HANS: A Handwritten Answer Sheet Dataset for Noisy Hybrid Document Parsing

arXiv:2610.12363 · cs.AI, cs.CV · Submitted 2026-10-08 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "HANS: A Handwritten Answer Sheet Dataset for Noisy Hybrid Document Parsing".

Jane: The gist The introduction introduces HANS, a novel dataset for handwritten answer sheets that captures complex, real-world educational scenarios including mathematical expressions, natural language text,

Tom: First, who's behind it and why it matters.

Paper summary: Jane: We've talked about HANS and how it sets a new standard for document parsing benchmarks. So, what is the overall message here about this work by Xiazhen Wu and his team in "HANS: A Handwritten Answer Sheet Dataset for Noisy Hybrid Document Parsing"?

Tom: The main thing they’re arguing is that by building HANS, they’ve created a dataset that directly addresses the limitations we see in current document parsing methods.

Lu: They are providing fine-grained annotations across content level and spatial levels for noise elements like strikethroughs and deletions, which builds a reliable foundation for more robust recognition research into student answer sheets.

Meng: This means that future AI in education, specifically intelligent grading systems, can handle the real variability of student submissions much better because the input data won't just be a sanitized version of reality.

Lalam: This work pushes the vision in AI toward systems that can understand authentic student work with all its complexities.

Tom: It suggests that if we want AI to truly grade complex handwritten processes, we need datasets that mirror the actual, messy process students go through while answering problems on an exam.

Jane: And the authors are essentially saying this dataset is a necessary step toward making those systems actually work in a classroom setting, not just in a controlled lab environment.

Conclusion: Tom: So we've looked at HANS and how NA-GOT handles the noise, but now let's talk about what this paper is actually called and who wrote it.

Jane: This whole thing is about "HANS: A Handwritten Answer Sheet Dataset for Noisy Hybrid Document Parsing." It sounds like a mouthful, but basically, it's a big collection of messy student answer sheets.

Tom: Yeah, the authors are Xiazhen Wu and his team. They built this dataset specifically because existing benchmarks just aren't good enough for real-world stuff.

Lu: They show that they took actual math exams from a Chinese school district and turned those into data, including all the weird stuff like scribbles and corrections.

Meng: So what does this mean practically? It means that when we build AI to grade student work, it's going to be much better at handling the actual way kids write their answers.

Lalam: It shifts the focus from just recognizing neat text to understanding a whole process—how students think and solve problems while writing.

Tom: Exactly. Because this dataset has such detailed annotations on things like strikethroughs, it gives researchers a solid place to test how much noise an AI can actually handle.

Jane: And the whole point is that if we can build models that work well with this kind of data, then intelligent grading systems become much more reliable for education.

Lu: The implications are huge because it shows that we need to stop making datasets in a vacuum and start making ones that reflect the real mess of human learning.

Meng: It’s about moving from controlled testing to something that actually works when you throw real student papers at it during a grading cycle.

Lalam: And for the future, this points toward AI that can recognize the *thinking* behind a handwritten answer sheet, not just the final output.

Xiazhen Wu, Wansong Qin, Yangbin Zheng, Liangda Fang, Zhan Li, Xiujie Huang, Liushen Zhou, Quanlong Guan

Jinan University

cs.AI, cs.CV

Submitted: 2026-10-08

Updated: 2026-10-08

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 77/100

The gist: The gist The introduction introduces HANS, a novel dataset for handwritten answer sheets that captures complex, real-world educational scenarios including mathematical expressions, natural language

Key concepts

HANS Dataset
HANS is a novel dataset specifically built for real-world educational answer sheets. It captures the complexity of student work, including mathematical expressions, natural language text, hand-drawn tables containing mixed content, and specific noise artifacts like corrections and deletions. This makes it much more realistic than standard benchmarks.
Lightweight Noise Suppression Module (LNSM)
This module is the first stage of noise reduction in NA-GOT. It analyzes visual patches from the input image and predicts a 'noise confidence score' for each patch using a small neural network. This score is then used to suppress the magnitude of the visual tokens, effectively cleaning up noisy areas before main recognition.
Weighted Attention Correction (WAC)
This second stage modifies how the model focuses on information during decoding. It introduces a negative bias into the attention mechanism based on noise probabilities. This forces the decoder to pay less attention to regions flagged as having high noise probability, leading to more accurate recognition.
End-to-End Framework (NA-GOT)
NA-GOT is a complete system that handles both noise suppression and recognition in one go. It combines the LNSM for initial cleaning with WAC within the decoder's attention layers. This unified approach allows the model to learn how to recognize handwritten answers even when they are heavily corrupted by real-world noise.

Terminology

Summary

The gist The introduction introduces HANS, a novel dataset for handwritten answer sheets that captures complex, real-world educational scenarios including mathematical expressions, natural language text, and diverse noise artifacts like strikethroughs and deletions.

HANS Dataset Construction

  1. HANS is 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 The paper introduces 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.

  2. The data are sourced from eighth-grade mathematics quality assessment examinations conducted over three semesters in a municipal school district in China The data are sourced from eighth-grade mathematics quality assessment examinations conducted over three semesters in a municipal school district in China.

  3. The dataset contains 5,213 answer question images, covering 745 students across 18 distinct problem types The dataset contains 5,213 answer question images, covering 745 students across 18 distinct problem types.

  4. The writing patterns in HANS exhibit considerable complexity, with natural language explanations and mathematical expressions frequently appearing in interleaved configurations The writing patterns in HANS exhibit considerable complexity, with natural language explanations and mathematical expressions frequently appearing in interleaved configurations.

  5. The dataset includes a non-trivial proportion of hand-drawn tables, whose cells typically contain both natural language and mathematical content simultaneously The dataset includes a non-trivial proportion of hand-drawn tables, whose cells typically contain both natural language and mathematical content simultaneously.

  6. HANS provides fine-grained annotations across distinct writing modalities, including content-level annotations for expressions, text, and tables HANS provides fine-grained annotations across distinct writing modalities, including content-level annotations for expressions, text, and tables.

  7. Spatial annotations are provided for noise elements such as strikethroughs and deletions Spatial annotations are provided for noise elements such as strikethroughs and deletions.

NA-GOT Framework

The paper proposes NA-GOT, an end-to-end noise-aware recognition framework extending the GOT [21] architecture We further propose NA-GOT, an end-to-end noise-aware recognition framework extending the GOT [21] architecture.

This framework achieves two-stage noise suppression through a mechanism that addresses interference from real-world noise artifacts The proposed framework suppresses noise interference through a two-stage mechanism that addresses interference from real-world noise artifacts.

The first stage involves designing a Lightweight Noise Suppression Module (LNSM) that attenuates noise at the feature level In the first stage, we design a Lightweight Noise Suppression Module (LNSM) that attenuates noise at the feature level.

The second stage introduces a noise-aware attention mechanism that reduces attention weights assigned to noise-bearing regions via learned noise bias terms 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.

Lightweight Noise Suppression Module (LNSM)

The LNSM is constructed by first converting all annotations into a binary pixel mask matrix M according to the original image resolution We first construct patch-level supervision signals aligned with the visual patch tokens from the noise annotations.

This mask matrix M is uniformly resized to a resolution of 1024×1024, followed by 16×16 average pooling to compute the noise occupancy ratio over the patch grid This mask matrix M is uniformly resized to a resolution of 1024×1024, followed by 16×16 average pooling to compute the noise occupancy ratio over the patch grid.

The resulting grid is flattened into patch-level noise labels mg t in one-to-one correspondence with the visual tokens The resulting grid is flattened into patch-level noise labels mg t in one-to-one correspondence with the visual tokens.

The Noise Mask Head (NMH) employs a twolayer MLP to predict a noise confidence score p for each patch token The Noise Mask Head (NMH) employs a twolayer MLP to predict a noise confidence score p for each patch token.

This design enables the model to adaptively predict noise distributions during inference without relying on manual annotations or additional noise inputs This design enables the model to adaptively predict noise distributions during inference without relying on manual annotations or additional noise inputs.

The representational magnitude of noise tokens is then suppressed via the equation V˜ = V ⊙ 1 − α · p Specifically, the noise confidence scores p are mapped to gating coefficients, which are applied to scale the visual tokens on a per-token basis: V˜ = V ⊙ 1 − α · p.

Weighted Attention Correction (WAC)

The WAC introduces a negative bias derived from the noise probabilities p output by NMH into the decoder’s self-attention layers Weighted Attention Correction (WAC) introduces a negative bias derived from the noise probabilities p output by NMH into the decoder’s self-attention layers.

The bias term is defined as b = −β · p, where β is a hyperparameter controlling the penalty strength Since NA-GOT concatenates visual patch tokens and text tokens into a unified sequence, cross-modal alignment occurs within the decoder.

This mechanism penalizes noise patches during attention computation, directing the decoder’s focus toward valid handwriting regions A higher noise probability yields a larger bias, resulting in a lower attention weight for the corresponding patch following the softmax operation.

Experimental Results

Extensive experimental results demonstrate that HANS poses substantial challenges to existing methods Extensive experimental results demonstrate that HANS poses substantial challenges to existing methods.

NA-GOT achieves significant improvements in both accuracy and stability for answer sheet problem-solving process recognition NA-GOT achieves significant improvements in both accuracy and stability for answer sheet problem-solving process recognition.

The proposed method consistently surpasses all baselines across General VLMs, Specialized VLMs, and Pipeline Tools NA-GOT consistently surpasses all baselines across General VLMs, Specialized VLMs, and Pipeline Tools.

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 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.

The full NA-GOT model achieves the best overall performance across all evaluation metrics The full model achieves the best overall performance across all evaluation metrics.

NA-GOT demonstrates stronger robustness and generation stability in noisy areas, making fewer errors in misinterpreting erasure textures or strikethrough lines as characters or symbols NA-GOT demonstrates stronger robustness and generation stability in noisy areas, making fewer errors in misinterpreting erasure textures or strikethrough lines as characters or symbols.

Overall, NA-GOT provides more stable, structured, and accurate recognition outputs Overall, NA-GOT provides more stable, structured, and accurate recognition outputs.

Ablation Study Insights

The ablation study shows that the w/o WAC variant achieves a BLEU score of 0.589 and an Edit Distance of 0.314 The w/o WAC variant achieves a BLEU score of 0.589 and an Edit Distance of 0.314.

Improvements for AI systems

  1. Bold Header: HANS Dataset Construction

This dataset explicitly constructed for real-world educational scenarios encompasses mathematical expressions, natural language text, hand-drawn tables, and diverse noise patterns including corrections and deletions, which provides a reliable foundation for robust recognition research that existing benchmarks lack.

  1. Bold Header: NA-GOT Framework Implementation

The proposed NA-GOT framework achieves two-stage noise suppression through a Lightweight Noise Suppression Module (LNSM) that attenuates noise at the feature level followed by a noise-aware attention mechanism that reduces attention weights assigned to noise-bearing regions via learned noise bias terms.

  1. Bold Header: Enhanced Recognition Capability

The improved AI system can perform accurate recognition of students’ handwritten problem-solving processes, including step-wise scoring, error cause analysis, and personalized feedback, by achieving significant improvements in both accuracy and stability for answer sheet problem-solving process recognition compared to previous methods.

Sources

Related papers