Source-linked AI summary
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
TL;DR
In mathematics education, intelligent systems must recognize students’ derivation steps and natural-language explanations to understand and diagnose problem-solving processes. The paper introduces HANS with fine-grained process and noise annotations, and NA-GOT, which combines feature-level noise suppression with noise-aware attention. HANS poses substantial challenges to existing methods, while NA-GOT significantly improves answer-process recognition accuracy and output stability.
Problem
In mathematics education, intelligent systems must recognize students’ derivation steps and natural-language explanations to understand and diagnose problem-solving processes.
Source: §1 · p. 1Method
The paper introduces HANS with fine-grained process and noise annotations, and NA-GOT, which combines feature-level noise suppression with noise-aware attention.
Source: §1 · p. 3Results
HANS poses substantial challenges to existing methods, while NA-GOT significantly improves answer-process recognition accuracy and output stability.
Source: §1 · p. 3, §6 · p. 14Takeaways & Limitations
Recognizing answer-sheet processes supports step-wise scoring, error-cause analysis, and personalized feedback in intelligent grading.
Source: §1 · p. 2
Abstract
from arXiv · showhide
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.
1 Introduction
Existing benchmarks leave student answer sheets, mixed multi-line handwriting, and correction noise underrepresented. HANS addresses this gap, and NA-GOT adds two-stage noise suppression for answer-process recognition.
1 Introduction
- Existing benchmarks largely target printed documents, clean text lines, or isolated formulas, leaving noisy, multi-line answer-sheet processes underrepresented.Recognizing students’ derivations and explanations matters for step-wise scoring, error-cause analysis, and personalized feedback.Source: §1 · pp. 1–2
- HANS brings together mathematical expressions, natural language, hand-drawn tables, and correction artifacts with fine-grained content and spatial annotations.The annotations cover noise such as strikethroughs and deletions, supporting noise-aware recognition research.Source: §1 · p. 3
- NA-GOT extends GOT with feature-level suppression through LNSM and decoder-stage attention that reduces weights on noise-bearing regions.Its two-stage design uses learned noise bias terms in attention.Source: §1 · p. 3
- HANS challenges existing methods, while NA-GOT significantly outperforms them for answer-process recognition.The framework also improves recognition accuracy and output stability.Source: §1 · p. 3
2 Related Work
Prior datasets and models support document parsing and handwriting recognition, but handwritten educational content and systematic noise challenges remain limited. The section also frames dataset comparisons by sample count, granularity, modality, and noise annotation.
2.1 Document Parsing and Handwriting Recognition Datasets
- Document datasets support layout and form parsing, while handwriting benchmarks focus mainly on offline text or formula transcription and offer limited modality diversity.The reviewed datasets contain relatively little handwritten content and largely lack systematic investigation of noise-related challenges.Source: §2.1 · p. 4
2.2 Document Parsing and Handwriting Recognition Models
- Generative document models, layout-content models, handwriting recognizers, and formula-recognition systems provide approaches for heterogeneous content and structured transcription.The section concludes that educational answer sheets still require more specialized datasets and adaptive frameworks for robustness and accuracy.Source: §2.2 · p. 4
- Table 1 compares datasets by handwritten-sample count, page-or-line granularity, and the presence of handwritten formulas, text, and noise annotations.Its caption defines HS, Level, HME, HT, and NA.Source: Table 1 · p. 5
3 HANS Dataset
HANS comprises scanned middle-school mathematics answer regions with mixed handwriting and real-world correction artifacts. Its process and noise annotations are produced through a staged pipeline of automatic pre-annotation, human correction, and expert inspection.
3.1 Dataset Construction and Statistics
- HANS contains 5,213 answer-question images from 745 students across 18 problem types, with multi-line mixed content, hand-drawn tables, and correction noise.The source is eighth-grade mathematics assessments spanning three semesters in one municipal school district in China.Source: §3.1 · p. 5, Figure 2 · p. 6
3.2 Annotation Details
- Each answer region has a complete visible-process transcription that preserves line breaks, plus pixel-level annotations for erasures and strikethroughs.Deleted content is excluded from the final transcription; noise boundaries are fitted conservatively where noise overlaps valid writing.Source: §3.2 · p. 6, Problem-Solving Process Recognition… · p. 6, Noise Region Detection Annotations · pp. 6–7
3.3 Annotation Pipeline
- HANS annotations pass through automatic pre-annotation, annotator correction, and expert quality inspection.YOLOv8 localizes correction regions and DeepSeek-OCR drafts content; annotators revise both, then domain experts check accuracy and consistency.Source: §3.3 · p. 7
4 Method
NA-GOT extends GOT with two-stage noise suppression: LNSM attenuates noise at the feature level, while decoder noise-aware attention lowers weights on noise-bearing regions.
4.1 Lightweight Noise Suppression Module
- Noise annotations become patch-level occupancy labels, supervising NMH confidence predictions from visual features that require no manual annotations or extra noise inputs at inference.Source: §4.1 · pp. 8–9
- LNSM maps noise confidences to per-token gates that scale visual features; α controls gating strength.Source: §4.1 · p. 9
4.2 Weighted Attention Correction
- WAC adds a negative bias derived from NMH noise probabilities to decoder self-attention, lowering attention on higher-probability noise patches and directing focus toward valid handwriting.Source: §4.2 · p. 9
4.3 Loss Function
- Training combines autoregressive recognition cross-entropy with positive-weighted binary cross-entropy noise-mask loss, scaled by λmask.Source: §4.3 · pp. 9–10
5 Experiments
Experiments on HANS compare NA-GOT with representative document-recognition methods, test its noise-aware components, and examine key hyperparameters. Results show stronger overall recognition and greater stability in noisy answer-sheet regions.
5.1 Experimental Setup
- Experiments use an 8:2 HANS train-test split, with noisy samples comprising three-quarters of training data, and evaluate six sequence-alignment metrics against three baseline categories.GOT is fine-tuned on the full HANS training set for comparison; training uses patch-level noise supervision for the noise module.Source: Dataset and Evaluation Metrics · p. 10, Baselines · p. 10, Implementation Details · p. 10
5.2 Comparison to Previous Approaches
- NA-GOT consistently surpasses general VLMs, specialized VLMs, and pipeline tools on HANS, with especially pronounced gains over fine-tuned GOT on heavily noisy samples.Qwen3-VL is the strongest baseline, while directly transferred GOT performs relatively poorly on this task.Source: §5.2 · pp. 10–11, Table 2 · p. 11
5.3 Ablation Study
- The full NA-GOT model performs best across all evaluation metrics, while ablations show complementary benefits from token gating and decoder-level WAC.Without WAC, the model scores 0.589 BLEU and 0.314 Edit Distance; token gating contributes more to robust recognition than attention correction alone.Source: §5.3 · pp. 11–12, Table 3 · p. 12
5.4 Hyperparameter experimentation
- With α=0.7 and β=2.0, NA-GOT achieves optimal overall performance; weaker suppression leaves noise influential, while stronger suppression can impair valid features.The reported trade-offs are insufficient feature suppression at α=0.5 and excessive suppression at α=0.9; β=1.0 under-penalizes noise patches and β=3.0 mildly over-suppresses valid patches.Source: §5.4 · p. 12, Table 4 · p. 12, Table 5 · p. 13
5.5 Visual Analysis
- In visual examples, NA-GOT makes fewer noise-related errors and better preserves line order and formula structure than the compared models.The other models exhibit errors such as misreading near noise boundaries, unstable repetitive decoding, or omitted and disjointed content.Source: §5.5 · pp. 12–14, Figure 4 · p. 13
6 Conclusion
The paper introduces HANS for recognizing handwritten answer-sheet problem-solving processes and proposes NA-GOT for noise-aware recognition. The authors report that HANS challenges existing methods and NA-GOT improves recognition accuracy and output stability.
6 Conclusion
- HANS is a dataset for answer-sheet problem-solving recognition, with mathematical expressions, natural language, hand-drawn tables, noise artifacts, and fine-grained annotations.The paper describes it as a foundation for robust recognition research.Source: §6 · p. 14
- NA-GOT is an end-to-end framework that suppresses noise in two stages using a lightweight noise suppression module and noise-aware attention.Source: §6 · p. 14
- Experiments show that HANS challenges existing methods and NA-GOT improves recognition accuracy and output stability for answer-process transcription.The authors present it as a solution for handwritten document parsing in educational scenarios.Source: §6 · p. 14