patch grid. The resulting grid is subsequently flattened into patch-level noise la- bels m g t in one-to-one correspondence with the visual tokens, where each entry in m g t represents the proportion of noise pixels within the corresponding patch. We takes the patch-level noise labels m g t and the output features V of the visual encoder as inputs to the Noise Mask Head (NMH), which determines whether each patch constitutes a noise patch. Concretely, NMH employs a two- layer MLP to predict a noise confidence score p for each patch token: HANS: Noisy Hybrid Document Parsing9 p = Ļ(MLP(V )) This design enables the model to adaptively predict noise distributions during inference without relying on manual annotations or additional noise inputs, while simultaneously providing stable noise priors for subsequent token gating and the decoderās noise-aware attention bias, thereby enhancing the robustness of downstream sequence generation. Leveraging the obtained noise confidence scores p, we further suppress the representational magnitude of noise tokens at the visual token level. Specifically, the noise confidence scores p are mapped to gating coefficients, which are applied to scale the visual tokens on a per-token basis: Ģ V = V ā 1ā Ī±Ā· p where ā denotes element-wise multiplication, and Ī± is a hyperparameter controlling the gating strength. 4.2 Weighted Attention Correction While LNSM modulates the representational magnitude of visual tokens, the decoder may nevertheless assign substantial attention weights to noise tokens. Since NA-GOT concatenates visual patch tokens and text tokens into a unified sequence, cross-modal alignment occurs within the decoder. To suppress the interference of noise visual tokens on the generation process, Weighted Attention Correction (WAC) introduces a negative bias derived from the noise probabilities p output by NMH into the decoderās self-attention layers: b =āĪ²Ā· p where b denotes the bias term and Ī² is a hyperparameter controlling the penalty strength. A higher noise probability yields a larger bias, resulting in a lower attention weight for the corresponding patch following the softmax oper- ation. This mechanism penalizes noise patches during attention computation, di- recting the decoderās focus toward valid handwriting regions, thereby improving both the accuracy and structural integrity of the generated Markdown output in complex handwritten answer sheet scenarios. 4.3 Loss Function The NMH predicts a noise confidence score p for each visual patch token, and the noise labels m gt are constructed from the noise annotations. A binary cross- entropy loss with positive sample weighting is applied to compute the noise mask loss L mask . Given the input image and the corresponding prompt, the model autoregressively generates the target sequence, with cross-entropy loss 10X. Wu et al. computed as the primary recognition loss L ocr . The final training objective is defined as: L =L ocr + Ī» mask L mask where Ī» mask denotes the noise branch weight, controlling the influence of noise supervision on the primary recognition task. 5 Experiments 5.1 Experimental Setup Implementation Details We employ the proposed NA-GOT model for both training and evaluation. For the LNSM, patch-level noise supervision signals are generated from noise annotations during training to supervise the NMH in predicting noise distributions. The key hyperparameters adopted are Ī± = 0.7 and Ī² = 2.0. Dataset and Evaluation Metrics We conduct all experiments on the pro- posed HANS dataset to evaluate whether the proposed method effectively ad- dresses the robustness challenges inherent to answer sheet recognition. The dataset is partitioned into training and test sets at an 8:2 ratio, with samples containing noise regions comprising three-quarters of the training set. Evaluation is performed using sequence alignment metrics commonly adopted in document parsing research, including BLEU, Edit Distance, METEOR, Precision, Recall, and F1-score. Baselines To comprehensively evaluate the proposed method, we select three categories of representative baselines for comparison: (1) General VLMs: Qwen3- VL [1] and InternVL3.5 [19]; (2) Specialized VLMs: DeepSeek-OCR2, PaddleOCR- VL1.5 [3], MinerU2.7.6-VLM [18], and GOT; (3) Pipeline Tools: PaddleOCR P-StructureV3 [4] and MinerU2.7.6-Pipeline. To provide a rigorous compar- ison, GOT is additionally fine-tuned on the full HANS training set prior to evaluation, thereby demonstrating the superiority of the proposed approach. 5.2 Comparison to Previous Approaches We present a comprehensive comparison of NA-GOT against prior methods on the HANS dataset in Tab. 2, with all baselines organized into three groups by model category. NA-GOT consistently surpasses all baselines across General VLMs, Specialized VLMs, and Pipeline Tools. Pipeline Tools exhibit notably inferior performance on this task, while General VLMs and Specialized VLMs achieve moderate performance. Among these, Qwen3-VL delivers the strongest results, indicating that powerful general-purpose VLMs possess considerable doc- ument understanding and generation capabilities, yet remain susceptible to noise HANS: Noisy Hybrid Document Parsing11 Table 2: Experimental results of NA-GOT and baseline models on the HANS dataset. MethodBLEUā Edit Distanceā METEORā Precisionā Recallā F1-scoreā PaddleOCR P-V30.177 0.6590.3360.6240.546 0.566 MinerU2.7.6-pipeline 0.064 0.7340.2180.4820.431 0.439 Qwen3-VL0.433 0.4450.5860.7340.731 0.724 InternVL3.50.196 0.6500.3640.6600.554 0.650 Paddle-VL1.50.349 0.5260.4970.7220.665 0.678 MinerU2.7.6-VLM0.274 0.5840.4850.6110.677 0.629 Deepseek-OCR20.338 0.5330.5050.6550.665 0.649 GOT0.098 0.7890.2310.4260.470 0.409 GOT(fine tuned)0.577 0.3240.6870.8120.806 0.801 NA-GOT(fine tuned) 0.594 0.3060.7010.821 0.813 0.810 interference in the challenging noisy answer sheet scenario. GOT transferred di- rectly to this dataset yields relatively low performance, suggesting a substantial domain shift between this task distribution and the original training data ā particularly with respect to noise patterns and the heterogeneous output re- quirements imposed by interleaved formula-text handwriting. This further un- derscores the value of HANS in filling a critical data gap in the document parsing domain. In contrast, GOT fine-tuned on HANS achieves significant performance gains, validating the effectiveness of our annotations in enabling the model to rapidly adapt to the target data distribution. Compared to the fine-tuned GOT baseline, our proposed method achieves consistent improvements across all eval- uation metrics, with particularly pronounced gains on heavily noisy samples. These results demonstrate that the proposed Lightweight Noise Suppression Module(LNSM) effectively learns to extract noise-invariant features, while the Noise-aware attention mechanism penalizes noise-bearing positions at the atten- tion layer, thereby mitigating noise-induced misalignment and repetitive gener- ation. 5.3 Ablation Study We conduct an ablation study by removing individual noise-aware components from the full NA-GOT model to evaluate their respective contributions, as re- ported in Tab. 3. The w/o WAC variant achieves a BLEU score of 0.589 and an Edit Distance of 0.314, indicating that the noise prediction and token gat- ing mechanism alone can effectively suppress noise interference at the feature level. The w/o token gating variant yields lower BLEU and F1-score than the w/o WAC variant, suggesting that feature-level token gating contributes more substantially to robust recognition than decoder-level attention correction alone. Nevertheless, both ablated variants underperform the full NA-GOT model. The full model achieves the best overall performance across all evaluation metrics, 12X. Wu et al. Table 3: Ablation study of token gating and Weighted Attention Correction (WAC) in NA-GOT. MethodBLEUā Edit Distanceā METEORā Precisionā Recallā F1-scoreā NA-GOT0.594 0.3060.7010.8210.813 0.810 w/o WAC0.589 0.3140.6980.8200.809 0.808 w/o Token Gating 0.580 0.3210.6910.8150.808 0.803 Table 4: Analysis of the influence of hyperparameterĪ± on model performance. Ī± BLEUā Edit Distanceā METEORā Precisionā Recallā F1-scoreā 0.5 0.592 0.3070.6980.8180.806 0.807 0.7 0.594 0.3060.7010.8210.813 0.810 0.9 0.593 0.3060.6990.8200.810 0.809 demonstrating that token gating and WAC provide complementary benefits at the representation and decoding stages, respectively. 5.4 Hyperparameter experimentation Furthermore, we conduct ablation experiments on the two key hyperparame- ters, Ī± and Ī². As reported in Tab. 4, the gating strength coefficient Ī± is found to exert a significant influence on model performance. When Ī± = 0.5, noise suppression is insufficient, allowing noise tokens to propagate into the decoder with considerable magnitude, resulting in marginally lower BLEU and F1-score. When Ī± = 0.9, excessive suppression impairs a portion of valid features, lead- ing to slight degradation in BLEU and METEOR. The model achieves optimal overall performance at Ī± = 0.7, indicating that moderate representation-level noise suppression effectively reduces noise-induced erroneous generation without sacrificing coverage of critical information. The hyperparameter Ī² governs the penalty strength imposed on noise patches at the attention layer, with results reported in Tab. 5. When Ī² = 1.0, the penalty is insufficient, causing the model to produce erroneous outputs in certain noise- bearing regions. When Ī² = 3.0, the excessively large penalty introduces mild over-suppression, marginally weakening the decoderās attention allocation to- ward valid patches. At Ī² = 2.0, the model exhibits a stronger tendency to ag- gregate information from non-noise regions, yielding optimal performance across all evaluation metrics. 5.5 Visual Analysis The recognition results from four modelsāNA-GOT, Qwen3-VL, DeepSeek- OCR2, and MinerU2.7.6-pipelineāare compared in Figure 4. NA-GOT demon- strates stronger robustness and generation stability in noisy areas, making fewer HANS: Noisy Hybrid Document Parsing13 Table 5: Analysis of the influence of hyperparameterĪ² on model performance. Ī² BLEUā Edit Distanceā METEORā Precisionā Recallā F1-scoreā 1.0 0.593 0.3070.7000.8190.811 0.810 2.0 0.594 0.3060.7010.8210.813 0.810 3.0 0.591 0.3080.6970.8170.807 0.808 Fig. 4: Visualization Results.Figure 4 presents a sample of answer sheet image recog- nition. The upper part shows the original answer sheet image, with the red box indicat- ing the noise areas due to student corrections. The lower part displays the recognition results from four modelsāNA-GOT, Qwen3-VL, DeepSeek-OCR2, and MinerU2.7.6- pipelineāwith red boxes highlighting the recognition errors caused by noise regions in the image errors in misinterpreting erasure textures or strikethrough lines as characters or symbols. It maintains a reasonable line-level structure and derivation or- der.Additionally, NA-GOT shows a higher fidelity to structured formulas, with fractions, parentheses, and equality chains more consistent, and key variables and operators more coherent, thus reducing semantic bias caused by omissions, mistakes, or structural damage. In contrast, general multimodal models like Qwen3-VL, although generally readable, are more prone to misreading similar-looking variables and numbers near noise boundaries, and tend to experience local misalignment or omission in complex derivations. Specialized VLMs, such as DeepSeek-OCR2, are more prone to decoding instability or structural misreadings under strong noise condi- tions. For instance, in the red box, DeepSeek-OCR2 shows noticeable repetitive outputs, rendering the entire derivation unusable. Pipeline methods, such as MinerU2.7.6-pipeline, are more likely to miss content or make incorrect identifi- cations, leading to missing steps, disjointed content, or even unrelated structures being introduced. Overall, this visualization case demonstrates that NA-GOT effectively mit- igates the misattention and misalignment issues induced by noise, significantly 14X. Wu et al. suppressing the degradation caused by repetitive generation in long sequences. As a result, it provides more stable, structured, and accurate recognition out- puts, which are closer to being usable for subsequent automatic grading and diagnosis in real answer sheet scenarios. 6 Conclusion In this paper, we address the problem of recognizing studentsā handwritten problem-solving processes on answer sheets in smart education scenarios. We introduce HANS, the first dataset specifically constructed for this task, encom- passing mathematical expressions, natural language text, hand-drawn tables, and diverse noise artifacts including strikethroughs and deletions, with fine- grained annotations that provide a reliable foundation for robust recognition research. Furthermore, we propose NA-GOT, an end-to-end noise-aware recog- nition framework that achieves two-stage noise suppression through a lightweight noise suppression module and a noise-aware attention mechanism. Experimen- tal results demonstrate that the HANS dataset presents substantial challenges to existing methods, while NA-GOT achieves significant improvements in both recognition accuracy and output stability for problem-solving process transcrip- tion, offering an effective solution for handwritten document parsing in real-world educational scenarios. References 1. Bai, S., Cai, Y., Chen, R., Chen, K., Chen, X., Cheng, Z., Deng, L., Ding, W., Gao, C., Ge, C., et al.: Qwen3-vl technical report. arXiv preprint arXiv:2511.21631 (2025) 2. Blecher, L., Cucurull, G., Scialom, T., Stojnic, R.: Nougat: Neural optical under- standing for academic documents. arXiv preprint arXiv:2308.13418 (2023) 3. Cui, C., Sun, T., Liang, S., Gao, T., Zhang, Z., Liu, J., Wang, X., Zhou, C., Liu, H., Lin, M., et al.: Paddleocr-vl-1.5: Towards a multi-task 0.9 b vlm for robust in-the-wild document parsing. arXiv preprint arXiv:2601.21957 (2026) 4. Cui, C., Sun, T., Lin, M., Gao, T., Zhang, Y., Liu, J., Wang, X., Zhang, Z., Zhou, C., Liu, H., et al.: Paddleocr 3.0 technical report. arXiv preprint arXiv:2507.05595 (2025) 5. Gervais, P., Fadeeva, A., Maksai, A.: Mathwriting: A dataset for handwritten math- ematical expression recognition. In: Proceedings of the 31st ACM SIGKDD Con- ference on Knowledge Discovery and Data Mining V. 2. p. 5459ā5469 (2025) 6. Hussain, M.: Yolo-v1 to yolo-v8, the rise of yolo and its complementary nature toward digital manufacturing and industrial defect detection. Machines 11(7), 677 (2023) 7. Jaume, G., Ekenel, H.K., Thiran, J.P.: Funsd: A dataset for form understanding in noisy scanned documents. arXiv preprint arXiv:1905.13538 (2019) 8. Li, C., Guo, R., Zhou, J., An, M., Du, Y., Zhu, L., Liu, Y., Hu, X., Yu, D.: Pp-structurev2: A stronger document analysis system. arXiv preprint arXiv:2210.05391 (2022) HANS: Noisy Hybrid Document Parsing15 9. Li, M., Lv, T., Chen, J., Cui, L., Lu, Y., Florencio, D., Zhang, C., Li, Z., Wei, F.: Trocr: Transformer-based optical character recognition with pre-trained models. In: Proceedings of the AAAI conference on artificial intelligence. vol. 37, p. 13094ā 13102 (2023) 10. Li, Z., Liu, Y., Liu, Q., Ma, Z., Zhang, Z., Zhang, S., Guo, Z., Zhang, J., Wang, X., Bai, X.: Monkeyocr: Document parsing with a structure-recognition-relation triplet paradigm. arXiv preprint arXiv:2506.05218 (2025) 11. Liu, C., Wei, H., Chen, J., Kong, L., Ge, Z., Zhu, Z., Zhao, L., Sun, J., Han, C., Zhang, X.: Focus anywhere for fine-grained multi-page document understanding. arXiv preprint arXiv:2405.14295 (2024) 12. Marti, U.V., Bunke, H.: The iam-database: an english sentence database for offline handwriting recognition. International journal on document analysis and recogni- tion 5(1), 39ā46 (2002) 13. Ouyang, L., Qu, Y., Zhou, H., Zhu, J., Zhang, R., Lin, Q., Wang, B., Zhao, Z., Jiang, M., Zhao, X., et al.: Omnidocbench: Benchmarking diverse pdf document parsing with comprehensive annotations. In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition. p. 24838ā24848 (2025) 14. Pfitzmann, B., Auer, C., Dolfi, M., Nassar, A.S., Staar, P.: Doclaynet: A large human-annotated dataset for document-layout segmentation. In: Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining. p. 3743ā3751 (2022) 15. Poznanski, J., Rangapur, A., Borchardt, J., Dunkelberger, J., Huff, R., Lin, D., Wilhelm, C., Lo, K., Soldaini, L.: olmocr: Unlocking trillions of tokens in pdfs with vision language models. arXiv preprint arXiv:2502.18443 (2025) 16. Poznanski, J., Soldaini, L., Lo, K.: olmocr 2: Unit test rewards for document ocr. arXiv preprint arXiv:2510.19817 (2025) 17. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, Å., Polosukhin, I.: Attention is all you need. Advances in neural information pro- cessing systems 30 (2017) 18. Wang, B., Xu, C., Zhao, X., Ouyang, L., Wu, F., Zhao, Z., Xu, R., Liu, K., Qu, Y., Shang, F., et al.: Mineru: An open-source solution for precise document content extraction. arXiv preprint arXiv:2409.18839 (2024) 19. Wang, W., Gao, Z., Gu, L., Pu, H., Cui, L., Wei, X., Liu, Z., Jing, L., Ye, S., Shao, J., et al.: Internvl3. 5: Advancing open-source multimodal models in versatility, reasoning, and efficiency. arXiv preprint arXiv:2508.18265 (2025) 20. Wei, H., Kong, L., Chen, J., Zhao, L., Ge, Z., Yang, J., Sun, J., Han, C., Zhang, X.: Vary: Scaling up the vision vocabulary for large vision-language model. In: European Conference on Computer Vision. p. 408ā424. Springer (2024) 21. Wei, H., Liu, C., Chen, J., Wang, J., Kong, L., Xu, Y., Ge, Z., Zhao, L., Sun, J., Peng, Y., et al.: General ocr theory: Towards ocr-2.0 via a unified end-to-end model. arXiv preprint arXiv:2409.01704 (2024) 22. Wei, H., Sun, Y., Li, Y.: Deepseek-ocr: Contexts optical compression. arXiv preprint arXiv:2510.18234 (2025) 23. Xia, R., Ye, H., Yan, X., Liu, Q., Zhou, H., Chen, Z., Shi, B., Yan, J., Zhang, B.: Chartx & chartvlm: A versatile benchmark and foundation model for complicated chart reasoning. IEEE Transactions on Image Processing (2025) 24. Xie, Y., MouchĆØre, H., Simistira Liwicki, F., Rakesh, S., Saini, R., Nakagawa, M., Nguyen, C.T., Truong, T.N.: Icdar 2023 crohme: Competition on recognition of handwritten mathematical expressions. In: International Conference on Document Analysis and Recognition. p. 553ā565. Springer (2023) 16X. Wu et al. 25. Xu, Y., Li, M., Cui, L., Huang, S., Wei, F., Zhou, M.: Layoutlm: Pre-training of text and layout for document image understanding. In: Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining. p. 1192ā1200 (2020) 26. Yang, Z., Tang, J., Li, Z., Wang, P., Wan, J., Zhong, H., Liu, X., Yang, M., Wang, P., Bai, S., et al.: Cc-ocr: A comprehensive and challenging ocr benchmark for evaluating large multimodal models in literacy. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. p. 21744ā21754 (2025)