SEP-LLM: Professional QA in the SEP Domain Using Retrieval-Augmented LLMs
DOI:
https://doi.org/10.4108/eetsis.10653Keywords:
Standard Essential Patent(SEP), Retrieval-Augmented Generation (RAG), Knowledge Graph, Low-Rank Adaptation (LoRA), Instruction Fine-TuningAbstract
INTRODUCTION: Question answering tasks in the Standard Essential Patent (SEP) domain impose high demands on models for professional terminology comprehension, regulatory interpretation, and factual accuracy. Existing general-purpose large language models show limitations in this field, mainly in knowledge retrieval accuracy, semantic matching, and legal compliance of generated content. Therefore, there is an urgent need to develop a specialized intelligent QA system tailored for the SEP domain.
OBJECTIVES: This paper aims to develop an intelligent QA system for the SEP domain, SEP-LLM, to improve knowledge retrieval, semantic matching, and content compliance, providing high-quality automated answers to SEP-related questions.
METHODS: We collected and curated a large set of SEP-related regulations, technical standards, and judicial cases to build a high-quality QA dataset. Leveraging the LightRAG framework, a large language model was used to extract entities and relationships from documents, constructing a structured SEP knowledge graph with incremental updates to ensure dynamic completeness. In retrieval, a two-layer strategy addresses both fine-grained entity queries and broader thematic searches, improving accuracy and coverage. In generation, DeepSeek-LLM-7B was fine-tuned with LoRA on SEP-specific instructions and terminology, enhancing the model’s understanding and generation capabilities while significantly reducing training and inference resource requirements.
RESULTS: Experimental results demonstrate that SEP-LLM significantly outperforms leading general-purpose models, including GPT-4o and Qwen3-235B, across three key metrics: BLEU-4, ROUGE-L, and Accuracy. These findings underscore its superior performance and promising potential for professional Quality Assurance within SEP domain.
CONCLUSION: The LightRAG-based SEP-LLM system effectively enhances knowledge retrieval, semantic understanding, and compliance in SEP QA tasks, demonstrating the potential of retrieval-augmented generation techniques in specialized domains and providing a practical solution for intelligent information services in the SEP field.
References
[1] YANG Z, WU X. Measurement of information disclosure level in standard-essential patent databases[J]. China Invention & Patents, 2025, 22(1): 4-15.
[2] KANG S. Challenges in the Determination of Standard-Essential Patents[EB]. (2024-01-02).
[3] SONG D, VOLD A, MADAN K, et al. Multi-label legal document classification: A deep learning-based approach with label-attention and domain-specific pre-training[J]. Information Systems, 2022, 106: 101718.
[4] WANG C, LUO X. A Legal Question Answering System Based on BERT[C]. Proceedings of the 2021 5th International Conference on Computer Science and Artificial Intelligence. New York, NY, USA: Association for Computing Machinery, 2022: 278-283.
[5] LIGA D, ROBALDO L. Fine-tuning GPT-3 for legal rule classification[J]. Computer Law & Security Review, 2023, 51: 105864.
[6] Mansurova A, Mansurova A, Nugumanova A. QA-RAG: Exploring LLM reliance on external knowledge[J]. Big Data and Cognitive Computing, 2024, 8(9): 115.
[7] LEWIS P, PEREZ E, PIKTUS A, et al. Retrieval-augmented generation for knowledge-intensive NLP tasks[C]//Proceedings of the 34th International Conference on Neural Information Processing Systems. Red Hook, NY, USA: Curran Associates Inc., 2020: 9459-9474.
[8] ZHAO Q, WANG R, CEN Y, et al. LongRAG: A Dual-Perspective Retrieval-Augmented Generation Paradigm for Long-Context Question Answering[C]. AL-ONAIZAN Y, BANSAL M, CHEN Y N. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. Miami, Florida, USA: Association for Computational Linguistics, 2024: 22600-22632.
[9] GOKDEMIR O, SIEBENSCHUH C, BRACE A, et al. HiPerRAG: High-Performance Retrieval Augmented Generation for Scientific Insights[C]. Proceedings of the Platform for Advanced Scientific Computing Conference. 2025: 1-13.
[10] LIM W, LI Z, KIM G, et al. MacRAG: Compress, Slice, and Scale-up for Multi-Scale Adaptive Context RAG[A]. arXiv, 2025.
[11] JIN J, LI X, DONG G, et al. Hierarchical Document Refinement for Long-context Retrieval-augmented Generation[A]. arXiv, 2025.
[12] PENTHEROUDAKIS C. TECHNICAL AND PRACTICAL ASPECTS RELATED TO PATENT QUALITY IN THE CONTEXT OF STANDARD ESSENTIAL PATENTS-An exploratory case study for WIPO[J].
[13] BARON J, POHLMANN T. Mapping standards to patents using declarations of standard-essential patents[J]. Journal of Economics & Management Strategy, 2018, 27(3): 504-534.
[14] TYAGI A, CHOPRA S. Standard Essential Patents (SEP’s)-Issues & Challenges in Developing Economies[J]. Journal of Intellectual Property Rights, 2017, 22: 121-135.
[15] GUO Z, XIA L, YU Y, et al. LightRAG: Simple and Fast Retrieval-Augmented Generation[A]. arXiv, 2025.
[16] NAZAR W, NAZAR G, KAMIŃSKA A, et al. How to Design, Create, and Evaluate an Instruction-Tuning Dataset for Large Language Model Training in Health Care: Tutorial From a Clinical Perspective[J]. Journal of Medical Internet Research, 2025, 27: e70481.
[17] MA Y, MIZUKI S, FUJII K, et al. Building Instruction-Tuning Datasets from Human-Written Instructions with Open-Weight Large Language Models[A]. arXiv, 2025. DOI:10.48550/arXiv.2503.23714.
[18] ABDALLA M, KASEM M S, MAHMOUD M, et al. ReceiptQA: A Question-Answering Dataset for Receipt Understanding[J]. Mathematics, 2025, 13(11).
[19] LIU Q, NIU Z, LIU S, et al. iTRI-QA: a Toolset for Customized Question-Answer Dataset Generation Using Language Models for Enhanced Scientific Research[A]. arXiv, 2025.
[20] YANG H, ZHANG Y, XU J, et al. Unveiling the Generalization Power of Fine-Tuned Large Language Models[A]. arXiv, 2024.
[21] CHUNG H W, HOU L, LONGPRE S, et al. Scaling instruction-finetuned language models[J]. Journal of Machine Learning Research, 2024, 25(70): 1-53.
[22] KUMAR A, RAGHUNATHAN A, JONES R, et al. Fine-tuning can distort pretrained features and underperform out-of-distribution[J]. arXiv preprint arXiv:2202.10054, 2022.
[23] HU E J, SHEN Y, WALLIS P, et al. Lora: Low-rank adaptation of large language models.[J]. ICLR, 2022, 1(2): 3.
Downloads
Published
Issue
Section
License
Copyright (c) 2026 Chenchen Guo, Kehao Wang, Dianhui Mao, Yunlong Xiong, Yiwen Lyu, Junhua Chen
This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.
This is an open access article distributed under the terms of the CC BY-NC-SA 4.0, which permits copying, redistributing, remixing, transformation, and building upon the material in any medium so long as the original work is properly cited.
