<p>Deploying state-of-the-art document understanding models remains resource-intensive and impractical in many real-world scenarios, particularly where labeled data is scarce and computational budgets are constrained. To address these challenges, this work proposes a novel approach towards parameter-efficient document understanding models capable of adapting to specific tasks and document types without the need for labeled data. Specifically, we propose an approach coined <i>SlimDoc</i> to distill multimodal document transformer encoder models into smaller student models, using internal signals at different training stages, followed by external signals. Our approach is inspired by TinyBERT and adapted to the domain of document understanding transformers. We demonstrate SlimDoc to outperform both a single-stage distillation and a direct fine-tuning of the student. Experimental results across six document understanding datasets demonstrate our approach’s effectiveness: Our distilled student models achieve on average <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="10032_2025_542_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="43" /> </InlineMediaObject> <EquationSource Format="TEX">\(93.0\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>93.0</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> of the teacher’s performance, while the fine-tuned students achieve <InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="10032_2025_542_Article_IEq2.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="43" /> </InlineMediaObject> <EquationSource Format="TEX">\(87.0\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>87.0</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> of the teacher’s performance. Without requiring any labeled data, we create a compact student which achieves <InlineEquation ID="IEq3"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="10032_2025_542_Article_IEq3.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="43" /> </InlineMediaObject> <EquationSource Format="TEX">\(96.0\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>96.0</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> of the performance of its supervised-distilled counterpart and <InlineEquation ID="IEq4"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="10032_2025_542_Article_IEq4.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="43" /> </InlineMediaObject> <EquationSource Format="TEX">\(86.2\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>86.2</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> of the performance of a supervised-fine-tuned teacher model. We demonstrate our distillation approach to pick up on document geometry and to be effective on the two popular document understanding models LiLT and LayoutLMv3. Our implementation and training data is available at <a href="https://github.com/marcel-lamott/SlimDoc">https://github.com/marcel-lamott/SlimDoc</a>.</p>

错误:搜索内容不能为空,请输入英文关键词
错误:关键词超出字数限制,请精简
高级检索

SlimDoc: lightweight distillation of document transformer models

  • Marcel Lamott,
  • Muhammad Armaghan Shakir,
  • Adrian Ulges,
  • Yves-Noel Weweler,
  • Faisal Shafait

摘要

Deploying state-of-the-art document understanding models remains resource-intensive and impractical in many real-world scenarios, particularly where labeled data is scarce and computational budgets are constrained. To address these challenges, this work proposes a novel approach towards parameter-efficient document understanding models capable of adapting to specific tasks and document types without the need for labeled data. Specifically, we propose an approach coined SlimDoc to distill multimodal document transformer encoder models into smaller student models, using internal signals at different training stages, followed by external signals. Our approach is inspired by TinyBERT and adapted to the domain of document understanding transformers. We demonstrate SlimDoc to outperform both a single-stage distillation and a direct fine-tuning of the student. Experimental results across six document understanding datasets demonstrate our approach’s effectiveness: Our distilled student models achieve on average \(93.0\%\) 93.0 % of the teacher’s performance, while the fine-tuned students achieve \(87.0\%\) 87.0 % of the teacher’s performance. Without requiring any labeled data, we create a compact student which achieves \(96.0\%\) 96.0 % of the performance of its supervised-distilled counterpart and \(86.2\%\) 86.2 % of the performance of a supervised-fine-tuned teacher model. We demonstrate our distillation approach to pick up on document geometry and to be effective on the two popular document understanding models LiLT and LayoutLMv3. Our implementation and training data is available at https://github.com/marcel-lamott/SlimDoc.