Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Layout-Aware Parsing Meets Efficient LLMs: A Unified, Scalable Framework for Resume Information Extraction and Evaluation

Fanwei Zhu, Jinke Yu, Zulong Chen, Ying Zhou, Junhao Ji, Zhibo Yang, Yuxue Zhang, Haoyuan Hu, Zhenghao Liu

TL;DR

This paper tackles industrial-scale resume information extraction under three challenges: diverse layouts, high LLM latency, and scarce evaluation tools. It introduces a three-stage framework that (i) normalizes layouts via a layout-aware parser, (ii) performs efficient extraction with parallelized, instruction-tuned LLMs using index-based pointers, and (iii) evaluates results through a robust two-stage automated framework employing the Hungarian algorithm for alignment and multi-strategy field matching. The approach achieves state-of-the-art accuracy with compact models (e.g., Qwen-0.6B-SFT) and real-time performance, demonstrated on synthetic and real Alibaba resume datasets, and deployed in Alibaba’s HR system with high throughput. The work also contributes open-source pipelines and benchmark datasets to advance practical research in layout-aware document understanding and production-ready information extraction.

Abstract

Automated resume information extraction is critical for scaling talent acquisition, yet its real-world deployment faces three major challenges: the extreme heterogeneity of resume layouts and content, the high cost and latency of large language models (LLMs), and the lack of standardized datasets and evaluation tools. In this work, we present a layout-aware and efficiency-optimized framework for automated extraction and evaluation that addresses all three challenges. Our system combines a fine-tuned layout parser to normalize diverse document formats, an inference-efficient LLM extractor based on parallel prompting and instruction tuning, and a robust two-stage automated evaluation framework supported by new benchmark datasets. Extensive experiments show that our framework significantly outperforms strong baselines in both accuracy and efficiency. In particular, we demonstrate that a fine-tuned compact 0.6B LLM achieves top-tier accuracy while significantly reducing inference latency and computational cost. The system is fully deployed in Alibaba's intelligent HR platform, supporting real-time applications across its business units.

Layout-Aware Parsing Meets Efficient LLMs: A Unified, Scalable Framework for Resume Information Extraction and Evaluation

TL;DR

This paper tackles industrial-scale resume information extraction under three challenges: diverse layouts, high LLM latency, and scarce evaluation tools. It introduces a three-stage framework that (i) normalizes layouts via a layout-aware parser, (ii) performs efficient extraction with parallelized, instruction-tuned LLMs using index-based pointers, and (iii) evaluates results through a robust two-stage automated framework employing the Hungarian algorithm for alignment and multi-strategy field matching. The approach achieves state-of-the-art accuracy with compact models (e.g., Qwen-0.6B-SFT) and real-time performance, demonstrated on synthetic and real Alibaba resume datasets, and deployed in Alibaba’s HR system with high throughput. The work also contributes open-source pipelines and benchmark datasets to advance practical research in layout-aware document understanding and production-ready information extraction.

Abstract

Automated resume information extraction is critical for scaling talent acquisition, yet its real-world deployment faces three major challenges: the extreme heterogeneity of resume layouts and content, the high cost and latency of large language models (LLMs), and the lack of standardized datasets and evaluation tools. In this work, we present a layout-aware and efficiency-optimized framework for automated extraction and evaluation that addresses all three challenges. Our system combines a fine-tuned layout parser to normalize diverse document formats, an inference-efficient LLM extractor based on parallel prompting and instruction tuning, and a robust two-stage automated evaluation framework supported by new benchmark datasets. Extensive experiments show that our framework significantly outperforms strong baselines in both accuracy and efficiency. In particular, we demonstrate that a fine-tuned compact 0.6B LLM achieves top-tier accuracy while significantly reducing inference latency and computational cost. The system is fully deployed in Alibaba's intelligent HR platform, supporting real-time applications across its business units.

Paper Structure

This paper contains 28 sections, 8 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Approach
  4. Layout-Aware Parsing and Regeneration
  5. PDF Parser: Text Extraction & Fusion
  6. Layout Regeneration: Fine-tuned Layout Parser & Unification
  7. Parallelized Instruction-Tuned LLM Extractor
  8. Parallelized Extraction via Task Decomposition.
  9. Model Selection and Fine-Tuning.
  10. Post-Processing and Data Refinement
  11. Two-Stage Automatic Evaluation
  12. Experiments
  13. Experimental Settings
  14. Datasets
  15. Methods for Comparison
  16. ...and 13 more sections

Figures (8)

  • Figure 1: Overview of the layout-aware, LLM-powered resume extraction and evaluation pipeline.
  • Figure 2: Impact of Hyper-parameters.
  • Figure 3: System Deployment and Serving Framework.
  • Figure 4: Prompt for Extracting Basic Information in JSON Format.
  • Figure 5: Prompt for Extracting Education Background in JSON Format.
  • ...and 3 more figures